add qwen3

vllm-v0.6.2/vllm/__init__.py

@@ -0,0 +1,58 @@
+"""vLLM: a high-throughput and memory-efficient inference engine for LLMs"""
+
+import os
+os.environ['PYTORCH_CNDEV_BASED_MLU_CHECK'] = '1'
+os.environ['CN_NOTIFIER_POOL_MAX'] = "1000"
+
+from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
+from vllm.engine.async_llm_engine import AsyncLLMEngine
+from vllm.engine.llm_engine import LLMEngine
+from vllm.entrypoints.llm import LLM
+from vllm.executor.ray_utils import initialize_ray_cluster
+from vllm.inputs import PromptType, TextPrompt, TokensPrompt
+from vllm.model_executor.models import ModelRegistry
+from vllm.outputs import (CompletionOutput, EmbeddingOutput,
+                          EmbeddingRequestOutput, RequestOutput)
+from vllm.pooling_params import PoolingParams
+from vllm.sampling_params import SamplingParams
+
+from .version import (__version__, __version_tuple__,
+                      __vllm_mlu_version__, __torch_version__)
+
+
+from vllm.platforms import current_platform
+
+if current_platform.is_mlu():
+    try:
+        import vllm_mlu
+        print("\033[0;32mApply vllm_mlu success, running in performance version !\033[0m")
+    except ModuleNotFoundError:
+        print("\033[0;31mApply vllm_mlu failed, running in basic version !\033[0m")
+    except Exception as e:
+        print("\033[0;31mApply vllm_mlu failed!\033[0m")
+        raise Exception(e)
+
+
+__version__ = f"{__version__}+mlu{__vllm_mlu_version__}.pt{__torch_version__}"
+
+
+__all__ = [
+    "__version__",
+    "__version_tuple__",
+    "LLM",
+    "ModelRegistry",
+    "PromptType",
+    "TextPrompt",
+    "TokensPrompt",
+    "SamplingParams",
+    "RequestOutput",
+    "CompletionOutput",
+    "EmbeddingOutput",
+    "EmbeddingRequestOutput",
+    "LLMEngine",
+    "EngineArgs",
+    "AsyncLLMEngine",
+    "AsyncEngineArgs",
+    "initialize_ray_cluster",
+    "PoolingParams",
+]

vllm-v0.6.2/vllm/_custom_ops.py

@@ -0,0 +1,984 @@
+import contextlib
+import functools
+import importlib
+from typing import TYPE_CHECKING, List, Optional, Tuple, Union
+
+import torch
+import torch.library
+
+import vllm.envs as envs
+from vllm.logger import init_logger
+from vllm.platforms import current_platform
+from vllm.scalar_type import ScalarType
+
+logger = init_logger(__name__)
+
+if (not current_platform.is_tpu()
+        and not current_platform.is_hpu()
+        and not current_platform.is_mlu()):
+    try:
+        import vllm._C
+    except ImportError as e:
+        logger.warning("Failed to import from vllm._C with %r", e)
+
+if current_platform.is_rocm():
+    import vllm._rocm_C  # noqa: F401
+
+supports_moe_ops = False
+with contextlib.suppress(ImportError):
+    import vllm._moe_C  # noqa: F401
+    supports_moe_ops = True
+
+# neuron has torch version that doesn't even have impl_abstract
+if TYPE_CHECKING or current_platform.is_neuron():
+
+    def register_fake(fn):
+        return lambda name: fn
+else:
+    try:
+        from torch.library import register_fake
+    except ImportError:
+        from torch.library import impl_abstract as register_fake
+
+
+def hint_on_error(fn):
+
+    @functools.wraps(fn)
+    def wrapper(*args, **kwargs):
+        try:
+            return fn(*args, **kwargs)
+
+        except NotImplementedError as e:
+            msg = (
+                "Error in calling custom op %s: %s\n"
+                "Not implemented or built, mostly likely because the current current device "
+                "does not support this kernel (less likely TORCH_CUDA_ARCH_LIST was set "
+                "incorrectly while building)")
+            logger.error(msg, fn.__name__, e)
+            raise NotImplementedError(msg % (fn.__name__, e)) from e
+        except AttributeError as e:
+            msg = (
+                "Error in calling custom op %s: %s\n"
+                "Possibly you have built or installed an obsolete version of vllm.\n"
+                "Please try a clean build and install of vllm,"
+                "or remove old built files such as vllm/*cpython*.so and build/ ."
+            )
+            logger.error(msg, fn.__name__, e)
+            raise e
+
+    return wrapper
+
+
+# activation ops
+def silu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
+    torch.ops._C.silu_and_mul(out, x)
+
+
+def gelu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
+    torch.ops._C.gelu_and_mul(out, x)
+
+
+def gelu_tanh_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
+    torch.ops._C.gelu_tanh_and_mul(out, x)
+
+
+def fatrelu_and_mul(out: torch.Tensor,
+                    x: torch.Tensor,
+                    threshold: float = 0.0) -> None:
+    torch.ops._C.fatrelu_and_mul(out, x, threshold)
+
+
+def gelu_fast(out: torch.Tensor, x: torch.Tensor) -> None:
+    torch.ops._C.gelu_fast(out, x)
+
+
+def gelu_new(out: torch.Tensor, x: torch.Tensor) -> None:
+    torch.ops._C.gelu_new(out, x)
+
+
+def gelu_quick(out: torch.Tensor, x: torch.Tensor) -> None:
+    torch.ops._C.gelu_quick(out, x)
+
+
+# page attention ops
+def paged_attention_v1(
+    out: torch.Tensor,
+    query: torch.Tensor,
+    key_cache: torch.Tensor,
+    value_cache: torch.Tensor,
+    num_kv_heads: int,
+    scale: float,
+    block_tables: torch.Tensor,
+    seq_lens: torch.Tensor,
+    block_size: int,
+    max_seq_len: int,
+    alibi_slopes: Optional[torch.Tensor],
+    kv_cache_dtype: str,
+    k_scale: float,
+    v_scale: float,
+    tp_rank: int = 0,
+    blocksparse_local_blocks: int = 0,
+    blocksparse_vert_stride: int = 0,
+    blocksparse_block_size: int = 64,
+    blocksparse_head_sliding_step: int = 0,
+) -> None:
+    torch.ops._C.paged_attention_v1(
+        out, query, key_cache, value_cache, num_kv_heads, scale, block_tables,
+        seq_lens, block_size, max_seq_len, alibi_slopes, kv_cache_dtype,
+        k_scale, v_scale, tp_rank, blocksparse_local_blocks,
+        blocksparse_vert_stride, blocksparse_block_size,
+        blocksparse_head_sliding_step)
+
+
+def paged_attention_v2(
+    out: torch.Tensor,
+    exp_sum: torch.Tensor,
+    max_logits: torch.Tensor,
+    tmp_out: torch.Tensor,
+    query: torch.Tensor,
+    key_cache: torch.Tensor,
+    value_cache: torch.Tensor,
+    num_kv_heads: int,
+    scale: float,
+    block_tables: torch.Tensor,
+    seq_lens: torch.Tensor,
+    block_size: int,
+    max_seq_len: int,
+    alibi_slopes: Optional[torch.Tensor],
+    kv_cache_dtype: str,
+    k_scale: float,
+    v_scale: float,
+    tp_rank: int = 0,
+    blocksparse_local_blocks: int = 0,
+    blocksparse_vert_stride: int = 0,
+    blocksparse_block_size: int = 64,
+    blocksparse_head_sliding_step: int = 0,
+) -> None:
+    torch.ops._C.paged_attention_v2(
+        out, exp_sum, max_logits, tmp_out, query, key_cache, value_cache,
+        num_kv_heads, scale, block_tables, seq_lens, block_size, max_seq_len,
+        alibi_slopes, kv_cache_dtype, k_scale, v_scale, tp_rank,
+        blocksparse_local_blocks, blocksparse_vert_stride,
+        blocksparse_block_size, blocksparse_head_sliding_step)
+
+
+def paged_attention_rocm(
+    out: torch.Tensor,
+    exp_sum: torch.Tensor,
+    max_logits: torch.Tensor,
+    tmp_out: torch.Tensor,
+    query: torch.Tensor,
+    key_cache: torch.Tensor,
+    value_cache: torch.Tensor,
+    num_kv_heads: int,
+    scale: float,
+    block_tables: torch.Tensor,
+    seq_lens: torch.Tensor,
+    block_size: int,
+    max_seq_len: int,
+    alibi_slopes: Optional[torch.Tensor],
+    kv_cache_dtype: str,
+    k_scale: float,
+    v_scale: float,
+) -> None:
+    torch.ops._rocm_C.paged_attention(out, exp_sum, max_logits, tmp_out, query,
+                                      key_cache, value_cache, num_kv_heads,
+                                      scale, block_tables, seq_lens,
+                                      block_size, max_seq_len, alibi_slopes,
+                                      kv_cache_dtype, k_scale, v_scale)
+
+
+# pos encoding ops
+def rotary_embedding(
+    positions: torch.Tensor,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    head_size: int,
+    cos_sin_cache: torch.Tensor,
+    is_neox: bool,
+) -> None:
+    torch.ops._C.rotary_embedding(positions, query, key, head_size,
+                                  cos_sin_cache, is_neox)
+
+
+def batched_rotary_embedding(positions: torch.Tensor, query: torch.Tensor,
+                             key: torch.Tensor, head_size: int,
+                             cos_sin_cache: torch.Tensor, is_neox: bool,
+                             rot_dim: int,
+                             cos_sin_cache_offsets: torch.Tensor) -> None:
+    torch.ops._C.batched_rotary_embedding(positions, query, key, head_size,
+                                          cos_sin_cache, is_neox, rot_dim,
+                                          cos_sin_cache_offsets)
+
+
+# layer norm ops
+def rms_norm(out: torch.Tensor, input: torch.Tensor, weight: torch.Tensor,
+             epsilon: float) -> None:
+    torch.ops._C.rms_norm(out, input, weight, epsilon)
+
+
+def fused_add_rms_norm(input: torch.Tensor, residual: torch.Tensor,
+                       weight: torch.Tensor, epsilon: float) -> None:
+    torch.ops._C.fused_add_rms_norm(input, residual, weight, epsilon)
+
+
+def advance_step_flashattn(num_seqs: int, num_queries: int, block_size: int,
+                           input_tokens: torch.Tensor,
+                           sampled_token_ids: torch.Tensor,
+                           input_positions: torch.Tensor,
+                           seq_lens: torch.Tensor, slot_mapping: torch.Tensor,
+                           block_tables: torch.Tensor) -> None:
+    """Advance a step on GPU for existing inputs for a multi-step runner"""
+    return torch.ops._C.advance_step_flashattn(num_seqs, num_queries,
+                                               block_size, input_tokens,
+                                               sampled_token_ids,
+                                               input_positions, seq_lens,
+                                               slot_mapping, block_tables)
+
+
+def advance_step_flashinfer(num_seqs: int, num_queries: int, block_size: int,
+                            input_tokens: torch.Tensor,
+                            sampled_token_ids: torch.Tensor,
+                            input_positions: torch.Tensor,
+                            seq_lens: torch.Tensor, slot_mapping: torch.Tensor,
+                            block_tables: torch.Tensor,
+                            paged_kv_indices: torch.Tensor,
+                            paged_kv_indptr: torch.Tensor,
+                            paged_kv_last_page_len: torch.Tensor,
+                            block_table_bound: torch.Tensor) -> None:
+
+    return torch.ops._C.advance_step_flashinfer(
+        num_seqs, num_queries, block_size, input_tokens, sampled_token_ids,
+        input_positions, seq_lens, slot_mapping, block_tables,
+        paged_kv_indices, paged_kv_indptr, paged_kv_last_page_len,
+        block_table_bound)
+
+
+# quantization ops
+# awq
+def awq_dequantize(qweight: torch.Tensor, scales: torch.Tensor,
+                   zeros: torch.Tensor, split_k_iters: int, thx: int,
+                   thy: int) -> torch.Tensor:
+    if envs.VLLM_USE_TRITON_AWQ:
+        from vllm.model_executor.layers.quantization.awq_triton import (
+            awq_dequantize_triton)
+        return awq_dequantize_triton(qweight, scales, zeros)
+    return torch.ops._C.awq_dequantize(qweight, scales, zeros, split_k_iters,
+                                       thx, thy)
+
+
+def awq_gemm(input: torch.Tensor, qweight: torch.Tensor, qzeros: torch.Tensor,
+             scales: torch.Tensor, split_k_iters: int) -> torch.Tensor:
+    if envs.VLLM_USE_TRITON_AWQ:
+        from vllm.model_executor.layers.quantization.awq_triton import (
+            awq_gemm_triton)
+        return awq_gemm_triton(input, qweight, qzeros, scales, split_k_iters)
+    return torch.ops._C.awq_gemm(input, qweight, qzeros, scales, split_k_iters)
+
+
+# gptq
+def gptq_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
+              b_gptq_qzeros: torch.Tensor, b_gptq_scales: torch.Tensor,
+              b_g_idx: torch.Tensor, use_exllama: bool,
+              bit: int) -> torch.Tensor:
+    return torch.ops._C.gptq_gemm(a, b_q_weight, b_gptq_qzeros, b_gptq_scales,
+                                  b_g_idx, use_exllama, bit)
+
+
+if hasattr(torch.ops._C, "gptq_gemm"):
+
+    @register_fake("_C::gptq_gemm")
+    def _gptq_gemm_fake(a: torch.Tensor, b_q_weight: torch.Tensor,
+                        b_gptq_qzeros: torch.Tensor,
+                        b_gptq_scales: torch.Tensor, b_g_idx: torch.Tensor,
+                        use_exllama: bool, bit: int) -> torch.Tensor:
+        return torch.empty((a.size(0), b_q_weight.size(1)),
+                           dtype=a.dtype,
+                           device=a.device)
+
+
+def gptq_shuffle(q_weight: torch.Tensor, q_perm: torch.Tensor,
+                 bit: int) -> None:
+    torch.ops._C.gptq_shuffle(q_weight, q_perm, bit)
+
+
+# marlin
+def marlin_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
+                b_scales: torch.Tensor, workspace: torch.Tensor, size_m: int,
+                size_n: int, size_k: int) -> torch.Tensor:
+    return torch.ops._C.marlin_gemm(a, b_q_weight, b_scales, workspace, size_m,
+                                    size_n, size_k)
+
+
+# marlin_24
+def gptq_marlin_24_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
+                        b_meta: torch.Tensor, b_scales: torch.Tensor,
+                        workspace: torch.Tensor, b_q_type: ScalarType,
+                        size_m: int, size_n: int, size_k: int) -> torch.Tensor:
+    return torch.ops._C.gptq_marlin_24_gemm(a, b_q_weight, b_meta, b_scales,
+                                            workspace, b_q_type.id, size_m,
+                                            size_n, size_k)
+
+
+if hasattr(torch.ops._C, "gptq_marlin_24_gemm"):
+
+    @register_fake("_C::gptq_marlin_24_gemm")
+    def _gptq_marlin_24_gemm_fake(a: torch.Tensor, b_q_weight: torch.Tensor,
+                                  b_meta: torch.Tensor, b_scales: torch.Tensor,
+                                  workspace: torch.Tensor,
+                                  b_q_type: ScalarType, size_m: torch.SymInt,
+                                  size_n: torch.SymInt,
+                                  size_k: torch.SymInt) -> torch.Tensor:
+        return torch.empty((size_m, size_n), device=a.device, dtype=a.dtype)
+
+    @register_fake("_C::gptq_marlin_gemm")
+    def _gptq_marlin_gemm_fake(a: torch.Tensor,
+                               b_q_weight: torch.Tensor,
+                               b_scales: torch.Tensor,
+                               b_zeros: torch.Tensor,
+                               g_idx: torch.Tensor,
+                               perm: torch.Tensor,
+                               workspace: torch.Tensor,
+                               b_q_type: ScalarType,
+                               size_m: torch.SymInt,
+                               size_n: torch.SymInt,
+                               size_k: torch.SymInt,
+                               is_k_full: bool,
+                               has_zp: bool = False,
+                               use_fp32_reduce: bool = False) -> torch.Tensor:
+        return torch.empty((size_m, size_n), device=a.device, dtype=a.dtype)
+
+    @register_fake("_C::ggml_dequantize")
+    def _ggml_dequantize_fake(W: torch.Tensor, quant_type: int,
+                              m: torch.SymInt,
+                              n: torch.SymInt) -> torch.Tensor:
+        return torch.empty((m, n), dtype=torch.float16, device=W.device)
+
+    @register_fake("_C::ggml_mul_mat_vec_a8")
+    def _ggml_mul_mat_vec_a8_fake(
+        W: torch.Tensor,
+        X: torch.Tensor,
+        quant_type: int,
+        row: torch.SymInt,
+    ) -> torch.Tensor:
+        return torch.empty((1, row), dtype=torch.float16, device=W.device)
+
+    @register_fake("_C::ggml_mul_mat_a8")
+    def _ggml_mul_mat_a8_fake(
+        W: torch.Tensor,
+        X: torch.Tensor,
+        quant_type: int,
+        row: torch.SymInt,
+    ) -> torch.Tensor:
+        batch = X.size(0)
+        return torch.empty((batch, row), dtype=torch.float16, device=W.device)
+
+    @register_fake("_C::marlin_qqq_gemm")
+    def _marlin_qqq_gemm_fake(a: torch.Tensor, b_q_weight: torch.Tensor,
+                              s_tok: torch.Tensor, s_ch: torch.Tensor,
+                              s_group: torch.Tensor, workspace: torch.Tensor,
+                              size_m: torch.SymInt, size_n: torch.SymInt,
+                              size_k: torch.SymInt) -> torch.Tensor:
+        return torch.empty((size_m, size_n),
+                           dtype=torch.float16,
+                           device=a.device)
+
+    @register_fake("_C::marlin_gemm")
+    def _marlin_gemm_fake(a: torch.Tensor, b_q_weight: torch.Tensor,
+                          b_scales: torch.Tensor, workspace: torch.Tensor,
+                          size_m: torch.SymInt, size_n: torch.SymInt,
+                          size_k: torch.SymInt) -> torch.Tensor:
+        return torch.empty((size_m, size_n),
+                           dtype=torch.float16,
+                           device=a.device)
+
+    @register_fake("_C::awq_dequantize")
+    def _awq_dequantize_fake(qweight: torch.Tensor, scales: torch.Tensor,
+                             zeros: torch.Tensor, split_k_iters: torch.SymInt,
+                             thx: int, thy: int) -> torch.Tensor:
+        in_c = qweight.size(0)
+        qout_c = qweight.size(1)
+        out_c = qout_c * 8
+        return torch.empty((in_c, out_c),
+                           dtype=scales.dtype,
+                           device=scales.device)
+
+    @register_fake("_C::awq_gemm")
+    def _awq_gemm_fake(input: torch.Tensor, qweight: torch.Tensor,
+                       qzeros: torch.Tensor, scales: torch.Tensor,
+                       split_k_iters: torch.SymInt) -> torch.Tensor:
+        num_in_feats = input.size(0)
+        return torch.empty((split_k_iters, num_in_feats, qweight.size(1) * 8),
+                           dtype=input.dtype,
+                           device=input.device).sum(0)
+
+    @register_fake("_C::aqlm_gemm")
+    def _aqlm_gemm_fake(input: torch.Tensor, codes: torch.Tensor,
+                        codebooks: torch.Tensor, scales: torch.Tensor,
+                        codebook_partition_sizes: List[int],
+                        bias: Optional[torch.Tensor]) -> torch.Tensor:
+        out_features = codes.size(0) * codebooks.size(2)
+        flat_input = input.reshape((-1, input.size(-1)))
+        flat_output = torch.empty((flat_input.size(0), out_features),
+                                  dtype=input.dtype,
+                                  device=input.device)
+
+        output_sizes = list(input.shape)
+        output_sizes.pop()
+        output_sizes.append(-1)
+        return flat_output.reshape(tuple(output_sizes))
+
+    @register_fake("_C::aqlm_dequant")
+    def _aqlm_dequant_fake(
+            codes: torch.Tensor, codebooks: torch.Tensor,
+            codebook_partition_sizes: List[int]) -> torch.Tensor:
+        in_features = codes.size(1) * 8
+        out_features = codes.size(0)
+        return torch.empty((out_features, in_features),
+                           dtype=codebooks.dtype,
+                           device=codebooks.device)
+
+    @register_fake("_C::fp8_marlin_gemm")
+    def _fp8_marlin_gemm_fake(a: torch.Tensor, b_q_weight: torch.Tensor,
+                              b_scales: torch.Tensor, workspace: torch.Tensor,
+                              num_bits: int, size_m: torch.SymInt,
+                              size_n: torch.SymInt,
+                              size_k: torch.SymInt) -> torch.Tensor:
+        return torch.empty((size_m, size_n), dtype=a.dtype, device=a.device)
+
+    @register_fake("_C::machete_gemm")
+    def machete_gemm_fake(
+        a: torch.Tensor,
+        # Should be the tensor returned by machete_prepack_B
+        b_q: torch.Tensor,
+        b_type: ScalarType,
+        b_scales: Optional[torch.Tensor] = None,
+        b_zeros: Optional[torch.Tensor] = None,
+        b_group_size: Optional[int] = None,
+        c: Optional[torch.Tensor] = None,
+        alpha: Optional[float] = None,
+        beta: Optional[float] = None,
+        schedule: Optional[str] = None,
+    ) -> torch.Tensor:
+        m = a.size(0)
+        n = b_q.size(1)
+        return torch.empty((m, n), device=a.device, dtype=a.dtype)
+
+    @register_fake("_C::machete_prepack_B")
+    def machete_prepack_B_fake(b_q_weight: torch.Tensor,
+                               b_type: ScalarType) -> torch.Tensor:
+        return torch.empty_like(b_q_weight,
+                                memory_format=torch.contiguous_format)
+
+
+# cutlass
+def cutlass_scaled_mm_supports_fp8(cuda_device_capability: int) -> bool:
+    return torch.ops._C.cutlass_scaled_mm_supports_fp8(cuda_device_capability)
+
+
+def cutlass_scaled_mm(a: torch.Tensor,
+                      b: torch.Tensor,
+                      scale_a: torch.Tensor,
+                      scale_b: torch.Tensor,
+                      out_dtype: torch.dtype,
+                      bias: Optional[torch.Tensor] = None) -> torch.Tensor:
+    assert (b.shape[0] % 16 == 0 and b.shape[1] % 16 == 0)
+    assert (out_dtype is torch.bfloat16 or out_dtype is torch.float16)
+    assert bias is None or bias.shape[0] == b.shape[
+        1] and bias.dtype == out_dtype
+
+    m = a.shape[0]
+    n = b.shape[1]
+
+    if current_platform.is_rocm():
+        triton_scaled_mm_module = importlib.import_module(
+            "vllm.model_executor.layers.quantization.compressed_tensors."
+            "triton_scaled_mm")
+        triton_scaled_mm = triton_scaled_mm_module.triton_scaled_mm
+        return triton_scaled_mm(a, b, scale_a, scale_b, out_dtype, bias)
+
+    out = torch.empty((m, n), dtype=out_dtype, device=a.device)
+
+    torch.ops._C.cutlass_scaled_mm(out, a, b, scale_a, scale_b, bias)
+
+    return out
+
+
+def cutlass_scaled_mm_azp(a: torch.Tensor,
+                          b: torch.Tensor,
+                          scale_a: torch.Tensor,
+                          scale_b: torch.Tensor,
+                          out_dtype: torch.dtype,
+                          azp_adj: torch.Tensor,
+                          azp: Optional[torch.Tensor] = None,
+                          bias: Optional[torch.Tensor] = None) -> torch.Tensor:
+    """
+    :param azp_adj: In the per-tensor case, this should include the azp.
+    Always per-channel.
+    :param azp: Only set in the per-token case. Per-token if set.
+    """
+    assert (b.shape[0] % 16 == 0 and b.shape[1] % 16 == 0)
+    assert (out_dtype is torch.bfloat16 or out_dtype is torch.float16)
+    assert bias is None or bias.numel(
+    ) == b.shape[1] and bias.dtype == out_dtype
+    assert azp is None or azp.numel() == a.shape[0]
+
+    m = a.shape[0]
+    n = b.shape[1]
+    out = torch.empty((m, n), dtype=out_dtype, device=a.device)
+
+    torch.ops._C.cutlass_scaled_mm_azp(out, a, b, scale_a, scale_b, azp_adj,
+                                       azp, bias)
+    return out
+
+
+# aqlm
+def aqlm_gemm(input: torch.Tensor, codes: torch.Tensor,
+              codebooks: torch.Tensor, scales: torch.Tensor,
+              codebook_partition_sizes: List[int],
+              bias: Optional[torch.Tensor]) -> torch.Tensor:
+    return torch.ops._C.aqlm_gemm(input, codes, codebooks, scales,
+                                  codebook_partition_sizes, bias)
+
+
+def aqlm_dequant(codes: torch.Tensor, codebooks: torch.Tensor,
+                 codebook_partition_sizes: List[int]) -> torch.Tensor:
+    return torch.ops._C.aqlm_dequant(codes, codebooks,
+                                     codebook_partition_sizes)
+
+
+# gptq_marlin
+def gptq_marlin_repack(b_q_weight: torch.Tensor, perm: torch.Tensor,
+                       size_k: int, size_n: int,
+                       num_bits: int) -> torch.Tensor:
+    return torch.ops._C.gptq_marlin_repack(b_q_weight, perm, size_k, size_n,
+                                           num_bits)
+
+
+# gptq_marlin
+def awq_marlin_repack(b_q_weight: torch.Tensor, size_k: int, size_n: int,
+                      num_bits: int) -> torch.Tensor:
+    return torch.ops._C.awq_marlin_repack(b_q_weight, size_k, size_n, num_bits)
+
+
+def gptq_marlin_moe_repack(b_q_weight: torch.Tensor, perm: torch.Tensor,
+                           size_k: int, size_n: int,
+                           num_bits: int) -> torch.Tensor:
+    num_experts = b_q_weight.shape[0]
+    assert size_k % 16 == 0
+    output = torch.empty((num_experts, size_k // 16, size_n * (num_bits // 2)),
+                         device=b_q_weight.device,
+                         dtype=b_q_weight.dtype)
+    for e in range(num_experts):
+        output[e] = torch.ops._C.gptq_marlin_repack(b_q_weight[e], perm[e],
+                                                    size_k, size_n, num_bits)
+    return output
+
+
+def awq_marlin_moe_repack(b_q_weight: torch.Tensor, perm: torch.Tensor,
+                          size_k: int, size_n: int,
+                          num_bits: int) -> torch.Tensor:
+    num_experts = b_q_weight.shape[0]
+    assert size_k % 16 == 0
+    output = torch.empty((num_experts, size_k // 16, size_n * (num_bits // 2)),
+                         device=b_q_weight.device,
+                         dtype=b_q_weight.dtype)
+    for e in range(num_experts):
+        output[e] = torch.ops._C.awq_marlin_repack(b_q_weight[e], size_k,
+                                                   size_n, num_bits)
+    return output
+
+
+def gptq_marlin_gemm(a: torch.Tensor,
+                     b_q_weight: torch.Tensor,
+                     b_scales: torch.Tensor,
+                     b_zeros: torch.Tensor,
+                     g_idx: torch.Tensor,
+                     perm: torch.Tensor,
+                     workspace: torch.Tensor,
+                     b_q_type: ScalarType,
+                     size_m: int,
+                     size_n: int,
+                     size_k: int,
+                     is_k_full: bool,
+                     has_zp: bool = False,
+                     use_fp32_reduce: bool = False) -> torch.Tensor:
+    return torch.ops._C.gptq_marlin_gemm(a, b_q_weight, b_scales, b_zeros,
+                                         g_idx, perm, workspace, b_q_type.id,
+                                         size_m, size_n, size_k, is_k_full,
+                                         has_zp, use_fp32_reduce)
+
+
+# fp8 marlin
+def fp8_marlin_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
+                    b_scales: torch.Tensor, workspace: torch.Tensor,
+                    num_bits: int, size_m: int, size_n: int,
+                    size_k: int) -> torch.Tensor:
+    return torch.ops._C.fp8_marlin_gemm(a, b_q_weight, b_scales, workspace,
+                                        num_bits, size_m, size_n, size_k)
+
+
+# machete
+def machete_supported_schedules(b_type: ScalarType) -> List[str]:
+    return torch.ops._C.machete_supported_schedules(b_type.id)
+
+
+def machete_gemm(
+    a: torch.Tensor,
+    b_q: torch.Tensor,  # Should be the tensor returned by machete_prepack_B
+    b_type: ScalarType,
+    b_scales: Optional[torch.Tensor] = None,
+    b_zeros: Optional[torch.Tensor] = None,
+    b_group_size: Optional[int] = None,
+    c: Optional[torch.Tensor] = None,
+    alpha: Optional[float] = None,
+    beta: Optional[float] = None,
+    schedule: Optional[str] = None,
+) -> torch.Tensor:
+    return torch.ops._C.machete_gemm(a, b_q, b_type.id, b_scales, b_zeros,
+                                     b_group_size, c, alpha, beta, schedule)
+
+
+def machete_prepack_B(b_q_weight: torch.Tensor,
+                      b_type: ScalarType) -> torch.Tensor:
+    return torch.ops._C.machete_prepack_B(b_q_weight, b_type.id)
+
+
+if hasattr(torch.ops._C, "permute_cols"):
+
+    @register_fake("_C::permute_cols")
+    def _permute_cols_fake(a: torch.Tensor,
+                           perm: torch.Tensor) -> torch.Tensor:
+        return torch.empty_like(a)
+
+
+def permute_cols(a: torch.Tensor, perm: torch.Tensor) -> torch.Tensor:
+    return torch.ops._C.permute_cols(a, perm)
+
+
+# fp8
+def scaled_fp8_quant(
+    input: torch.Tensor,
+    scale: Optional[torch.Tensor] = None,
+    num_token_padding: Optional[int] = None,
+    scale_ub: Optional[torch.Tensor] = None,
+    use_per_token_if_dynamic: bool = False,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Quantize input tensor to FP8 and return quantized tensor and scale.
+
+    This function supports both static and dynamic quantization: If you
+    provide the scale, it will use static scaling and if you omit it,
+    the scale will be determined dynamically. The function also allows
+    optional padding of the output tensors for downstream kernels that
+    will benefit from padding.
+
+    Args:
+        input: The input tensor to be quantized to FP8
+        scale: Optional scaling factor for the FP8 quantization
+        scale_ub: Optional upper bound for scaling factor in dynamic
+            per token case
+        num_token_padding: If specified, pad the first dimension
+            of the output to at least this value.
+        use_per_token_if_dynamic: Whether to do per_tensor or per_token
+            in the dynamic quantization case.
+
+    Returns:
+        Tuple[torch.Tensor, torch.Tensor]: The output tensor in FP8 and
+            scaling factor.
+    """
+    # This code assumes batch_dim and num_tokens are flattened
+    assert (input.ndim == 2)
+    shape: Union[Tuple[int, int], torch.Size] = input.shape
+    # For rocm, the output fp8 dtype is torch.float_e3m3fnuz
+    out_dtype: torch.dtype = torch.float8_e4m3fnuz \
+            if current_platform.is_rocm() else torch.float8_e4m3fn
+    if num_token_padding:
+        shape = (max(num_token_padding, input.shape[0]), shape[1])
+    output = torch.empty(shape, device=input.device, dtype=out_dtype)
+
+    if scale is None:
+        if use_per_token_if_dynamic:
+            scale = torch.empty((shape[0], 1),
+                                device=input.device,
+                                dtype=torch.float32)
+            torch.ops._C.dynamic_per_token_scaled_fp8_quant(
+                output, input, scale, scale_ub)
+        else:
+            scale = torch.zeros(1, device=input.device, dtype=torch.float32)
+            torch.ops._C.dynamic_scaled_fp8_quant(output, input, scale)
+    else:
+        # num_token_padding not implemented for this case
+        assert (scale.numel() == 1 or num_token_padding is None)
+        torch.ops._C.static_scaled_fp8_quant(output, input, scale)
+
+    return output, scale
+
+
+# int8
+def scaled_int8_quant(
+    input: torch.Tensor,
+    scale: Optional[torch.Tensor] = None,
+    azp: Optional[torch.Tensor] = None,
+    symmetric: bool = True
+) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]:
+    """
+    Quantize the input tensor to int8 and return the quantized tensor and scale, and maybe azp.
+
+    Args:
+        input: The input tensor to be quantized to int8.
+        scale: Optional scaling factor for the int8 quantization.
+            When not provided, we invoke dynamic-per-token quantization.
+        azp: Optional zero-point for the int8 quantization.
+            Must be provided for asymmetric quantization if `scale` is provided.
+        symmetric: Whether to use symmetric quantization (scale only, azp ignored).
+
+    Returns:
+      Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]] : Output int8 tensor, scales, and optionally azp.
+    """
+    output = torch.empty_like(input, dtype=torch.int8)
+    if scale is not None:
+        # static-per-tensor quantization.
+        assert symmetric == (
+            azp is
+            None), "azp must only be provided for asymmetric quantization."
+        torch.ops._C.static_scaled_int8_quant(output, input, scale, azp)
+        return output, scale, azp
+
+    # dynamic-per-token quantization.
+    input_scales = torch.empty((input.numel() // input.shape[-1], 1),
+                               device=input.device,
+                               dtype=torch.float32)
+    input_azp = None if symmetric else torch.empty_like(input_scales,
+                                                        dtype=torch.int32)
+    torch.ops._C.dynamic_scaled_int8_quant(output, input, input_scales,
+                                           input_azp)
+    return output, input_scales, input_azp
+
+
+# qqq ops
+def marlin_qqq_gemm(a: torch.Tensor, b_q_weight: torch.Tensor,
+                    s_tok: torch.Tensor, s_ch: torch.Tensor,
+                    s_group: torch.Tensor, workspace: torch.Tensor,
+                    size_m: int, size_n: int, size_k: int) -> torch.Tensor:
+    return torch.ops._C.marlin_qqq_gemm(a, b_q_weight, s_tok, s_ch, s_group,
+                                        workspace, size_m, size_n, size_k)
+
+
+# gguf
+def ggml_dequantize(W: torch.Tensor, quant_type: int, m: int,
+                    n: int) -> torch.Tensor:
+    return torch.ops._C.ggml_dequantize(W, quant_type, m, n)
+
+
+def ggml_mul_mat_vec_a8(
+    W: torch.Tensor,
+    X: torch.Tensor,
+    quant_type: int,
+    row: int,
+) -> torch.Tensor:
+    return torch.ops._C.ggml_mul_mat_vec_a8(W, X, quant_type, row)
+
+
+def ggml_mul_mat_a8(
+    W: torch.Tensor,
+    X: torch.Tensor,
+    quant_type: int,
+    row: int,
+) -> torch.Tensor:
+    return torch.ops._C.ggml_mul_mat_a8(W, X, quant_type, row)
+
+
+# mamba
+def causal_conv1d_fwd(x: torch.Tensor, weight: torch.Tensor,
+                      bias_: Optional[torch.Tensor],
+                      conv_states: Optional[torch.Tensor],
+                      query_start_loc: Optional[torch.Tensor],
+                      cache_indices: Optional[torch.Tensor],
+                      has_initial_state: Optional[torch.Tensor],
+                      silu_activation: bool, pad_slot_id: int):
+    torch.ops._C.causal_conv1d_fwd(x, weight, bias_, conv_states,
+                                   query_start_loc, cache_indices,
+                                   has_initial_state, silu_activation,
+                                   pad_slot_id)
+
+
+def causal_conv1d_update(x: torch.Tensor, conv_state: torch.Tensor,
+                         weight: torch.Tensor, bias_: Optional[torch.Tensor],
+                         silu_activation: bool,
+                         cache_seqlens: Optional[torch.Tensor],
+                         conv_state_indices: Optional[torch.Tensor],
+                         pad_slot_id: int):
+    torch.ops._C.causal_conv1d_update(x, conv_state, weight, bias_,
+                                      silu_activation, cache_seqlens,
+                                      conv_state_indices, pad_slot_id)
+
+
+def selective_scan_fwd(u: torch.Tensor, delta: torch.Tensor, A: torch.Tensor,
+                       B: torch.Tensor, C: torch.Tensor,
+                       D_: Optional[torch.Tensor], z_: Optional[torch.Tensor],
+                       delta_bias_: Optional[torch.Tensor],
+                       delta_softplus: bool,
+                       query_start_loc: Optional[torch.Tensor],
+                       cache_indices: Optional[torch.Tensor],
+                       has_initial_state: Optional[torch.Tensor],
+                       ssm_states: torch.Tensor, pad_slot_id: int):
+    torch.ops._C.selective_scan_fwd(u, delta, A, B, C, D_, z_, delta_bias_,
+                                    delta_softplus, query_start_loc,
+                                    cache_indices, has_initial_state,
+                                    ssm_states, pad_slot_id)
+
+
+# moe
+def moe_sum(input: torch.Tensor, output: torch.Tensor):
+    torch.ops._moe_C.moe_sum(input, output)
+
+
+def moe_align_block_size(topk_ids: torch.Tensor, num_experts: int,
+                         block_size: int, sorted_token_ids: torch.Tensor,
+                         experts_ids: torch.Tensor,
+                         num_tokens_post_pad: torch.Tensor) -> None:
+    torch.ops._moe_C.moe_align_block_size(topk_ids, num_experts, block_size,
+                                          sorted_token_ids, experts_ids,
+                                          num_tokens_post_pad)
+
+
+def topk_softmax(topk_weights: torch.Tensor, topk_ids: torch.Tensor,
+                 token_expert_indicies: torch.Tensor,
+                 gating_output: float) -> None:
+    torch.ops._moe_C.topk_softmax(topk_weights, topk_ids,
+                                  token_expert_indicies, gating_output)
+
+
+if supports_moe_ops and hasattr(torch.ops._moe_C, "marlin_gemm_moe"):
+
+    @register_fake("_moe_C::marlin_gemm_moe")
+    def marlin_gemm_moe_fake(a: torch.Tensor, b_q_weights: torch.Tensor,
+                             sorted_ids: torch.Tensor,
+                             topk_weights: torch.Tensor,
+                             topk_ids: torch.Tensor, b_scales: torch.Tensor,
+                             b_zero_points: torch.Tensor, g_idx: torch.Tensor,
+                             perm: torch.Tensor, workspace: torch.Tensor,
+                             b_q_type: ScalarType, size_m: torch.SymInt,
+                             size_n: torch.SymInt, size_k: torch.SymInt,
+                             is_k_full: bool, num_experts: int, topk: int,
+                             moe_block_size: int, replicate_input: bool,
+                             apply_weights: bool) -> torch.Tensor:
+        return torch.empty((size_m, topk, size_n),
+                           dtype=a.dtype,
+                           device=a.device)
+
+
+def reshape_and_cache(
+    key: torch.Tensor,
+    value: torch.Tensor,
+    key_cache: torch.Tensor,
+    value_cache: torch.Tensor,
+    slot_mapping: torch.Tensor,
+    kv_cache_dtype: str,
+    k_scale: float,
+    v_scale: float,
+) -> None:
+    torch.ops._C_cache_ops.reshape_and_cache(key, value, key_cache,
+                                             value_cache, slot_mapping,
+                                             kv_cache_dtype, k_scale, v_scale)
+
+
+def reshape_and_cache_flash(
+    key: torch.Tensor,
+    value: torch.Tensor,
+    key_cache: torch.Tensor,
+    value_cache: torch.Tensor,
+    slot_mapping: torch.Tensor,
+    kv_cache_dtype: str,
+    k_scale: float,
+    v_scale: float,
+) -> None:
+    torch.ops._C_cache_ops.reshape_and_cache_flash(key, value, key_cache,
+                                                   value_cache, slot_mapping,
+                                                   kv_cache_dtype, k_scale,
+                                                   v_scale)
+
+
+def copy_blocks(key_caches: List[torch.Tensor],
+                value_caches: List[torch.Tensor],
+                block_mapping: torch.Tensor) -> None:
+    torch.ops._C_cache_ops.copy_blocks(key_caches, value_caches, block_mapping)
+
+
+def swap_blocks(src: torch.Tensor, dst: torch.Tensor,
+                block_mapping: torch.Tensor) -> None:
+    torch.ops._C_cache_ops.swap_blocks(src, dst, block_mapping)
+
+
+def convert_fp8(output: torch.Tensor,
+                input: torch.Tensor,
+                scale: float = 1.0,
+                kv_dtype: str = "fp8") -> None:
+    torch.ops._C_cache_ops.convert_fp8(output, input, scale, kv_dtype)
+
+
+def get_device_attribute(attribute: int, device: int) -> int:
+    return torch.ops._C_cuda_utils.get_device_attribute(attribute, device)
+
+
+def get_max_shared_memory_per_block_device_attribute(device: int) -> int:
+    # ruff: noqa: E501
+    return torch.ops._C_cuda_utils.get_max_shared_memory_per_block_device_attribute(
+        device)
+
+
+# custom ar
+def init_custom_ar(ipc_tensors: List[torch.Tensor], rank_data: torch.Tensor,
+                   rank: int, full_nvlink: bool) -> int:
+    return torch.ops._C_custom_ar.init_custom_ar(ipc_tensors, rank_data, rank,
+                                                 full_nvlink)
+
+
+def all_reduce(fa: int, inp: torch.Tensor, out: torch.Tensor, reg_buffer: int,
+               reg_buffer_sz_bytes: int) -> None:
+    torch.ops._C_custom_ar.all_reduce(fa, inp, out, reg_buffer,
+                                      reg_buffer_sz_bytes)
+
+
+def dispose(fa: int) -> None:
+    torch.ops._C_custom_ar.dispose(fa)
+
+
+def meta_size() -> int:
+    return torch.ops._C_custom_ar.meta_size()
+
+
+def register_buffer(fa: int, ipc_tensors: List[int]) -> None:
+    return torch.ops._C_custom_ar.register_buffer(fa, ipc_tensors)
+
+
+def get_graph_buffer_ipc_meta(fa: int) -> Tuple[List[int], List[int]]:
+    return torch.ops._C_custom_ar.get_graph_buffer_ipc_meta(fa)
+
+
+def register_graph_buffers(fa: int, handles: List[List[int]],
+                           offsets: List[List[int]]) -> None:
+    torch.ops._C_custom_ar.register_graph_buffers(fa, handles, offsets)
+
+
+# temporary fix for https://github.com/vllm-project/vllm/issues/5456
+# TODO: remove this in v0.6.0
+names_and_values = globals()
+names_and_values_to_update = {}
+# prepare variables to avoid dict size change during iteration
+k, v, arg = None, None, None
+fn_type = type(lambda x: x)
+for k, v in names_and_values.items():
+    # find functions that are defined in this file and have torch.Tensor
+    # in their annotations. `arg == "torch.Tensor"` is used to handle
+    # the case when users use `import __annotations__` to turn type
+    # hints into strings.
+    if isinstance(v, fn_type) \
+        and v.__code__.co_filename == __file__ \
+        and any(arg is torch.Tensor or arg == "torch.Tensor"
+                for arg in v.__annotations__.values()):
+        names_and_values_to_update[k] = hint_on_error(v)
+
+names_and_values.update(names_and_values_to_update)
+del names_and_values_to_update, names_and_values, v, k, fn_type

vllm-v0.6.2/vllm/_ipex_ops.py

@@ -0,0 +1,226 @@
+from typing import List, Optional, Tuple
+
+import torch
+
+from vllm.logger import init_logger
+
+logger = init_logger(__name__)
+
+try:
+    import intel_extension_for_pytorch as ipex
+except ImportError as e:
+    logger.warning("Import error msg: %s", e.msg)
+
+
+class ipex_ops:
+
+    @staticmethod
+    def _reshape_activation_tensor(
+            x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+        num = x.size(0)
+        d = x.size(1) // 2
+        x = x.reshape(num, 2, d)
+        x1, x2 = torch.chunk(x, chunks=2, dim=1)
+        x1 = x1.reshape(num, d)
+        x2 = x2.reshape(num, d)
+        return x1, x2
+
+    @staticmethod
+    def silu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
+        ipex.llm.functional.silu_and_mul(x, out)
+
+    @staticmethod
+    def gelu_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
+        ipex.llm.functional.gelu_and_mul(x, out)
+
+    @staticmethod
+    def gelu_tanh_and_mul(out: torch.Tensor, x: torch.Tensor) -> None:
+        ipex.llm.functional.gelu_and_mul(x, out)
+
+    @staticmethod
+    def gelu_fast(x: torch.Tensor) -> torch.Tensor:
+        return torch.nn.functional.gelu(x)
+
+    @staticmethod
+    def gelu_new(x: torch.Tensor) -> torch.Tensor:
+        return torch.nn.functional.gelu(x)
+
+    @staticmethod
+    def gelu_quick(out: torch.Tensor, x: torch.Tensor) -> None:
+        ipex.llm.functional.gelu_quick(x, out)
+
+    @staticmethod
+    def paged_attention_v1(
+        out: torch.Tensor,
+        query: torch.Tensor,
+        key_cache: torch.Tensor,
+        value_cache: torch.Tensor,
+        num_kv_heads: int,
+        scale: float,
+        block_tables: torch.Tensor,
+        context_lens: torch.Tensor,
+        block_size: int,
+        max_context_len: int,
+        alibi_slopes: Optional[torch.Tensor],
+        kv_cache_dtype: str,
+        k_scale: float,
+        v_scale: float,
+        tp_rank: int = 0,
+        blocksparse_local_blocks: int = 0,
+        blocksparse_vert_stride: int = 0,
+        blocksparse_block_size: int = 64,
+        blocksparse_head_sliding_step: int = 0,
+    ) -> None:
+        assert kv_cache_dtype == "auto"
+        num_heads = out.size(1)
+        num_queries_per_tokens = num_heads // num_kv_heads
+        ipex.llm.modules.PagedAttention.single_query_kv_attention(
+            out,
+            query.contiguous(),
+            key_cache.view_as(value_cache),
+            value_cache,
+            num_queries_per_tokens,
+            scale,
+            block_tables,
+            context_lens,
+            block_size,
+            max_context_len,
+            alibi_slopes,
+        )
+
+    @staticmethod
+    def paged_attention_v2(
+        out: torch.Tensor,
+        exp_sum: torch.Tensor,
+        max_logits: torch.Tensor,
+        tmp_out: torch.Tensor,
+        query: torch.Tensor,
+        key_cache: torch.Tensor,
+        value_cache: torch.Tensor,
+        num_kv_heads: int,
+        scale: float,
+        block_tables: torch.Tensor,
+        context_lens: torch.Tensor,
+        block_size: int,
+        max_context_len: int,
+        alibi_slopes: Optional[torch.Tensor],
+        kv_cache_dtype: str,
+        k_scale: float,
+        v_scale: float,
+        tp_rank: int = 0,
+        blocksparse_local_blocks: int = 0,
+        blocksparse_vert_stride: int = 0,
+        blocksparse_block_size: int = 64,
+        blocksparse_head_sliding_step: int = 0,
+    ) -> None:
+        assert kv_cache_dtype == "auto"
+        num_heads = out.size(1)
+        num_queries_per_tokens = num_heads // num_kv_heads
+        ipex.llm.modules.PagedAttention.single_query_kv_attention(
+            out,
+            query.contiguous(),
+            key_cache.view_as(value_cache),
+            value_cache,
+            num_queries_per_tokens,
+            scale,
+            block_tables,
+            context_lens,
+            block_size,
+            max_context_len,
+            alibi_slopes,
+        )
+
+    @staticmethod
+    def rotary_embedding(
+        positions: torch.Tensor,  # [batch_size, seq_len]
+        query: torch.Tensor,  # [batch_size, seq_len, num_heads*head_size]
+        key: torch.Tensor,  # [batch_size, seq_len, num_kv_heads*head_size]
+        head_size: int,
+        cos_sin_cache: torch.Tensor,  # [cos_sin_dim, rot_dim]
+        is_neox: bool,
+    ) -> None:
+        rot_dim = cos_sin_cache.size(1)
+        ipex.llm.functional.rotary_embedding_batched(positions, query, key,
+                                                     head_size, cos_sin_cache,
+                                                     is_neox, rot_dim)
+
+    @staticmethod
+    def batched_rotary_embedding(positions: torch.Tensor, query: torch.Tensor,
+                                 key: torch.Tensor, head_size: int,
+                                 cos_sin_cache: torch.Tensor, is_neox: bool,
+                                 rot_dim: int,
+                                 cos_sin_cache_offsets: torch.Tensor) -> None:
+        ipex.llm.functional.rotary_embedding_batched(positions, query, key,
+                                                     head_size, cos_sin_cache,
+                                                     is_neox, rot_dim,
+                                                     cos_sin_cache_offsets)
+
+    @staticmethod
+    def rms_norm(input: torch.Tensor, weight: torch.Tensor,
+                 epsilon: float) -> torch.Tensor:
+        return ipex.llm.functional.rms_norm(input, weight, epsilon)
+
+    @staticmethod
+    def fused_add_rms_norm(input: torch.Tensor, residual: torch.Tensor,
+                           weight: torch.Tensor, epsilon: float) -> None:
+        tmp = ipex.llm.functional.add_rms_norm(residual, input, weight, None,
+                                               epsilon, True)
+        input.copy_(tmp)
+
+    @staticmethod
+    def varlen_attention(
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        out: torch.Tensor,
+        seqlen_q: torch.Tensor,
+        seqlen_k: torch.Tensor,
+        max_seqlen_q: int,
+        max_seqlen_k: int,
+        pdropout: float,
+        softmax_scale: float,
+        zero_tensors: bool,
+        is_causal: bool,
+        return_softmax: bool,
+        gen_: torch.Generator,
+        logits_soft_cap: float,
+    ) -> None:
+        ipex.llm.functional.varlen_attention(query.contiguous(),
+                                             key.contiguous(),
+                                             value.contiguous(), out,
+                                             seqlen_q.int(), seqlen_k.int(),
+                                             max_seqlen_q, max_seqlen_k,
+                                             pdropout, softmax_scale,
+                                             zero_tensors, is_causal,
+                                             return_softmax, gen_,
+                                             logits_soft_cap)
+
+    @staticmethod
+    def reshape_and_cache(
+        key: torch.Tensor,
+        value: torch.Tensor,
+        key_cache: torch.Tensor,
+        value_cache: torch.Tensor,
+        slot_mapping: torch.Tensor,
+        kv_cache_dtype: str,
+        k_scale: float,
+        v_scale: float,
+    ) -> None:
+        assert kv_cache_dtype == "auto"
+        ipex.llm.modules.PagedAttention.reshape_and_cache(
+            key, value, key_cache, value_cache, slot_mapping)
+
+    @staticmethod
+    def copy_blocks(key_caches: List[torch.Tensor],
+                    value_caches: List[torch.Tensor],
+                    block_mapping: torch.Tensor) -> None:
+        torch.xpu.copy_blocks(  # type: ignore
+            key_caches,
+            value_caches,
+            block_mapping,
+        )
+
+    @staticmethod
+    def swap_blocks(src: torch.Tensor, dst: torch.Tensor,
+                    block_mapping: torch.Tensor) -> None:
+        torch.xpu.swap_blocks(src, dst, block_mapping)  # type: ignore

vllm-v0.6.2/vllm/_mlu_ops.py

@@ -0,0 +1,778 @@
+from typing import List, Optional, Tuple
+
+import torch
+
+import math
+import triton
+import triton.language as tl
+
+from vllm.logger import init_logger
+
+logger = init_logger(__name__)
+
+try:
+    import torch_mlu_ops as tmo
+except ImportError as e:
+    logger.warning("Failed to import from TMO OPS with %r", e)
+
+
+def rotary_embedding(
+    input: torch.Tensor,
+    sin_cache: torch.Tensor,
+    cos_cache: torch.Tensor,
+    position_ids: Optional[torch.Tensor],
+    cu_seqlens: Optional[torch.Tensor],
+    interleaved: bool,
+    discrete: bool,
+    dynamic_ntk: bool,
+    max_seqlen: int,
+) -> torch.Tensor:
+    return tmo.apply_rotary(
+                input, sin_cache, cos_cache,
+                position_ids, cu_seqlens, interleaved,
+                discrete, dynamic_ntk, max_seqlen)
+
+
+def fused_rms_norm(
+    x: torch.Tensor,
+    residual: torch.Tensor,
+    gamma: torch.Tensor,
+    beta: torch.Tensor,
+    bias: torch.Tensor,
+    eps: float,
+    store_output_before_norm: bool,
+    quant_scale: torch.Tensor = None,
+    dynamic_quant: bool = False,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+    return tmo.fused_rms_norm(
+                x, residual, gamma, beta, bias,
+                eps, store_output_before_norm, quant_scale,
+                None, dynamic_quant)
+
+
+def fused_layer_norm(
+    x: torch.Tensor,
+    residual: torch.Tensor,
+    gamma: torch.Tensor,
+    beta: torch.Tensor,
+    bias: torch.Tensor,
+    eps: float,
+    store_output_before_norm: bool,
+    quant_scale: torch.Tensor = None,
+    dynamic_quant: bool = False,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[torch.Tensor]]:
+    return tmo.fused_layer_norm(
+                x, residual, gamma, beta, bias,
+                eps, store_output_before_norm, quant_scale,
+                None, dynamic_quant)
+
+
+def flash_attention(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    out: torch.Tensor,
+    cu_seq_lens_q: torch.Tensor,
+    cu_seq_lens_kv: torch.Tensor,
+    alibi_slope: torch.Tensor,
+    attn_bias: torch.Tensor,
+    max_seq_len_q: int,
+    max_seq_len_kv: int,
+    softmax_scale: float,
+    is_causal: bool,
+    window_size_left: int = -1,
+    window_size_right: int = -1,
+    compute_dtype: torch.dtype = torch.float,
+    return_lse: bool = False,
+    block_tables: torch.Tensor = None,
+    k_cache_quant_scale: torch.Tensor = None,
+    v_cache_quant_scale: torch.Tensor = None
+) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+    return tmo.flash_attention(
+        q, k, v, out,
+        cu_seq_lens_q, cu_seq_lens_kv,
+        alibi_slope, attn_bias,
+        max_seq_len_q, max_seq_len_kv,
+        softmax_scale, is_causal,
+        window_size_left, window_size_right,
+        compute_dtype, return_lse,
+        block_tables, k_cache_quant_scale,
+        v_cache_quant_scale)
+
+
+def split_head_nums(q_head_num, kv_head_num, max_q_head_num):
+    """
+    对 q_head_num 进行切分，使得：
+    1. 切分后的 q_head_num 最大值不超过 max_q_head_num
+    2. kv_head_num 按 q_head_num 相同份数拆分；
+    3. 每个切分后的 q_head_num 可以被对应的 kv_head_num 整除；
+    4. 若 kv_head_num < 1，则调整为 1。
+
+    参数：
+    - q_head_num: int, 需要切分的 q_head_num。
+    - kv_head_num: int, 需要切分的 kv_head_num。
+    - max_q_head_num: int, 支持切分后最大的 q_head_num
+
+    返回：
+    - q_splits: list, 切分后的 q_head_num。
+    - kv_splits: list, 切分后的 kv_head_num。
+    """
+    if q_head_num <= 0 or kv_head_num <= 0:
+        return "q_head_num 和 kv_head_num 必须是正整数！"
+
+    q_splits = []
+    kv_splits = []
+
+    # 剩余值
+    remaining_q = q_head_num
+    remaining_kv = kv_head_num
+
+    while remaining_q > 0:
+        # 尝试切分 q_head_num，最大值不超过 max_q_head_num
+        for q_part in range(min(max_q_head_num, remaining_q), 0, -1):
+            # 确保 q_part 能被分配并且对应的 kv_part >= 1
+            if remaining_q % q_part == 0:
+                kv_part = max(remaining_kv // (remaining_q // q_part), 1)  # 确保 kv_part >= 1
+                if q_part % kv_part == 0:  # 确保 q_part 可以被 kv_part 整除
+                    # 记录切分值
+                    q_splits.append(q_part)
+                    kv_splits.append(kv_part)
+                    remaining_q -= q_part
+                    remaining_kv -= kv_part
+                    break
+        else:
+            err_msg = f"Unable to find split method for q_head_num:{q_head_num}, kv_head_num:{kv_head_num}"
+            raise RuntimeError(err_msg)
+
+    return q_splits, kv_splits
+
+
+def repeat_elements(input_list, n):
+    """
+    将列表的每个成员连续重复 n 次。
+
+    参数：
+    - input_list: list，输入的列表。
+    - n: int，每个元素需要重复的次数。
+
+    返回：
+    - list，包含重复元素的新列表。
+    """
+    if not isinstance(input_list, list) or not isinstance(n, int) or n < 0:
+        raise ValueError("输入必须是一个列表，并且重复次数 n 必须是大于或等于 0 的整数。")
+    
+    # 使用列表推导式重复每个元素 n 次
+    return [item for item in input_list for _ in range(n)]
+
+
+def single_query_cached_kv_attn(
+    q: torch.Tensor,
+    k_cache: torch.Tensor,
+    v_cache: torch.Tensor,
+    out: torch.Tensor,
+    block_tables: torch.Tensor,
+    context_lens: torch.Tensor,
+    k_cache_quant_scale: Optional[torch.Tensor],
+    v_cache_quant_scale: Optional[torch.Tensor],
+    alibi_slopes: Optional[torch.Tensor],
+    max_contxt_len: int,
+    windows_size_left: int,
+    windows_size_right: int,
+    softmax_scale: float,
+    q_head_dim: Optional[int] = 2,
+    kv_head_dim: Optional[int] = 1,
+    seq_q_dim: Optional[int] = 1,
+    max_seq_q_mul_q_divide_kv: Optional[int] = 48,
+) -> None:
+    # FIXME(chenxiaobing): TMO only support windows_size_right = -1 yet.
+    windows_size_right = -1
+
+    # singleQwithkvCache limits seq_q * q_divide_kv <= max_seq_q_mul_q_divide_kv now.
+    # When the limitation is fixed, we should delete the split process.
+    seq_q = q.shape[seq_q_dim]
+    q_head_num = q.shape[q_head_dim]
+    kv_head_num = k_cache.shape[kv_head_dim]
+    q_divide_kv = q_head_num // kv_head_num
+    if seq_q * q_divide_kv <= max_seq_q_mul_q_divide_kv:
+        tmo.single_query_cached_kv_attn(
+            q, k_cache, v_cache, out,
+            block_tables, context_lens,
+            k_cache_quant_scale, v_cache_quant_scale,
+            alibi_slopes, max_contxt_len,
+            windows_size_left, windows_size_right, softmax_scale)
+    else:
+        max_q_head_num = max_seq_q_mul_q_divide_kv * kv_head_num // seq_q
+        q_head_num_sizes, kv_head_num_sizes = split_head_nums(q_head_num, kv_head_num, max_q_head_num)
+        parts_num = len(q_head_num_sizes)
+        q_parts = torch.split(q, q_head_num_sizes, dim=q_head_dim)
+        out_parts = torch.split(out, q_head_num_sizes, dim=q_head_dim)
+        alibi_slopes_parts = [None] * parts_num
+        if alibi_slopes:
+            alibi_slopes_parts = torch.split(alibi_slopes, q_head_num_sizes, dim=0)
+
+        kv_parts_num = parts_num
+        if parts_num > kv_head_num:
+            assert parts_num % kv_head_num == 0, f"parts_num:{parts_num} need by divided by kv_head_num:{kv_head_num} when parts_num > kv_head_num"
+            kv_parts_num = kv_head_num
+            kv_head_num_sizes = kv_head_num_sizes[:kv_parts_num]
+
+        if len(kv_head_num_sizes) > 1:
+            k_cache_parts = torch.split(k_cache, kv_head_num_sizes, dim=kv_head_dim)
+            v_cache_parts = torch.split(v_cache, kv_head_num_sizes, dim=kv_head_dim)
+            k_cache_quant_scale_parts = [None] * kv_parts_num
+            v_cache_quant_scale_parts = [None] * kv_parts_num
+            if k_cache_quant_scale:
+                k_cache_quant_scale_dim = 1 if k_cache_quant_scale.dim() == 2 else kv_head_dim
+                k_cache_quant_scale_parts = torch.split(k_cache_quant_scale, kv_head_num_sizes, dim=k_cache_quant_scale_dim)
+            if v_cache_quant_scale:
+                v_cache_quant_scale_dim = 1 if v_cache_quant_scale.dim() == 2 else kv_head_dim
+                v_cache_quant_scale_parts = torch.split(v_cache_quant_scale, kv_head_num_sizes, dim=v_cache_quant_scale_dim)
+        else:
+            k_cache_parts = [k_cache]
+            v_cache_parts = [v_cache]
+            k_cache_quant_scale_parts = [k_cache_quant_scale]
+            v_cache_quant_scale_parts = [v_cache_quant_scale]
+
+        if parts_num > kv_parts_num:
+            repeate_num = parts_num // kv_parts_num
+            k_cache_parts = repeat_elements(k_cache_parts, repeate_num)
+            v_cache_parts = repeat_elements(v_cache_parts, repeate_num)
+            k_cache_quant_scale_parts = repeat_elements(k_cache_quant_scale_parts, repeate_num)
+            v_cache_quant_scale_parts = repeat_elements(v_cache_quant_scale_parts, repeate_num)
+
+        for q_value, k_cache_value, v_cache_value, out_value, k_cache_quant_scale_value, v_cache_quant_scale_value, alibi_slopes_value in zip(
+                q_parts, k_cache_parts, v_cache_parts, out_parts, k_cache_quant_scale_parts, v_cache_quant_scale_parts,
+                alibi_slopes_parts):
+            tmo.single_query_cached_kv_attn(
+                q_value, k_cache_value.contiguous(), v_cache_value.contiguous(), out_value,
+                block_tables, context_lens,
+                k_cache_quant_scale_value, v_cache_quant_scale_value,
+                alibi_slopes_value, max_contxt_len,
+                windows_size_left, windows_size_right, softmax_scale)
+
+
+def reshape_linear_cache(
+    key: torch.Tensor,
+    value: Optional[torch.Tensor],
+    key_cache: torch.Tensor,
+    value_cache: Optional[torch.Tensor],
+    context_lengths: torch.Tensor,
+    max_context_len: int,
+    packed: bool,
+    context_seq_offset: Optional[torch.Tensor],
+    cache_bs_id: Optional[torch.Tensor],
+    cache_seqlen_offset: Optional[torch.Tensor],
+) -> None:
+    tmo.reshape_linear_cache(
+        key, value,
+        key_cache, value_cache,
+        context_lengths, max_context_len,
+        packed, context_seq_offset,
+        cache_bs_id, cache_seqlen_offset)
+
+
+def reshape_paged_cache(
+    k: torch.Tensor,
+    v: torch.Tensor,
+    k_cache: torch.Tensor,
+    v_cache: torch.Tensor,
+    slot_mapping: torch.Tensor
+) -> None:
+    tmo.reshape_paged_cache(k, v, k_cache, v_cache, slot_mapping)
+
+
+def swap_blocks(
+    dst: torch.Tensor,
+    src: torch.Tensor,
+    block_mapping: torch.Tensor
+) -> None:
+    # FIXME: Remove this conversion after
+    # tmo.swap_blocks support block_mapping tensor.
+    block_mapping = block_mapping.tolist()
+    block_mapping = {src: dst for src, dst in block_mapping}
+    return tmo.swap_blocks(dst, src, block_mapping)
+
+
+def copy_blocks(
+    k_caches: List[torch.Tensor],
+    v_caches: List[torch.Tensor],
+    block_mapping: torch.Tensor
+) -> None:
+    # FIXME: Remove this conversion after
+    # tmo.swap_blocks support block_mapping tensor.
+    block_mapping = block_mapping.tolist()
+    result_dict = {}
+    for row in block_mapping:
+        key = row[0]
+        values = row[1:]
+        if key in result_dict:
+            result_dict[key].extend(values)
+        else:
+            result_dict[key] = values
+    return tmo.copy_blocks(k_caches, v_caches, result_dict)
+
+
+def ffn(
+    input: torch.Tensor,
+    up_fc_weight: torch.Tensor,
+    up_fc_bias: Optional[torch.Tensor],
+    down_proj_weight: torch.Tensor,
+    down_proj_bias: Optional[torch.Tensor],
+    gate_up_proj_weight: Optional[torch.Tensor] = None,
+    gate_up_proj_bias: Optional[torch.Tensor] = None,
+    act_mode: str = "none"
+) -> torch.Tensor:
+    return tmo.ffn(input, up_fc_weight, up_fc_bias, down_proj_weight, down_proj_bias,
+                   gate_up_proj_weight, gate_up_proj_bias, act_mode)
+
+
+def active(
+    input: torch.Tensor,
+    act_mode: str,
+    is_gated: bool
+) -> torch.Tensor:
+    return tmo.active(input, act_mode, is_gated)
+
+
+def fused_moe(
+    hidden_states: torch.Tensor,
+    gating_output: torch.Tensor,
+    w1: torch.Tensor,
+    w2: torch.Tensor,
+    bias1: Optional[torch.Tensor],
+    bias2: Optional[torch.Tensor],
+    residual: Optional[torch.Tensor],
+    input_smooth: Optional[torch.Tensor],
+    act_smooth: Optional[torch.Tensor],
+    w1_scale: Optional[torch.Tensor],
+    w2_scale: Optional[torch.Tensor],
+    topk: int,
+    renormalize: bool,
+    gated: bool,
+    act_mode: str,
+    start_expert_id: int = 0,
+    block_n: int = 0,
+    cncl_comm: int = 0
+) -> torch.Tensor:
+    return tmo.fused_moe(
+        hidden_states, gating_output,
+        w1, w2, bias1, bias2, residual,
+        input_smooth, act_smooth,
+        w1_scale, w2_scale, topk,
+        renormalize, gated, act_mode, start_expert_id,
+        block_n, cncl_comm)
+
+
+def matmul(
+    a: torch.Tensor,
+    b: torch.Tensor,
+    bias: Optional[torch.Tensor] = None,
+    c: Optional[torch.Tensor] = None,
+    act_mode: str = 'none',
+    alpha: float = 1.0,
+    beta: float = .0
+) -> torch.Tensor:
+    return tmo.matmul(a, b, bias, c, act_mode, alpha, beta)
+
+
+def weight_only_quant_matmul(
+    a: torch.Tensor,
+    b: torch.Tensor,
+    scale: torch.Tensor,
+    zero: torch.Tensor = None,
+    bias: torch.Tensor = None,
+    c: torch.Tensor = None,
+    act_mode: str = "none",
+    quant_bit_size: int = 8,
+    alpha: float = 1.0,
+    beta: float = 1.0
+) -> torch.Tensor:
+    return tmo.weight_only_quant_matmul(
+                a, b,
+                scale, zero, bias, c,
+                act_mode, quant_bit_size, alpha, beta)
+
+
+def smooth_quant_matmul(
+    a: torch.Tensor,
+    a_scale: torch.Tensor,
+    b: torch.Tensor,
+    b_scale: torch.Tensor,
+    dtype: torch.dtype,
+    bias: torch.Tensor = None,
+    c: torch.Tensor = None,
+    act_mode: str = "none",
+    alpha: float = 1.0,
+    beta: float = 1.0
+) -> torch.Tensor:
+    return tmo.smooth_quant_matmul(
+                a, a_scale,
+                b, b_scale,
+                dtype, bias, c,
+                act_mode, alpha, beta)
+
+
+def per_token_smooth_quantize(
+    x: torch.Tensor,
+    smooth: torch.Tensor,
+    zero: torch.Tensor = None,
+    token_count: torch.Tensor = None
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    return tmo.per_token_smooth_quantize(x, smooth, zero, token_count)
+
+
+def quantize(
+    x: torch.Tensor,
+    scale: torch.Tensor,
+    zero: torch.Tensor = None
+) -> torch.Tensor:
+    return tmo.quantize(x, scale, zero)
+
+def quant_to_paged_cache(
+    k: torch.Tensor,
+    v: torch.Tensor,
+    k_cache: torch.Tensor,
+    v_cache: torch.Tensor,
+    k_cache_quant_scale: torch.Tensor,
+    v_cache_quant_scale: torch.Tensor,
+    slot_mapping: torch.Tensor,
+) -> None:
+    return tmo.quant_to_paged_cache(
+        k, v, k_cache, v_cache, k_cache_quant_scale, v_cache_quant_scale, slot_mapping
+    )
+
+
+def quant_to_linear_cache(
+    key: torch.Tensor,
+    value: Optional[torch.Tensor],
+    key_cache: torch.Tensor,
+    value_cache: Optional[torch.Tensor],
+    key_cache_quant_scale: torch.Tensor,
+    value_cache_quant_scale: Optional[torch.Tensor],
+    context_lengths: torch.Tensor,
+    max_context_len: int,
+    packed: bool,
+    context_seq_offset: Optional[torch.Tensor],
+    cache_bs_id: Optional[torch.Tensor],
+    cache_seqlen_offset: Optional[torch.Tensor],
+) -> None:
+    return tmo.quant_to_linear_cache(
+        key,
+        value,
+        key_cache,
+        value_cache,
+        key_cache_quant_scale,
+        value_cache_quant_scale,
+        context_lengths,
+        max_context_len,
+        packed,
+        context_seq_offset,
+        cache_bs_id,
+        cache_seqlen_offset,
+    )
+
+
+def advance_step(num_seqs: int,
+                 num_queries: int,
+                 block_size: int,
+                 input_tokens: torch.Tensor,
+                 sampled_token_ids: torch.Tensor,
+                 input_positions: torch.Tensor,
+                 seq_lens: torch.Tensor,
+                 slot_mapping: torch.Tensor,
+                 block_tables: torch.Tensor,
+                 TILE_SIZE: int = 64) -> None:
+    """
+    Advance a step on MLU for existing inputs for a multi-step runner, which
+    will update input_tokens/seq_lens/input_positions/slot_mapping inplace.
+    """
+    def verify_tensor(
+        name: str,
+        tensor: torch.Tensor,
+        size_0: int,
+        size_1: int,
+        dtype: torch.dtype,
+    ):
+        """
+        Auxiliary function to check whether input is valid.
+        """
+        size_0_cond = (size_0 == -1 or tensor.size(0) == size_0)
+        size_1_cond = (size_1 == -1 or tensor.size(1) == size_1)
+        if not (size_0_cond and size_1_cond and tensor.is_contiguous and tensor.dtype == dtype):
+            raise ValueError(
+                f"The input to advance_step is invalid with tensor name = {name}, "
+                f"shape = {tensor.shape}, "
+                f"is_cont = {tensor.is_contiguous()}, "
+                f"type = {tensor.dtype}, "
+                f"is not as expected: shape[{size_0}, {size_1}], type = {dtype}"
+            )
+
+
+    @triton.jit
+    def _triton_advance_step(input_tokens_ptr,
+                             sampled_token_ids_ptr,
+                             input_positions_ptr,
+                             seq_lens_ptr,
+                             slot_mapping_ptr,
+                             block_tables_ptr,
+                             block_tables_stride,
+                             num_seqs,
+                             num_queries,
+                             block_size,
+                             TILE_SIZE: tl.constexpr,
+    ):
+        """
+        The triton implementation of advance step.
+        Reference: https://github.com/vllm-project/vllm/blob/v0.6.1/csrc/prepare_inputs/advance_step.cu#L14-L55
+        """
+        # Set meta info.
+        pid = tl.program_id(axis=0)
+        offsets = pid * TILE_SIZE + tl.arange(0, TILE_SIZE)
+        mask = offsets < num_queries
+
+        # Update input_tokens.
+        sampled_token_ids = tl.load(sampled_token_ids_ptr + offsets, mask=mask)
+        tl.store(input_tokens_ptr + offsets, sampled_token_ids, mask=mask)
+
+        seq_lens = tl.load(seq_lens_ptr + offsets, mask=mask)
+        next_seq_lens = seq_lens + 1
+        next_input_pos = next_seq_lens - 1
+
+        # Update seq_lens.
+        tl.store(seq_lens_ptr + offsets, next_seq_lens, mask=mask)
+
+        # Update input_positions.
+        tl.store(input_positions_ptr + offsets, next_input_pos, mask=mask)
+
+        # Calculate slot num.
+        block_index = next_input_pos // block_size
+        block_offset = next_input_pos % block_size
+        block_tables = tl.load(block_tables_ptr + block_tables_stride * offsets + block_index, mask=mask)
+        slot_num = block_tables * block_size + block_offset
+
+        # Update slot_mapping.
+        tl.store(slot_mapping_ptr + offsets, slot_num, mask=mask)
+
+
+    verify_tensor("input_tokens", input_tokens, num_seqs, -1, torch.int64)
+    verify_tensor("sampled_token_ids", sampled_token_ids, num_queries, 1, torch.int64)
+    verify_tensor("input_positions", input_positions, num_seqs, -1, torch.int32)
+    verify_tensor("seq_lens", seq_lens, num_seqs, -1, torch.int32)
+    verify_tensor("slot_mapping", slot_mapping, num_seqs, -1, torch.int32)
+    verify_tensor("block_tables", block_tables, num_seqs, -1, torch.int32)
+
+    grid = (math.ceil(num_queries / TILE_SIZE), )
+    _triton_advance_step[grid](input_tokens,
+                               sampled_token_ids,
+                               input_positions,
+                               seq_lens,
+                               slot_mapping,
+                               block_tables,
+                               block_tables.stride(0),
+                               num_seqs,
+                               num_queries,
+                               block_size,
+                               TILE_SIZE)
+
+def preload(
+    weight: torch.Tensor,
+    size: int
+) -> None:
+    """
+    Preload weights of layer.
+
+    Args:
+        weight (torch.Tensor): Weight to preload。
+        size (int): Preload size (byte)。
+
+    Returns:
+        None
+    """
+    return tmo.preload(weight, size)
+
+
+def matmul_allreduce(
+    cncl_comm,
+    a: torch.Tensor,
+    b: torch.Tensor,
+    bias: Optional[torch.Tensor] = None,
+    c: Optional[torch.Tensor] = None,
+    alpha: float = 1.0,
+    beta: float = .0,
+    block_m: int = 0
+) -> torch.Tensor:
+    return tmo.matmul_allreduce(cncl_comm=cncl_comm,
+                                a=a, b=b,
+                                bias=bias, c=c,
+                                alpha=alpha,
+                                beta=beta,
+                                block_m=block_m)
+
+
+def smooth_quant_matmul_allreduce(
+    cncl_comm,
+    a: torch.Tensor,
+    a_scale: torch.Tensor,
+    b: torch.Tensor,
+    b_scale: torch.Tensor,
+    dtype: torch.dtype,
+    bias: torch.Tensor = None,
+    c: torch.Tensor = None,
+    alpha: float = 1.0,
+    beta: float = 1.0,
+    block_m: int = 0):
+    return tmo.smooth_quant_matmul_allreduce(
+                cncl_comm=cncl_comm,
+                a=a, a_scale=a_scale,
+                b=b, b_scale=b_scale,
+                dtype=dtype, bias=bias, c=c,
+                alpha=alpha, beta=beta, block_m=block_m)
+
+
+def quant_matmul_allreduce(
+    cncl_comm,
+    a_tensor: torch.Tensor,
+    a_scale: Optional[torch.Tensor],
+    a_zero: Optional[torch.Tensor],
+    b_tensor: torch.Tensor,
+    b_scale: Optional[torch.Tensor],
+    b_zero: Optional[torch.Tensor],
+    bias: Optional[torch.Tensor],
+    c_tensor: Optional[torch.Tensor],
+    c_scale: Optional[torch.Tensor],
+    c_zero: Optional[torch.Tensor],
+    gemm_output_scale: Optional[torch.Tensor],
+    gemm_output_zero: Optional[torch.Tensor],
+    data_type: Optional[str],
+    quant_algo: str,
+    a_quant_layout: str,
+    b_quant_layout: str,
+    quant_bit_size: int = 8,
+    alpha: float = 1.0,
+    beta: float = 1.0,
+    trans_a: bool = False,
+    trans_b: bool = True,
+    block_m: int = 0
+) -> torch.Tensor:
+    return tmo.quant_matmul_allreduce(
+        cncl_comm=cncl_comm, a_tensor=a_tensor, a_scale=a_scale, a_zero=a_zero,
+        b_tensor=b_tensor, b_scale=b_scale, b_zero=b_zero, bias=bias,
+        c_tensor=c_tensor, c_scale=c_scale, c_zero=c_zero,
+        gemm_output_scale=gemm_output_scale, gemm_output_zero=gemm_output_zero,
+        data_type=data_type, quant_algo=quant_algo,
+        a_quant_layout=a_quant_layout, b_quant_layout=b_quant_layout,
+        quant_bit_size=quant_bit_size,
+        alpha=alpha, beta=beta, trans_a=trans_a, trans_b=trans_b, block_m=block_m)
+
+
+def flash_attn_sq_mm_allreduce(
+    cncl_comm: int,
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    cu_seq_lens_q: Optional[torch.Tensor],
+    cu_seq_lens_kv: Optional[torch.Tensor],
+    alibi_slope: Optional[torch.Tensor],
+    attn_bias: Optional[torch.Tensor],
+    smooth: torch.Tensor,
+    weight: torch.Tensor,
+    weight_scale: torch.Tensor,
+    bias: Optional[torch.Tensor],
+    max_seq_len_q: int,
+    max_seq_len_kv: int,
+    softmax_scale: float,
+    is_causal: bool,
+    window_size_left: int = -1,
+    window_size_right: int = -1,
+    compute_dtype: torch.dtype = torch.float,
+    block_seq: int = 0) -> torch.Tensor:
+    return tmo.flash_attn_sq_mm_allreduce(cncl_comm, q, k, v,
+                                cu_seq_lens_q, cu_seq_lens_kv, alibi_slope, attn_bias, smooth, weight, weight_scale,
+                                bias, max_seq_len_q, max_seq_len_kv, softmax_scale, is_causal, window_size_left,
+                                window_size_right, compute_dtype, block_seq)
+
+#Moe inner kernels
+def moe_softmax_topk(input: torch.Tensor,
+                     topk: int,
+                     normalize: bool = False,
+       num_expert_group: int = -1,
+       topk_group: int = 0) -> Tuple[torch.Tensor]:
+    return tmo.moe_softmax_topk(input, topk, normalize, num_expert_group, topk_group)
+
+
+def moe_gen_idx(expert_id: torch.Tensor,
+                expert_num: int) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+    return tmo.moe_gen_idx(expert_id, expert_num)
+
+def moe_expand_input(input: torch.Tensor,
+                     gather_idx: torch.Tensor,
+                     cusum_token_count: Optional[torch.Tensor] = None,
+                     start_expert_id: int = 0,
+                     expert_size: int = 0) -> torch.Tensor:
+    return tmo.moe_expand_input(input, gather_idx,
+                                cusum_token_count,
+                                start_expert_id, expert_size)
+
+def moe_active(input: torch.Tensor,
+               act_mode: str,
+               is_gated: bool,
+               output: Optional[torch.Tensor] = None,
+               bias: Optional[torch.Tensor] = None,
+               cusum_token_count: Optional[torch.Tensor] = None,
+               start_expert_id: int = 0,
+               expert_size: int = 0) -> torch.Tensor:
+    return tmo.moe_active(input, act_mode, is_gated, output,
+                          bias, cusum_token_count,
+                          start_expert_id, expert_size)
+
+def group_gemm(a: torch.Tensor,
+               b: torch.Tensor,
+               m_list: torch.Tensor,
+               expand_idx: Optional[torch.Tensor],
+               c: Optional[torch.Tensor],
+               alpha: Optional[torch.Tensor],
+               beta: Optional[torch.Tensor],
+               max_m: int = 0
+               ) -> torch.Tensor:
+    return tmo.group_gemm(a, b, m_list, expand_idx,
+                              c, alpha, beta, max_m)
+
+def smooth_quant_group_gemm(a: torch.Tensor,
+                            b: torch.Tensor,
+                            m_list: torch.Tensor,
+                            expand_idx: Optional[torch.Tensor],
+                            c: Optional[torch.Tensor],
+                            alpha: Optional[torch.Tensor],
+                            beta: Optional[torch.Tensor],
+                            a_scale: torch.Tensor,
+                            b_scale: torch.Tensor,
+                            dtype,
+                            max_m: int = 0
+                            ) -> torch.Tensor:
+    return tmo.smooth_quant_group_gemm(a, b, m_list, expand_idx, c, alpha, beta,
+                                       a_scale, b_scale, dtype, max_m)
+
+def moe_combine_result(input: torch.Tensor,
+                       reduce_weight: torch.Tensor,
+                       gather_ids: torch.Tensor,
+                       residual: Optional[torch.Tensor],
+                       cusum_token_count: Optional[torch.Tensor],
+                       start_expert_id: int,
+                       expert_size: int,
+                       bias: Optional[torch.Tensor] = None) -> torch.Tensor:
+    return tmo.moe_combine_result(input, reduce_weight, gather_ids,
+                                  residual, cusum_token_count,
+                                  start_expert_id, expert_size, bias)
+
+def moe_quantize(x: torch.Tensor,
+                 smooth: torch.Tensor,
+                 zero: Optional[torch.Tensor] = None,
+                 token_count: Optional[torch.Tensor] = None,
+                 gather_index: Optional[torch.Tensor] = None,
+                 gather_index_start_position: Optional[torch.Tensor] = None,
+                 output: Optional[torch.Tensor] = None,
+                 output_scale: Optional[torch.Tensor] = None,
+                 dynamic_quant: bool = True
+                ) -> Tuple[torch.Tensor, torch.Tensor]:
+    return tmo.moe_quantize(x, smooth, zero, token_count, gather_index, gather_index_start_position,
+                            output, output_scale, dynamic_quant)

vllm-v0.6.2/vllm/adapter_commons/layers.py

@@ -0,0 +1,14 @@
+from dataclasses import dataclass
+from typing import Tuple
+
+
+@dataclass
+class AdapterMapping:
+    # Per every token in input_ids:
+    index_mapping: Tuple[int, ...]
+    # Per sampled token:
+    prompt_mapping: Tuple[int, ...]
+
+    def __post_init__(self):
+        self.index_mapping = tuple(self.index_mapping)
+        self.prompt_mapping = tuple(self.prompt_mapping)

vllm-v0.6.2/vllm/adapter_commons/models.py

@@ -0,0 +1,104 @@
+from abc import ABC, abstractmethod
+from typing import Any, Callable, Dict, Hashable, Optional, TypeVar
+
+from torch import nn
+
+from vllm.logger import init_logger
+from vllm.utils import LRUCache
+
+logger = init_logger(__name__)
+
+
+class AdapterModel(ABC):
+
+    def __init__(self, model_id=None):
+        self.id = model_id
+
+    @abstractmethod
+    def from_local_checkpoint(cls, model_dir, model_id=None, **kwargs):
+        # Common initialization code
+        # Load weights or embeddings from local checkpoint
+        raise NotImplementedError("Subclasses must implement this method.")
+
+
+T = TypeVar('T')
+
+
+class AdapterLRUCache(LRUCache[T]):
+
+    def __init__(self, capacity: int, deactivate_fn: Callable[[Hashable],
+                                                              None]):
+        super().__init__(capacity)
+        self.deactivate_fn = deactivate_fn
+
+    def _on_remove(self, key: Hashable, value: Optional[T]):
+        logger.debug("Removing adapter int id: %d", key)
+        self.deactivate_fn(key)
+        return super()._on_remove(key, value)
+
+
+class AdapterModelManager(ABC):
+
+    def __init__(
+        self,
+        model: nn.Module,
+    ):
+        """Create a AdapterModelManager and adapter for a given model.
+        Args:
+            model: the model to be adapted.
+        """
+        self.model: nn.Module = model
+        self._registered_adapters: Dict[int, Any] = {}
+        # Dict instead of a Set for compatibility with LRUCache.
+        self._active_adapters: Dict[int, None] = {}
+        self.adapter_type = 'Adapter'
+        self._last_mapping = None
+
+    def __len__(self) -> int:
+        return len(self._registered_adapters)
+
+    @property
+    @abstractmethod
+    def adapter_slots(self) -> int:
+        raise NotImplementedError
+
+    @property
+    @abstractmethod
+    def capacity(self) -> int:
+        raise NotImplementedError
+
+    @abstractmethod
+    def activate_adapter(self, adapter_id: int) -> bool:
+        raise NotImplementedError
+
+    @abstractmethod
+    def deactivate_adapter(self, adapter_id: int) -> bool:
+        raise NotImplementedError
+
+    @abstractmethod
+    def add_adapter(self, adapter: Any) -> bool:
+        raise NotImplementedError
+
+    @abstractmethod
+    def set_adapter_mapping(self, mapping: Any) -> None:
+        raise NotImplementedError
+
+    @abstractmethod
+    def remove_adapter(self, adapter_id: int) -> bool:
+        raise NotImplementedError
+
+    @abstractmethod
+    def remove_all_adapters(self) -> None:
+        raise NotImplementedError
+
+    @abstractmethod
+    def get_adapter(self, adapter_id: int) -> Optional[Any]:
+        raise NotImplementedError
+
+    @abstractmethod
+    def list_adapters(self) -> Dict[int, Any]:
+        raise NotImplementedError
+
+    @abstractmethod
+    def pin_adapter(self, adapter_id: int) -> bool:
+        raise NotImplementedError

vllm-v0.6.2/vllm/adapter_commons/request.py

@@ -0,0 +1,23 @@
+from abc import ABC, abstractmethod
+
+
+class AdapterRequest(ABC):
+    """
+    Base class for adapter requests.
+    """
+
+    @property
+    @abstractmethod
+    def adapter_id(self) -> int:
+        raise NotImplementedError
+
+    def __post_init__(self) -> None:
+        if self.adapter_id < 1:
+            raise ValueError(f"id must be > 0, got {self.adapter_id}")
+
+    def __eq__(self, value: object) -> bool:
+        return isinstance(
+            value, self.__class__) and self.adapter_id == value.adapter_id
+
+    def __hash__(self) -> int:
+        return hash(self.adapter_id)

vllm-v0.6.2/vllm/adapter_commons/utils.py

@@ -0,0 +1,90 @@
+from typing import Any, Callable, Dict, Optional, Set
+
+
+## model functions
+def deactivate_adapter(adapter_id: int, active_adapters: Dict[int, None],
+                       deactivate_func: Callable) -> bool:
+    if adapter_id in active_adapters:
+        deactivate_func(adapter_id)
+        active_adapters.pop(adapter_id)
+        return True
+    return False
+
+
+def add_adapter(adapter: Any, registered_adapters: Dict[int, Any],
+                capacity: int, add_func: Callable) -> bool:
+    if adapter.id not in registered_adapters:
+        if len(registered_adapters) >= capacity:
+            raise RuntimeError('No free adapter slots.')
+        add_func(adapter)
+        registered_adapters[adapter.id] = adapter
+        return True
+    return False
+
+
+def set_adapter_mapping(mapping: Any, last_mapping: Any,
+                        set_mapping_func: Callable) -> Any:
+    if last_mapping != mapping:
+        set_mapping_func(mapping)
+        return mapping
+    return last_mapping
+
+
+def remove_adapter(adapter_id: int, registered_adapters: Dict[int, Any],
+                   deactivate_func: Callable) -> bool:
+    deactivate_func(adapter_id)
+    return bool(registered_adapters.pop(adapter_id, None))
+
+
+def list_adapters(registered_adapters: Dict[int, Any]) -> Dict[int, Any]:
+    return dict(registered_adapters)
+
+
+def get_adapter(adapter_id: int,
+                registered_adapters: Dict[int, Any]) -> Optional[Any]:
+    return registered_adapters.get(adapter_id)
+
+
+## worker functions
+def set_active_adapters_worker(requests: Set[Any], mapping: Optional[Any],
+                               apply_adapters_func,
+                               set_adapter_mapping_func) -> None:
+    apply_adapters_func(requests)
+    set_adapter_mapping_func(mapping)
+
+
+def add_adapter_worker(adapter_request: Any, list_adapters_func,
+                       load_adapter_func, add_adapter_func,
+                       activate_adapter_func) -> bool:
+    if adapter_request.adapter_id in list_adapters_func():
+        return False
+    loaded_adapter = load_adapter_func(adapter_request)
+    loaded = add_adapter_func(loaded_adapter)
+    activate_adapter_func(loaded_adapter.id)
+    return loaded
+
+
+def apply_adapters_worker(adapter_requests: Set[Any], list_adapters_func,
+                          adapter_slots: int, remove_adapter_func,
+                          add_adapter_func) -> None:
+    models_that_exist = list_adapters_func()
+    models_map = {
+        adapter_request.adapter_id: adapter_request
+        for adapter_request in adapter_requests if adapter_request
+    }
+    if len(models_map) > adapter_slots:
+        raise RuntimeError(
+            f"Number of requested models ({len(models_map)}) is greater "
+            f"than the number of GPU model slots "
+            f"({adapter_slots}).")
+    new_models = set(models_map)
+    models_to_add = new_models - models_that_exist
+    models_to_remove = models_that_exist - new_models
+    for adapter_id in models_to_remove:
+        remove_adapter_func(adapter_id)
+    for adapter_id in models_to_add:
+        add_adapter_func(models_map[adapter_id])
+
+
+def list_adapters_worker(adapter_manager_list_adapters_func) -> Set[int]:
+    return set(adapter_manager_list_adapters_func())

vllm-v0.6.2/vllm/adapter_commons/worker_manager.py

@@ -0,0 +1,36 @@
+from abc import ABC, abstractmethod
+from typing import Any, Optional, Set
+
+import torch
+
+
+class AbstractWorkerManager(ABC):
+
+    def __init__(self, device: torch.device):
+        self.device = device
+
+    @property
+    @abstractmethod
+    def is_enabled(self) -> bool:
+        raise NotImplementedError
+
+    @abstractmethod
+    def set_active_adapters(self, requests: Set[Any],
+                            mapping: Optional[Any]) -> None:
+        raise NotImplementedError
+
+    @abstractmethod
+    def add_adapter(self, adapter_request: Any) -> bool:
+        raise NotImplementedError
+
+    @abstractmethod
+    def remove_adapter(self, adapter_id: int) -> bool:
+        raise NotImplementedError
+
+    @abstractmethod
+    def remove_all_adapters(self) -> None:
+        raise NotImplementedError
+
+    @abstractmethod
+    def list_adapters(self) -> Set[int]:
+        raise NotImplementedError

vllm-v0.6.2/vllm/assets/audio.py

@@ -0,0 +1,28 @@
+from dataclasses import dataclass
+from typing import Literal, Tuple
+from urllib.parse import urljoin
+
+import librosa
+import numpy as np
+
+from vllm.assets.base import get_vllm_public_assets, vLLM_S3_BUCKET_URL
+
+ASSET_DIR = "multimodal_asset"
+
+
+@dataclass(frozen=True)
+class AudioAsset:
+    name: Literal["winning_call", "mary_had_lamb"]
+
+    @property
+    def audio_and_sample_rate(self) -> Tuple[np.ndarray, int]:
+
+        audio_path = get_vllm_public_assets(filename=f"{self.name}.ogg",
+                                            s3_prefix=ASSET_DIR)
+        y, sr = librosa.load(audio_path, sr=None)
+        assert isinstance(sr, int)
+        return y, sr
+
+    @property
+    def url(self) -> str:
+        return urljoin(vLLM_S3_BUCKET_URL, f"{ASSET_DIR}/{self.name}.ogg")

vllm-v0.6.2/vllm/assets/base.py

@@ -0,0 +1,39 @@
+from functools import lru_cache
+from pathlib import Path
+from typing import Optional
+
+import vllm.envs as envs
+from vllm.connections import global_http_connection
+from vllm.envs import VLLM_IMAGE_FETCH_TIMEOUT
+
+vLLM_S3_BUCKET_URL = "https://vllm-public-assets.s3.us-west-2.amazonaws.com"
+
+
+def get_cache_dir() -> Path:
+    """Get the path to the cache for storing downloaded assets."""
+    path = Path(envs.VLLM_ASSETS_CACHE)
+    path.mkdir(parents=True, exist_ok=True)
+
+    return path
+
+
+@lru_cache
+def get_vllm_public_assets(filename: str,
+                           s3_prefix: Optional[str] = None) -> Path:
+    """
+    Download an asset file from ``s3://vllm-public-assets``
+    and return the path to the downloaded file.
+    """
+    asset_directory = get_cache_dir() / "vllm_public_assets"
+    asset_directory.mkdir(parents=True, exist_ok=True)
+
+    asset_path = asset_directory / filename
+    if not asset_path.exists():
+        if s3_prefix is not None:
+            filename = s3_prefix + "/" + filename
+        global_http_connection.download_file(
+            f"{vLLM_S3_BUCKET_URL}/{filename}",
+            asset_path,
+            timeout=VLLM_IMAGE_FETCH_TIMEOUT)
+
+    return asset_path

vllm-v0.6.2/vllm/assets/image.py

@@ -0,0 +1,30 @@
+from dataclasses import dataclass
+from typing import Literal
+
+import torch
+from PIL import Image
+
+from vllm.assets.base import get_vllm_public_assets
+
+VLM_IMAGES_DIR = "vision_model_images"
+
+
+@dataclass(frozen=True)
+class ImageAsset:
+    name: Literal["stop_sign", "cherry_blossom"]
+
+    @property
+    def pil_image(self) -> Image.Image:
+
+        image_path = get_vllm_public_assets(filename=f"{self.name}.jpg",
+                                            s3_prefix=VLM_IMAGES_DIR)
+        return Image.open(image_path)
+
+    @property
+    def image_embeds(self) -> torch.Tensor:
+        """
+        Image embeddings, only used for testing purposes with llava 1.5.
+        """
+        image_path = get_vllm_public_assets(filename=f"{self.name}.pt",
+                                            s3_prefix=VLM_IMAGES_DIR)
+        return torch.load(image_path, map_location="cpu")

vllm-v0.6.2/vllm/assets/video.py

@@ -0,0 +1,85 @@
+from dataclasses import dataclass
+from functools import lru_cache
+from typing import List, Literal
+
+import numpy as np
+import numpy.typing as npt
+from huggingface_hub import hf_hub_download
+from PIL import Image
+
+from vllm.multimodal.utils import (sample_frames_from_video,
+                                   try_import_video_packages)
+
+from .base import get_cache_dir
+
+
+@lru_cache
+def download_video_asset(filename: str) -> str:
+    """
+    Download and open an image from huggingface
+    repo: raushan-testing-hf/videos-test
+    """
+    video_directory = get_cache_dir() / "video-eample-data"
+    video_directory.mkdir(parents=True, exist_ok=True)
+
+    video_path = video_directory / filename
+    video_path_str = str(video_path)
+    if not video_path.exists():
+        video_path_str = hf_hub_download(
+            repo_id="raushan-testing-hf/videos-test",
+            filename=filename,
+            repo_type="dataset",
+            cache_dir=video_directory,
+        )
+    return video_path_str
+
+
+def video_to_ndarrays(path: str, num_frames: int = -1) -> npt.NDArray:
+    cv2, _ = try_import_video_packages()
+
+    cap = cv2.VideoCapture(path)
+    if not cap.isOpened():
+        raise ValueError(f"Could not open video file {path}")
+
+    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    frames = []
+    for i in range(total_frames):
+        ret, frame = cap.read()
+        if ret:
+            frames.append(frame)
+    cap.release()
+
+    frames = np.stack(frames)
+    frames = sample_frames_from_video(frames, num_frames)
+    if len(frames) < num_frames:
+        raise ValueError(f"Could not read enough frames from video file {path}"
+                         f" (expected {num_frames} frames, got {len(frames)})")
+    return frames
+
+
+def video_to_pil_images_list(path: str,
+                             num_frames: int = -1) -> List[Image.Image]:
+    cv2, _ = try_import_video_packages()
+    frames = video_to_ndarrays(path, num_frames)
+    return [
+        Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
+        for frame in frames
+    ]
+
+
+@dataclass(frozen=True)
+class VideoAsset:
+    name: Literal["sample_demo_1.mp4"]
+    num_frames: int = -1
+
+    @property
+    def pil_images(self) -> List[Image.Image]:
+        video_path = download_video_asset(self.name)
+        ret = video_to_pil_images_list(video_path, self.num_frames)
+        return ret
+
+    @property
+    def np_ndarrays(self) -> npt.NDArray:
+        video_path = download_video_asset(self.name)
+        ret = video_to_ndarrays(video_path, self.num_frames)
+        return ret

vllm-v0.6.2/vllm/attention/__init__.py

@@ -0,0 +1,17 @@
+from vllm.attention.backends.abstract import (AttentionBackend,
+                                              AttentionMetadata,
+                                              AttentionMetadataBuilder,
+                                              AttentionState, AttentionType)
+from vllm.attention.layer import Attention
+from vllm.attention.selector import get_attn_backend
+
+__all__ = [
+    "Attention",
+    "AttentionBackend",
+    "AttentionMetadata",
+    "AttentionType",
+    "AttentionMetadataBuilder",
+    "Attention",
+    "AttentionState",
+    "get_attn_backend",
+]

vllm-v0.6.2/vllm/attention/backends/abstract.py

@@ -0,0 +1,246 @@
+from abc import ABC, abstractmethod
+from contextlib import contextmanager
+from dataclasses import dataclass, fields
+from enum import Enum, auto
+from typing import (TYPE_CHECKING, Any, Dict, Generic, List, Optional, Set,
+                    Tuple, Type, TypeVar)
+
+import torch
+
+from vllm.multimodal import MultiModalPlaceholderMap
+
+if TYPE_CHECKING:
+    from vllm.worker.model_runner_base import (ModelRunnerBase,
+                                               ModelRunnerInputBase,
+                                               ModelRunnerInputBuilderBase)
+
+
+class AttentionType(Enum):
+    DECODER = auto()  # Decoder attention between previous layer Q/K/V
+    ENCODER = auto(
+    )  # Encoder attention between previous layer Q/K/V for encoder-decoder
+    ENCODER_ONLY = auto()  # Encoder attention between previous layer Q/K/V
+    ENCODER_DECODER = auto(
+    )  # Attention between dec. Q and enc. K/V for encoder-decoder
+
+
+class AttentionBackend(ABC):
+    """Abstract class for attention backends."""
+
+    @staticmethod
+    @abstractmethod
+    def get_name() -> str:
+        raise NotImplementedError
+
+    @staticmethod
+    @abstractmethod
+    def get_impl_cls() -> Type["AttentionImpl"]:
+        raise NotImplementedError
+
+    @staticmethod
+    @abstractmethod
+    def get_metadata_cls() -> Type["AttentionMetadata"]:
+        raise NotImplementedError
+
+    @staticmethod
+    @abstractmethod
+    def get_state_cls() -> Type["AttentionState"]:
+        raise NotImplementedError
+
+    @classmethod
+    def make_metadata(cls, *args, **kwargs) -> "AttentionMetadata":
+        return cls.get_metadata_cls()(*args, **kwargs)
+
+    @staticmethod
+    @abstractmethod
+    def get_builder_cls() -> Type["AttentionMetadataBuilder"]:
+        raise NotImplementedError
+
+    @classmethod
+    def make_metadata_builder(cls, *args,
+                              **kwargs) -> "AttentionMetadataBuilder":
+        return cls.get_builder_cls()(*args, **kwargs)
+
+    @staticmethod
+    @abstractmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        raise NotImplementedError
+
+    @staticmethod
+    @abstractmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: torch.Tensor,
+    ) -> None:
+        raise NotImplementedError
+
+    @staticmethod
+    @abstractmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: torch.Tensor,
+    ) -> None:
+        raise NotImplementedError
+
+    def advance_step(self, model_input: "ModelRunnerInputBase",
+                     sampled_token_ids: Optional[torch.Tensor],
+                     block_size: int, num_seqs: int, num_queries: int) -> None:
+        raise NotImplementedError
+
+
+@dataclass
+class AttentionMetadata:
+    """Attention metadata for prefill and decode batched together."""
+    # Total number of prefill requests.
+    num_prefills: int
+    # Number of prefill tokens.
+    num_prefill_tokens: int
+    # Number of decode tokens. Note that it is equivalent to the number of
+    # decode requests.
+    num_decode_tokens: int
+    # (num_tokens,). The indices of the token slots that input tokens will be
+    # stored into. E.g., if `slot_mapping` is [35, 2, 17] and the block size
+    # is 16, the three tokens are stored in the 3rd slot in block 2, 2nd slot
+    # in block 0, and 1st slot in block 1, respectively.
+    slot_mapping: torch.Tensor
+
+    # The index maps that relate multi-modal embeddings to the corresponding
+    # placeholders.
+    #
+    # N.B. These aren't really related to attention and don't belong on this
+    # type -- this is just a temporary solution to make them available to
+    # `model_executable`.
+    multi_modal_placeholder_index_maps: Optional[Dict[
+        str, MultiModalPlaceholderMap.IndexMap]]
+
+    @property
+    @abstractmethod
+    def prefill_metadata(self) -> Optional["AttentionMetadata"]:
+        """Return the attention metadata that's required to run prefill
+        attention."""
+        pass
+
+    @property
+    @abstractmethod
+    def decode_metadata(self) -> Optional["AttentionMetadata"]:
+        """Return the attention metadata that's required to run decode
+        attention."""
+        pass
+
+    def asdict_zerocopy(self,
+                        skip_fields: Optional[Set[str]] = None
+                        ) -> Dict[str, Any]:
+        """Similar to dataclasses.asdict, but avoids deepcopying."""
+        if skip_fields is None:
+            skip_fields = set()
+        # Note that if we add dataclasses as fields, they will need
+        # similar handling.
+        return {
+            field.name: getattr(self, field.name)
+            for field in fields(self) if field.name not in skip_fields
+        }
+
+
+T = TypeVar("T", bound=AttentionMetadata)
+
+
+class AttentionState(ABC, Generic[T]):
+    """Holds attention backend-specific objects reused during the
+    lifetime of the model runner."""
+
+    @abstractmethod
+    def __init__(self, runner: "ModelRunnerBase"):
+        ...
+
+    @abstractmethod
+    @contextmanager
+    def graph_capture(self, max_batch_size: int):
+        """Context manager used when capturing CUDA graphs."""
+        yield
+
+    @abstractmethod
+    def graph_clone(self, batch_size: int) -> "AttentionState[T]":
+        """Clone attention state to save in CUDA graph metadata."""
+        ...
+
+    @abstractmethod
+    def graph_capture_get_metadata_for_batch(
+            self,
+            batch_size: int,
+            is_encoder_decoder_model: bool = False) -> T:
+        """Get attention metadata for CUDA graph capture of batch_size."""
+        ...
+
+    @abstractmethod
+    def get_graph_input_buffers(
+            self,
+            attn_metadata: T,
+            is_encoder_decoder_model: bool = False) -> Dict[str, Any]:
+        """Get attention-specific input buffers for CUDA graph capture."""
+        ...
+
+    @abstractmethod
+    def prepare_graph_input_buffers(
+            self,
+            input_buffers: Dict[str, Any],
+            attn_metadata: T,
+            is_encoder_decoder_model: bool = False) -> None:
+        """In-place modify input buffers dict for CUDA graph replay."""
+        ...
+
+    @abstractmethod
+    def begin_forward(self, model_input: "ModelRunnerInputBase") -> None:
+        """Prepare state for forward pass."""
+        ...
+
+
+class AttentionMetadataBuilder(ABC, Generic[T]):
+    """Abstract class for attention metadata builders."""
+
+    @abstractmethod
+    def __init__(self, input_builder: "ModelRunnerInputBuilderBase") -> None:
+        raise NotImplementedError
+
+    @abstractmethod
+    def build(self, seq_lens: List[int], query_lens: List[int],
+              cuda_graph_pad_size: int, batch_size: int) -> T:
+        """Build attention metadata with on-device tensors."""
+        raise NotImplementedError
+
+
+class AttentionImpl(ABC, Generic[T]):
+
+    @abstractmethod
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: Optional[int] = None,
+        alibi_slopes: Optional[List[float]] = None,
+        sliding_window: Optional[int] = None,
+        kv_cache_dtype: str = "auto",
+        blocksparse_params: Optional[Dict[str, Any]] = None,
+        logits_soft_cap: Optional[float] = None,
+    ) -> None:
+        raise NotImplementedError
+
+    @abstractmethod
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: T,
+        k_scale: float = 1.0,
+        v_scale: float = 1.0,
+        attn_type: AttentionType = AttentionType.DECODER,
+    ) -> torch.Tensor:
+        raise NotImplementedError

vllm-v0.6.2/vllm/attention/backends/blocksparse_attn.py

@@ -0,0 +1,447 @@
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional, Tuple, Type
+
+import torch
+
+from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
+                                              AttentionMetadata, AttentionType)
+from vllm.attention.backends.utils import (CommonAttentionState,
+                                           CommonMetadataBuilder)
+from vllm.attention.ops.blocksparse_attention.interface import (
+    LocalStridedBlockSparseAttn, get_head_sliding_step)
+from vllm.attention.ops.paged_attn import PagedAttention
+from vllm.distributed import (get_tensor_model_parallel_rank,
+                              get_tensor_model_parallel_world_size)
+
+
+@dataclass
+class BlocksparseParams:
+    max_seqlen: int
+
+    # Num q heads per tensor-parallel rank/partition
+    num_heads: int  # per TP partition
+    # Num kv heads per tensor-parallel rank/partition
+    num_kv_heads: int
+
+    # block size used for blocksparse attention.
+    # This is the block_size used in `local_blocks`, `vert_stride`.
+    block_size: int
+
+    # Number of blocks for local attention, i.e., number of
+    # local attended tokens / `sparse_block_size`
+    local_blocks: int
+
+    # Attend to one block per every `vert_stride` blocks.
+    # Controlling the sparsity
+    vert_stride: int
+    """
+    If to use the same vertical stride offset for all heads, 
+    i.e., attend to the same block of tokens on all heads.
+    By default, it is False, i.e., attention on the non-local 
+    blocks depends on the `head_idx`, that is on
+    blocks satisfying 
+    `(block_idx + head_idx * head_sliding_step + 1) % vert_stride == 0`
+    where `head_sliding_step=max(1, int(vert_stride / num_total_heads))`,
+            `block_idx = position_id // sparse_block_size`.
+    See `..ops.blocksparse_attention.utils:get_sparse_attn_mask`
+    for more detail.
+    """
+    homo_head: bool = False
+
+    # If within a group, the kv offsets that each q attends is the same or no.
+    homo_head_group: bool = False
+
+    # Decided by homo_head and homo_head group
+    head_sliding_step: int = field(init=False)
+
+    # range of q heads to for a TP rank
+    active_head_range: Tuple = field(init=False)
+
+    def __post_init__(self):
+        assert self.block_size > 0
+        assert self.local_blocks >= 0
+        assert self.vert_stride >= 1
+        assert self.num_heads % self.num_kv_heads == 0
+
+        tp_size = get_tensor_model_parallel_world_size()
+        tp_rank = get_tensor_model_parallel_rank()
+        total_heads = tp_size * self.num_heads
+        total_kv_heads = tp_size * self.num_kv_heads
+
+        if self.homo_head:
+            self.head_sliding_step = 0
+        elif self.homo_head_group:
+            head_sliding_step = get_head_sliding_step(total_kv_heads,
+                                                      self.vert_stride)
+            # negative indicates sliding along kv heads, i.e., homo q group
+            self.head_sliding_step = -head_sliding_step
+        else:
+            self.head_sliding_step = get_head_sliding_step(
+                total_heads, self.vert_stride)
+
+        self.active_head_range = (
+            tp_rank * self.num_heads,
+            (tp_rank + 1) * self.num_heads,
+        )
+
+
+class BlocksparseFlashAttentionBackend(AttentionBackend):
+
+    @staticmethod
+    def get_impl_cls() -> Type["BlocksparseFlashAttentionImpl"]:
+        return BlocksparseFlashAttentionImpl
+
+    @staticmethod
+    def get_metadata_cls() -> Type["AttentionMetadata"]:
+        return BlocksparseFlashAttentionMetadata
+
+    @staticmethod
+    def get_builder_cls() -> Type["BlocksparseFlashAttentionMetadataBuilder"]:
+        return BlocksparseFlashAttentionMetadataBuilder
+
+    @staticmethod
+    def get_state_cls() -> Type["CommonAttentionState"]:
+        return CommonAttentionState
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        return PagedAttention.get_kv_cache_shape(num_blocks, block_size,
+                                                 num_kv_heads, head_size)
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: Dict[int, int],
+    ) -> None:
+        PagedAttention.swap_blocks(src_kv_cache, dst_kv_cache, src_to_dst)
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: Dict[int, List[int]],
+    ) -> None:
+        PagedAttention.copy_blocks(kv_caches, src_to_dists)
+
+
+@dataclass
+class BlocksparseFlashAttentionMetadata(AttentionMetadata):
+    """A copy of Metadata for FlashAttentionBackend,
+    to avoid having to install flash_attn.
+
+    NOTE: Any python object stored here is not updated when it is
+    cuda-graph replayed. If you have values that need to be changed
+    dynamically, it should be stored in tensor. The tensor has to be
+    updated from `CUDAGraphRunner.forward` API.
+    """
+    # (batch_size,). The sequence length per sequence. Sequence length means
+    # the computed tokens + new tokens None if it is a decoding.
+    seq_lens: Optional[List[int]]
+    # seq_lens stored as a tensor.
+    seq_lens_tensor: Optional[torch.Tensor]
+
+    # NOTE(sang): Definition of context_len, query_len, and seq_len.
+    # |---------- N-1 iteration --------|
+    # |---------------- N iteration ---------------------|
+    # |- tokenA -|......................|-- newTokens ---|
+    # |---------- context_len ----------|
+    # |-------------------- seq_len ----------------------|
+    #                                   |-- query_len ---|
+
+    # Maximum query length in the batch. None for decoding.
+    max_query_len: Optional[int]
+    # Maximum sequence length among prefill batch. 0 if there are decoding
+    # requests only.
+    max_prefill_seq_len: int
+    # Maximum sequence length among decode batch. 0 if there are prefill
+    # requests only.
+    max_decode_seq_len: int
+    # (batch_size + 1,). The cumulative subquery lengths of the sequences in
+    # the batch, used to index into subquery. E.g., if the subquery length
+    # is [4, 6], it is [0, 4, 10].
+    query_start_loc: Optional[torch.Tensor]
+    # (batch_size + 1,). The cumulative sequence lengths of the sequences in
+    # the batch, used to index into sequence. E.g., if the sequence length is
+    # [4, 6], it is [0, 4, 10].
+    seq_start_loc: Optional[torch.Tensor]
+    # (batch_size,) A tensor of context lengths (tokens that are computed
+    # so far).
+    context_lens_tensor: Optional[torch.Tensor]
+
+    # (batch_size, max_blocks_per_seq).
+    # Block addresses per sequence. (Seq id -> list of physical block)
+    # E.g., [0, 1, 2] means tokens are stored in 0th, 1st, and 2nd blocks
+    # in the kv cache. Each block can contain up to block_size tokens.
+    # 2nd dimensions are padded up to max_blocks_per_seq if it is cuda-graph
+    # captured.
+    block_tables: Optional[torch.Tensor]
+
+    # Whether or not if cuda graph is enabled.
+    # Cuda-graph is currently enabled for decoding only.
+    # TODO(woosuk): Move `use_cuda_graph` out since it's unrelated to attention.
+    use_cuda_graph: bool
+
+    # Max number of query tokens for among request in the batch.
+    max_decode_query_len: Optional[int] = None
+
+    _cached_prefill_metadata: Optional[
+        "BlocksparseFlashAttentionMetadata"] = None
+    _cached_decode_metadata: Optional[
+        "BlocksparseFlashAttentionMetadata"] = None
+
+    @property
+    def prefill_metadata(
+            self) -> Optional["BlocksparseFlashAttentionMetadata"]:
+        if self.num_prefills == 0:
+            return None
+
+        if self._cached_prefill_metadata is not None:
+            return self._cached_prefill_metadata
+
+        assert self.seq_lens is not None
+        assert self.seq_lens_tensor is not None
+        assert self.query_start_loc is not None
+        assert self.context_lens_tensor is not None
+        assert self.block_tables is not None
+        assert self.seq_start_loc is not None
+
+        self._cached_prefill_metadata = BlocksparseFlashAttentionMetadata(
+            num_prefills=self.num_prefills,
+            num_prefill_tokens=self.num_prefill_tokens,
+            num_decode_tokens=0,
+            slot_mapping=self.slot_mapping[:self.num_prefill_tokens],
+            multi_modal_placeholder_index_maps=self.
+            multi_modal_placeholder_index_maps,
+            seq_lens=self.seq_lens[:self.num_prefills],
+            seq_lens_tensor=self.seq_lens_tensor[:self.num_prefills],
+            max_query_len=self.max_query_len,
+            max_prefill_seq_len=self.max_prefill_seq_len,
+            max_decode_seq_len=0,
+            query_start_loc=self.query_start_loc[:self.num_prefills + 1],
+            seq_start_loc=self.seq_start_loc[:self.num_prefills + 1],
+            context_lens_tensor=self.context_lens_tensor[:self.num_prefills],
+            block_tables=self.block_tables[:self.num_prefills],
+            use_cuda_graph=False,
+        )
+        return self._cached_prefill_metadata
+
+    @property
+    def decode_metadata(self) -> Optional["BlocksparseFlashAttentionMetadata"]:
+        if self.num_decode_tokens == 0:
+            return None
+
+        if self._cached_decode_metadata is not None:
+            return self._cached_decode_metadata
+        assert self.block_tables is not None
+        assert self.seq_lens_tensor is not None
+
+        self._cached_decode_metadata = BlocksparseFlashAttentionMetadata(
+            num_prefills=0,
+            num_prefill_tokens=0,
+            num_decode_tokens=self.num_decode_tokens,
+            slot_mapping=self.slot_mapping[self.num_prefill_tokens:],
+            multi_modal_placeholder_index_maps=None,
+            seq_lens=None,
+            seq_lens_tensor=self.seq_lens_tensor[self.num_prefills:],
+            max_query_len=None,
+            max_prefill_seq_len=0,
+            max_decode_seq_len=self.max_decode_seq_len,
+            query_start_loc=None,
+            seq_start_loc=None,
+            context_lens_tensor=None,
+            block_tables=self.block_tables[self.num_prefills:],
+            use_cuda_graph=self.use_cuda_graph,
+        )
+        return self._cached_decode_metadata
+
+
+class BlocksparseFlashAttentionMetadataBuilder(
+        CommonMetadataBuilder[BlocksparseFlashAttentionMetadata]):
+
+    _metadata_cls = BlocksparseFlashAttentionMetadata
+
+
+class BlocksparseFlashAttentionImpl(AttentionImpl):
+    """
+    If the input tensors contain prompt tokens, the layout is as follows:
+    |<--------------- num_prompt_tokens -------------->|
+    |<--prompt_0-->|<--prompt_1-->|...|<--prompt_N-1-->|
+
+    Otherwise, the layout is as follows:
+    |<------------------ num_generation_tokens (M) ----------------->|
+    |<--generation_0-->|..........|<--generation_M-1-->|<--padding-->|
+
+    Generation tokens can contain padding when cuda-graph is used.
+    Currently, prompt tokens don't contain any padding.
+
+    The prompts might have different lengths, while the generation tokens
+    always have length 1.
+
+    """
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: Optional[List[float]],
+        sliding_window: Optional[int],
+        kv_cache_dtype: str,
+        blocksparse_params: Optional[Dict[str, Any]] = None,
+        logits_soft_cap: Optional[float] = None,
+    ) -> None:
+        assert blocksparse_params is not None
+        assert alibi_slopes is None, ValueError(
+            "Alibi not support for blocksparse flash attention.")
+        assert sliding_window is None, ValueError(
+            "sliding_window is invalid for blocksparse attention.")
+        assert logits_soft_cap is None, ValueError(
+            "logits_soft_cap is invalid for blocksparse attention.")
+
+        if "num_heads" not in blocksparse_params:
+            blocksparse_params["num_heads"] = num_heads
+        if "num_kv_heads" not in blocksparse_params:
+            blocksparse_params["num_kv_heads"] = num_kv_heads or num_heads
+        self.blocksparse_params = BlocksparseParams(**blocksparse_params)
+        self.kv_cache_dtype = kv_cache_dtype
+
+        self.num_heads = num_heads
+        self.head_size = head_size
+        self.scale = float(scale)
+        self.alibi_slopes = alibi_slopes
+        self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
+
+        assert self.num_heads % self.num_kv_heads == 0
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+
+        self.local_blocks = self.blocksparse_params.local_blocks
+        self.vert_stride = self.blocksparse_params.vert_stride
+        self.sparse_block_size = self.blocksparse_params.block_size
+        self.head_sliding_step = self.blocksparse_params.head_sliding_step
+
+        suppored_head_sizes = PagedAttention.get_supported_head_sizes()
+        if head_size not in suppored_head_sizes:
+            raise ValueError(
+                f"Head size {head_size} is not supported by PagedAttention. "
+                f"Supported head sizes are: {suppored_head_sizes}.")
+
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.tp_rank = get_tensor_model_parallel_rank()
+
+        total_num_heads = num_heads * self.tp_size
+        self.bs_attn = LocalStridedBlockSparseAttn(
+            total_num_heads,
+            self.blocksparse_params.max_seqlen,
+            self.blocksparse_params.local_blocks,
+            self.blocksparse_params.vert_stride,
+            self.blocksparse_params.block_size,
+            homo_head=self.blocksparse_params.homo_head,
+            active_head_range=self.blocksparse_params.active_head_range,
+        )
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: BlocksparseFlashAttentionMetadata,
+        k_scale: float = 1.0,
+        v_scale: float = 1.0,
+        attn_type: AttentionType = AttentionType.DECODER,
+    ) -> torch.Tensor:
+        """Forward pass with FlashAttention and PagedAttention.
+
+        Args:
+            query: shape = [num_tokens, num_heads * head_size]
+            key: shape = [num_tokens, num_kv_heads * head_size]
+            value: shape = [num_tokens, num_kv_heads * head_size]
+            kv_cache = [2, num_blocks, block_size * num_kv_heads * head_size]
+                NOTE: kv_cache will be an empty tensor with shape [0]
+                for profiling run.
+            attn_metadata: Metadata for attention.
+        Returns:
+            shape = [num_tokens, num_heads * head_size]
+        """
+        if attn_type != AttentionType.DECODER:
+            raise NotImplementedError("Encoder self-attention and "
+                                      "encoder/decoder cross-attention "
+                                      "are not implemented for "
+                                      "BlocksparseFlashAttentionImpl")
+
+        num_tokens, hidden_size = query.shape
+        # Reshape the query, key, and value tensors.
+        query = query.view(-1, self.num_heads, self.head_size)
+        key = key.view(-1, self.num_kv_heads, self.head_size)
+        value = value.view(-1, self.num_kv_heads, self.head_size)
+
+        if kv_cache.numel() > 0:
+            key_cache, value_cache = PagedAttention.split_kv_cache(
+                kv_cache, self.num_kv_heads, self.head_size)
+
+            # Reshape the input keys and values and store them in the cache.
+            # If kv_cache is not provided, the new key and value tensors are
+            # not cached. This happens during the initial memory profiling run.
+
+            PagedAttention.write_to_paged_cache(
+                key,
+                value,
+                key_cache,
+                value_cache,
+                attn_metadata.slot_mapping,
+                self.kv_cache_dtype,
+                k_scale,
+                v_scale,
+            )
+
+        if prefill_meta := attn_metadata.prefill_metadata:
+
+            # Prompt run.
+            # normal attention
+            # When block_tables are not filled, it means q and k are the
+            # prompt, and they have the same length.
+
+            assert kv_cache.numel() == 0 \
+                    or prefill_meta.block_tables is None \
+                    or prefill_meta.block_tables.numel() == 0, \
+                "Does not support prefix-enabled attention."
+
+            output = self.bs_attn(
+                q=query,
+                k=key,
+                v=value,
+                cu_seqlens_q=prefill_meta.seq_start_loc,
+                cu_seqlens_k=prefill_meta.seq_start_loc,
+                sm_scale=self.scale,
+            )
+
+        if decode_meta := attn_metadata.decode_metadata:
+            # Decoding run.
+            output = PagedAttention.forward_decode(
+                query,
+                key_cache,
+                value_cache,
+                decode_meta.block_tables,
+                decode_meta.seq_lens_tensor,
+                self.blocksparse_params.max_seqlen,
+                self.kv_cache_dtype,
+                self.num_kv_heads,
+                self.scale,
+                self.alibi_slopes,
+                k_scale,
+                v_scale,
+                tp_rank=self.tp_rank,
+                blocksparse_local_blocks=self.local_blocks,
+                blocksparse_vert_stride=self.vert_stride,
+                blocksparse_block_size=self.sparse_block_size,
+                blocksparse_head_sliding_step=self.head_sliding_step,
+            )
+
+        # Reshape the output tensor.
+        return output.view(num_tokens, hidden_size)

vllm-v0.6.2/vllm/attention/backends/flash_attn.py

@@ -0,0 +1,989 @@
+"""Attention layer with FlashAttention."""
+from collections import defaultdict
+from dataclasses import dataclass
+from itertools import accumulate
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Type
+
+import torch
+
+from vllm import _custom_ops as ops
+from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
+                                              AttentionMetadata,
+                                              AttentionMetadataBuilder,
+                                              AttentionType)
+from vllm.attention.backends.utils import (
+    PAD_SLOT_ID, CommonAttentionState, compute_slot_mapping,
+    compute_slot_mapping_start_idx, get_num_prefill_decode_query_kv_tokens,
+    get_seq_len_block_table_args, is_all_cross_attn_metadata_set,
+    is_all_encoder_attn_metadata_set, is_block_tables_empty)
+from vllm.forward_context import get_forward_context
+from vllm.multimodal import MultiModalPlaceholderMap
+from vllm.utils import (async_tensor_h2d, direct_register_custom_op,
+                        make_tensor_with_pad)
+
+if TYPE_CHECKING:
+    from vllm.worker.model_runner import (ModelInputForGPUBuilder,
+                                          ModelInputForGPUWithSamplingMetadata)
+
+from vllm.vllm_flash_attn import (flash_attn_varlen_func,
+                                  flash_attn_with_kvcache)
+
+
+class FlashAttentionBackend(AttentionBackend):
+
+    @staticmethod
+    def get_supported_head_sizes() -> List[int]:
+        return [32, 64, 96, 128, 160, 192, 224, 256]
+
+    @staticmethod
+    def get_name() -> str:
+        return "FLASH_ATTN"
+
+    @staticmethod
+    def get_impl_cls() -> Type["FlashAttentionImpl"]:
+        return FlashAttentionImpl
+
+    @staticmethod
+    def get_metadata_cls() -> Type["AttentionMetadata"]:
+        return FlashAttentionMetadata
+
+    @staticmethod
+    def get_builder_cls() -> Type["FlashAttentionMetadataBuilder"]:
+        return FlashAttentionMetadataBuilder
+
+    @staticmethod
+    def get_state_cls() -> Type["CommonAttentionState"]:
+        return CommonAttentionState
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        if block_size % 16 != 0:
+            raise ValueError("Block size must be a multiple of 16.")
+        return (2, num_blocks, block_size, num_kv_heads, head_size)
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: torch.Tensor,
+    ) -> None:
+        src_key_cache = src_kv_cache[0]
+        dst_key_cache = dst_kv_cache[0]
+        ops.swap_blocks(src_key_cache, dst_key_cache, src_to_dst)
+        src_value_cache = src_kv_cache[1]
+        dst_value_cache = dst_kv_cache[1]
+        ops.swap_blocks(src_value_cache, dst_value_cache, src_to_dst)
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: torch.Tensor,
+    ) -> None:
+        key_caches = [kv_cache[0] for kv_cache in kv_caches]
+        value_caches = [kv_cache[1] for kv_cache in kv_caches]
+
+        ops.copy_blocks(key_caches, value_caches, src_to_dists)
+
+
+@dataclass
+class FlashAttentionMetadata(AttentionMetadata):
+    """Metadata for FlashAttentionBackend.
+
+    NOTE: Any python object stored here is not updated when it is
+    cuda-graph replayed. If you have values that need to be changed
+    dynamically, it should be stored in tensor. The tensor has to be
+    updated from `CUDAGraphRunner.forward` API.
+    """
+    # (batch_size,). The sequence length per sequence. Sequence length means
+    # the computed tokens + new tokens None if it is a decoding.
+    seq_lens: Optional[List[int]]
+    # seq_lens stored as a tensor.
+    seq_lens_tensor: Optional[torch.Tensor]
+
+    # NOTE(sang): Definition of context_len, query_len, and seq_len.
+    # |---------- N-1 iteration --------|
+    # |---------------- N iteration ---------------------|
+    # |- tokenA -|......................|-- newTokens ---|
+    # |---------- context_len ----------|
+    # |-------------------- seq_len ---------------------|
+    #                                   |-- query_len ---|
+
+    # Maximum sequence length among prefill batch. 0 if there are decoding
+    # requests only.
+    max_prefill_seq_len: int
+    # Maximum sequence length among decode batch. 0 if there are prefill
+    # requests only.
+    max_decode_seq_len: int
+    # (batch_size,) A tensor of context lengths (tokens that are computed
+    # so far).
+    context_lens_tensor: Optional[torch.Tensor]
+
+    # (batch_size, max_blocks_per_seq).
+    # Block addresses per sequence. (Seq id -> list of physical block)
+    # E.g., [0, 1, 2] means tokens are stored in 0th, 1st, and 2nd blocks
+    # in the kv cache. Each block can contain up to block_size tokens.
+    # 2nd dimensions are padded up to max_blocks_per_seq if it is cuda-graph
+    # captured.
+    block_tables: Optional[torch.Tensor]
+
+    # Whether or not if cuda graph is enabled.
+    # Cuda-graph is currently enabled for decoding only.
+    # TODO(woosuk): Move `use_cuda_graph` out since it's unrelated to attention.
+
+    use_cuda_graph: bool
+
+    # Maximum query length in the batch.
+    max_query_len: Optional[int] = None
+
+    # Max number of query tokens among request in the batch.
+    max_decode_query_len: Optional[int] = None
+
+    # (batch_size + 1,). The cumulative subquery lengths of the sequences in
+    # the batch, used to index into subquery. E.g., if the subquery length
+    # is [4, 6], it is [0, 4, 10].
+    query_start_loc: Optional[torch.Tensor] = None
+    # (batch_size + 1,). The cumulative sequence lengths of the sequences in
+    # the batch, used to index into sequence. E.g., if the sequence length is
+    # [4, 6], it is [0, 4, 10].
+    seq_start_loc: Optional[torch.Tensor] = None
+
+    _cached_prefill_metadata: Optional["FlashAttentionMetadata"] = None
+    _cached_decode_metadata: Optional["FlashAttentionMetadata"] = None
+
+    # Begin encoder attn & enc/dec cross-attn fields...
+
+    # Encoder sequence lengths representation
+    encoder_seq_lens: Optional[List[int]] = None
+    encoder_seq_lens_tensor: Optional[torch.Tensor] = None
+    # (batch_size + 1,). The cumulative sequence lengths of the sequences in
+    # the batch, used to index into sequence. E.g., if the sequence length is
+    # [4, 6], it is [0, 4, 10].
+    encoder_seq_start_loc: Optional[torch.Tensor] = None
+    # Maximum sequence length among encoder sequences
+    max_encoder_seq_len: Optional[int] = None
+    # Number of tokens input to encoder
+    num_encoder_tokens: Optional[int] = None
+
+    # Cross-attention memory-mapping data structures: slot mapping
+    # and block tables
+    cross_slot_mapping: Optional[torch.Tensor] = None
+    cross_block_tables: Optional[torch.Tensor] = None
+
+    @property
+    def is_all_encoder_attn_metadata_set(self):
+        '''
+        All attention metadata required for encoder attention is set.
+        '''
+        return is_all_encoder_attn_metadata_set(self)
+
+    @property
+    def is_all_cross_attn_metadata_set(self):
+        '''
+        All attention metadata required for enc/dec cross-attention is set.
+
+        Superset of encoder attention required metadata.
+        '''
+        return is_all_cross_attn_metadata_set(self)
+
+    @property
+    def prefill_metadata(self) -> Optional["FlashAttentionMetadata"]:
+        if self.num_prefills == 0:
+            return None
+
+        if self._cached_prefill_metadata is not None:
+            return self._cached_prefill_metadata
+
+        assert ((self.seq_lens is not None)
+                or (self.encoder_seq_lens is not None))
+        assert ((self.seq_lens_tensor is not None)
+                or (self.encoder_seq_lens_tensor is not None))
+
+        # Compute some attn_metadata fields which default to None
+        query_start_loc = (None if self.query_start_loc is None else
+                           self.query_start_loc[:self.num_prefills + 1])
+        slot_mapping = (None if self.slot_mapping is None else
+                        self.slot_mapping[:self.num_prefill_tokens])
+        seq_lens = (None if self.seq_lens is None else
+                    self.seq_lens[:self.num_prefills])
+        seq_lens_tensor = (None if self.seq_lens_tensor is None else
+                           self.seq_lens_tensor[:self.num_prefills])
+        seq_start_loc = (None if self.seq_start_loc is None else
+                         self.seq_start_loc[:self.num_prefills + 1])
+        context_lens_tensor = (None if self.context_lens_tensor is None else
+                               self.context_lens_tensor[:self.num_prefills])
+        block_tables = (None if self.block_tables is None else
+                        self.block_tables[:self.num_prefills])
+
+        self._cached_prefill_metadata = FlashAttentionMetadata(
+            num_prefills=self.num_prefills,
+            num_prefill_tokens=self.num_prefill_tokens,
+            num_decode_tokens=0,
+            slot_mapping=slot_mapping,
+            multi_modal_placeholder_index_maps=self.
+            multi_modal_placeholder_index_maps,
+            seq_lens=seq_lens,
+            seq_lens_tensor=seq_lens_tensor,
+            max_query_len=self.max_query_len,
+            max_prefill_seq_len=self.max_prefill_seq_len,
+            max_decode_query_len=0,
+            max_decode_seq_len=0,
+            query_start_loc=query_start_loc,
+            seq_start_loc=seq_start_loc,
+            context_lens_tensor=context_lens_tensor,
+            block_tables=block_tables,
+            use_cuda_graph=False,
+            # Begin encoder & cross attn fields below...
+            encoder_seq_lens=self.encoder_seq_lens,
+            encoder_seq_lens_tensor=self.encoder_seq_lens_tensor,
+            encoder_seq_start_loc=self.encoder_seq_start_loc,
+            max_encoder_seq_len=self.max_encoder_seq_len,
+            cross_slot_mapping=self.cross_slot_mapping,
+            cross_block_tables=self.cross_block_tables)
+        return self._cached_prefill_metadata
+
+    @property
+    def decode_metadata(self) -> Optional["FlashAttentionMetadata"]:
+        if self.num_decode_tokens == 0:
+            return None
+
+        if self._cached_decode_metadata is not None:
+            return self._cached_decode_metadata
+        assert ((self.seq_lens_tensor is not None)
+                or (self.encoder_seq_lens_tensor is not None))
+
+        # Compute some attn_metadata fields which default to None
+        slot_mapping = (None if self.slot_mapping is None else
+                        self.slot_mapping[self.num_prefill_tokens:])
+        seq_lens_tensor = (None if self.seq_lens_tensor is None else
+                           self.seq_lens_tensor[self.num_prefills:])
+        block_tables = (None if self.block_tables is None else
+                        self.block_tables[self.num_prefills:])
+
+        self._cached_decode_metadata = FlashAttentionMetadata(
+            num_prefills=0,
+            num_prefill_tokens=0,
+            num_decode_tokens=self.num_decode_tokens,
+            slot_mapping=slot_mapping,
+            multi_modal_placeholder_index_maps=None,
+            seq_lens=None,
+            seq_lens_tensor=seq_lens_tensor,
+            max_decode_query_len=self.max_decode_query_len,
+            max_query_len=self.max_query_len,
+            max_prefill_seq_len=0,
+            max_decode_seq_len=self.max_decode_seq_len,
+            # Batch may be composed of prefill|decodes, adjust query start
+            # indices to refer to the start of decodes. E.g.
+            # in tokens:[3 prefills|6 decodes], query_start_loc=[3,9] => [0,6].
+            query_start_loc=(self.query_start_loc[self.num_prefills:] -
+                             self.query_start_loc[self.num_prefills])
+            if self.query_start_loc is not None else None,
+            seq_start_loc=self.seq_start_loc[self.num_prefills:]
+            if self.seq_start_loc is not None else None,
+            context_lens_tensor=None,
+            block_tables=block_tables,
+            use_cuda_graph=self.use_cuda_graph,
+            # Begin encoder & cross attn fields below...
+            encoder_seq_lens=self.encoder_seq_lens,
+            encoder_seq_lens_tensor=self.encoder_seq_lens_tensor,
+            encoder_seq_start_loc=self.encoder_seq_start_loc,
+            max_encoder_seq_len=self.max_encoder_seq_len,
+            cross_slot_mapping=self.cross_slot_mapping,
+            cross_block_tables=self.cross_block_tables)
+        return self._cached_decode_metadata
+
+    def advance_step(self,
+                     model_input: "ModelInputForGPUWithSamplingMetadata",
+                     sampled_token_ids: Optional[torch.Tensor],
+                     block_size: int,
+                     num_seqs: int,
+                     num_queries: int,
+                     turn_prefills_into_decodes: bool = False):
+        """
+        Update metadata in-place to advance one decode step.
+        """
+        # When using cudagraph, the num_seqs is padded to the next captured
+        # batch sized, but num_queries tracks the actual number of requests in
+        # the batch. For --enforce-eager mode, num_seqs == num_queries
+        if num_seqs != num_queries:
+            assert num_seqs > num_queries
+            assert self.use_cuda_graph
+
+        if turn_prefills_into_decodes:
+            # When Mutli-Step is enabled with Chunked-Prefill, prefills and
+            # decodes are scheduled together. In the first step, all the
+            # prefills turn into decodes. This update reflects that
+            # conversion.
+            assert self.num_decode_tokens + self.num_prefills == num_seqs
+            self.num_decode_tokens += self.num_prefills
+            self.num_prefills = 0
+            self.num_prefill_tokens = 0
+            self.max_prefill_seq_len = 0
+            self.max_query_len = 1
+
+            self.slot_mapping = self.slot_mapping[:num_seqs]
+        else:
+            assert self.seq_lens is not None
+            assert self.max_decode_seq_len == max(self.seq_lens)
+
+        assert self.num_prefills == 0
+        assert self.num_prefill_tokens == 0
+        assert self.num_decode_tokens == num_seqs
+        assert self.slot_mapping.shape == (num_seqs, )
+
+        assert self.seq_lens is not None
+        assert len(self.seq_lens) == num_seqs
+        assert self.seq_lens_tensor is not None
+        assert self.seq_lens_tensor.shape == (num_seqs, )
+        assert self.max_query_len == 1
+        assert self.max_prefill_seq_len == 0
+
+        assert self.query_start_loc is not None
+        assert self.query_start_loc.shape == (num_queries + 1, )
+        assert self.seq_start_loc is not None
+        assert self.seq_start_loc.shape == (num_seqs + 1, )
+
+        assert self.context_lens_tensor is not None
+        assert self.context_lens_tensor.shape == (num_queries, )
+
+        assert self.block_tables is not None
+        assert self.block_tables.shape[0] == num_seqs
+
+        # Update query lengths. Note that we update only queries and not seqs,
+        # since tensors may be padded due to captured cuda graph batch size
+        for i in range(num_queries):
+            self.seq_lens[i] += 1
+        self.max_decode_seq_len = max(self.seq_lens)
+
+        ops.advance_step_flashattn(num_seqs=num_seqs,
+                                   num_queries=num_queries,
+                                   block_size=block_size,
+                                   input_tokens=model_input.input_tokens,
+                                   sampled_token_ids=sampled_token_ids,
+                                   input_positions=model_input.input_positions,
+                                   seq_lens=self.seq_lens_tensor,
+                                   slot_mapping=self.slot_mapping,
+                                   block_tables=self.block_tables)
+
+
+class FlashAttentionMetadataBuilder(
+        AttentionMetadataBuilder[FlashAttentionMetadata]):
+
+    def __init__(self, input_builder: "ModelInputForGPUBuilder"):
+        self.slot_mapping: List[int] = []
+        self.prefill_seq_lens: List[int] = []
+        self.context_lens: List[int] = []
+        self.block_tables: List[List[int]] = []
+        self.curr_seq_lens: List[int] = []
+        self.multimodal_placeholder_maps: Dict[
+            str,
+            MultiModalPlaceholderMap] = defaultdict(MultiModalPlaceholderMap)
+        self.num_prefills = 0
+        self.num_prefill_tokens = 0
+        self.num_decode_tokens = 0
+        self.has_prefix_cache_hit = False
+
+        self.input_builder = input_builder
+        self.runner = input_builder.runner
+        self.sliding_window = input_builder.sliding_window
+        self.block_size = input_builder.block_size
+
+    def _add_seq_group(
+            self, inter_data: "ModelInputForGPUBuilder.InterDataForSeqGroup",
+            chunked_prefill_enabled: bool, prefix_cache_hit: bool):
+        """Add a sequence group to the metadata. Specifically update/append
+        1. context length.
+        2. block table.
+        3. slot mapping.
+        """
+        is_prompt = inter_data.is_prompt
+        block_tables = inter_data.block_tables
+
+        for (seq_id, token_len, seq_len, curr_seq_len, query_len, context_len,
+             curr_sliding_window_block) in zip(
+                 inter_data.seq_ids, [len(t) for t in inter_data.input_tokens],
+                 inter_data.orig_seq_lens, inter_data.seq_lens,
+                 inter_data.query_lens, inter_data.context_lens,
+                 inter_data.curr_sliding_window_blocks):
+            self.context_lens.append(context_len)
+
+            if is_prompt:
+                mm_maps = inter_data.multi_modal_placeholder_maps
+                if mm_maps:
+                    for modality, placeholders in mm_maps.items():
+                        self.multimodal_placeholder_maps[modality].extend(
+                            placeholders)
+
+                self.num_prefills += 1
+                self.num_prefill_tokens += token_len
+                self.prefill_seq_lens.append(seq_len)
+            else:
+                self.num_decode_tokens += query_len
+                self.curr_seq_lens.append(curr_seq_len)
+
+            # Compute block table.
+            # TODO(sang): Combine chunked prefill and prefix caching by
+            # only allowing multiple of block_size chunk size.
+            # NOTE: This only works for oooooooxxx style attention.
+            block_table = []
+            if prefix_cache_hit:
+                # NOTE(woosuk): For flash-attn, the block table should
+                # include the entries for the incoming prefill tokens.
+                block_table = block_tables[seq_id]
+            elif ((chunked_prefill_enabled or not is_prompt)
+                  and block_tables is not None):
+                if curr_sliding_window_block == 0:
+                    block_table = block_tables[seq_id]
+                else:
+                    block_table = block_tables[seq_id][
+                        -curr_sliding_window_block:]
+            self.block_tables.append(block_table)
+
+            # Compute slot mapping.
+            is_profile_run = is_block_tables_empty(block_tables)
+            start_idx = compute_slot_mapping_start_idx(is_prompt, query_len,
+                                                       context_len,
+                                                       self.sliding_window)
+            compute_slot_mapping(is_profile_run, self.slot_mapping, seq_id,
+                                 seq_len, context_len, start_idx,
+                                 self.block_size, inter_data.block_tables)
+
+    def _get_graph_runner_block_tables(
+            self, num_seqs: int,
+            block_tables: List[List[int]]) -> torch.Tensor:
+        # The shape of graph_block_tables is
+        # [max batch size, max context len // block size].
+        max_batch_size, max_blocks = self.runner.graph_block_tables.shape
+        assert max_batch_size >= num_seqs
+
+        graph_block_tables = self.runner.graph_block_tables[:num_seqs]
+        for i, block_table in enumerate(block_tables):
+            if block_table:
+                num_blocks = len(block_table)
+                if num_blocks <= max_blocks:
+                    graph_block_tables[i, :num_blocks] = block_table
+                else:
+                    # It may be possible to have more blocks allocated due
+                    # to lookahead slots of multi-step, however, they are
+                    # not used anyway, so can be safely ignored.
+                    graph_block_tables[
+                        i, :max_blocks] = block_table[:max_blocks]
+
+        return torch.from_numpy(graph_block_tables).to(
+            device=self.runner.device, non_blocking=True)
+
+    def build(self, seq_lens: List[int], query_lens: List[int],
+              cuda_graph_pad_size: int, batch_size: int):
+        """Build attention metadata with on-device tensors.
+
+        Args:
+            seq_lens: The maybe padded sequence lengths of the input sequences.
+            query_lens: The query lengths of the input sequences.
+            cuda_graph_pad_size: The padding size for cuda graph.
+                                 -1 if cuda graph is not used.
+            batch_size: The maybe padded batch size.
+        """
+        prefix_cache_hit = any([
+            inter_data.prefix_cache_hit
+            for inter_data in self.input_builder.inter_data_list
+        ])
+        for inter_data in self.input_builder.inter_data_list:
+            self._add_seq_group(inter_data,
+                                self.input_builder.chunked_prefill_enabled,
+                                prefix_cache_hit)
+
+        device = self.runner.device
+        use_captured_graph = cuda_graph_pad_size != -1
+
+        max_query_len = max(query_lens)
+        decode_query_lens = query_lens[self.num_prefills:]
+        if len(decode_query_lens) > 0:
+            max_decode_query_len = max(decode_query_lens)
+        else:
+            max_decode_query_len = 1
+        max_prefill_seq_len = max(self.prefill_seq_lens, default=0)
+        max_decode_seq_len = max(self.curr_seq_lens, default=0)
+        num_decode_tokens = self.num_decode_tokens
+        query_start_loc = list(accumulate(query_lens, initial=0))
+        seq_start_loc = list(accumulate(seq_lens, initial=0))
+
+        num_seqs = len(seq_lens)
+        if use_captured_graph:
+            self.slot_mapping.extend([PAD_SLOT_ID] * cuda_graph_pad_size)
+            self.block_tables.extend([] * cuda_graph_pad_size)
+            num_decode_tokens = batch_size - self.num_prefill_tokens
+            block_tables = self._get_graph_runner_block_tables(
+                num_seqs, self.block_tables)
+        else:
+            block_tables = make_tensor_with_pad(
+                self.block_tables,
+                pad=0,
+                dtype=torch.int,
+                device=device,
+            )
+        assert max_query_len > 0, ("query_lens: {}".format(query_lens))
+
+        assert device is not None
+        context_lens_tensor = async_tensor_h2d(self.context_lens, torch.int,
+                                               device, self.runner.pin_memory)
+        seq_lens_tensor = async_tensor_h2d(seq_lens, torch.int, device,
+                                           self.runner.pin_memory)
+        slot_mapping_tensor = async_tensor_h2d(self.slot_mapping, torch.long,
+                                               device, self.runner.pin_memory)
+        query_start_loc_tensor = async_tensor_h2d(query_start_loc, torch.int32,
+                                                  device,
+                                                  self.runner.pin_memory)
+        seq_start_loc_tensor = async_tensor_h2d(seq_start_loc, torch.int32,
+                                                device, self.runner.pin_memory)
+        placeholder_index_maps = {
+            modality: placeholder_map.index_map()
+            for modality, placeholder_map in
+            self.multimodal_placeholder_maps.items()
+        }
+
+        return FlashAttentionMetadata(
+            num_prefills=self.num_prefills,
+            slot_mapping=slot_mapping_tensor,
+            num_prefill_tokens=self.num_prefill_tokens,
+            num_decode_tokens=num_decode_tokens,
+            seq_lens=seq_lens,
+            multi_modal_placeholder_index_maps=placeholder_index_maps,
+            seq_lens_tensor=seq_lens_tensor,
+            max_query_len=max_query_len,
+            max_decode_query_len=max_decode_query_len,
+            max_prefill_seq_len=max_prefill_seq_len,
+            max_decode_seq_len=max_decode_seq_len,
+            query_start_loc=query_start_loc_tensor,
+            seq_start_loc=seq_start_loc_tensor,
+            context_lens_tensor=context_lens_tensor,
+            block_tables=block_tables,
+            use_cuda_graph=use_captured_graph,
+        )
+
+
+class FlashAttentionImpl(AttentionImpl):
+    """
+    If the input tensors contain prompt tokens, the layout is as follows:
+    |<--------------- num_prefill_tokens ----------------->|	
+    |<--prefill_0-->|<--prefill_1-->|...|<--prefill_N-1--->|
+
+    Otherwise, the layout is as follows:	
+    |<----------------- num_decode_tokens ------------------>|	
+    |<--decode_0-->|..........|<--decode_M-1-->|<--padding-->|
+
+    Generation tokens can contain padding when cuda-graph is used.
+    Currently, prompt tokens don't contain any padding.
+
+    The prompts might have different lengths, while the generation tokens
+    always have length 1.
+
+    If chunked prefill is enabled, prefill tokens and decode tokens can be
+    batched together in a flattened 1D query.
+
+    |<----- num_prefill_tokens ---->|<------- num_decode_tokens --------->|
+    |<-prefill_0->|...|<-prefill_N-1->|<--decode_0-->|...|<--decode_M-1-->|
+
+    Currently, cuda graph is disabled for chunked prefill, meaning there's no
+    padding between prefill and decode tokens.
+    """
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: Optional[List[float]],
+        sliding_window: Optional[int],
+        kv_cache_dtype: str,
+        blocksparse_params: Optional[Dict[str, Any]] = None,
+        logits_soft_cap: Optional[float] = None,
+    ) -> None:
+        if blocksparse_params is not None:
+            raise ValueError(
+                "FlashAttention does not support block-sparse attention.")
+        self.num_heads = num_heads
+        self.head_size = head_size
+        self.scale = float(scale)
+        self.num_kv_heads = num_kv_heads
+        if alibi_slopes is not None:
+            alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
+        self.alibi_slopes = alibi_slopes
+        self.sliding_window = ((sliding_window - 1,
+                                0) if sliding_window is not None else (-1, -1))
+        self.kv_cache_dtype = kv_cache_dtype
+        if logits_soft_cap is None:
+            # In flash-attn, setting logits_soft_cap as 0 means no soft cap.
+            logits_soft_cap = 0
+        self.logits_soft_cap = logits_soft_cap
+
+        assert self.num_heads % self.num_kv_heads == 0
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+
+        support_head_sizes = FlashAttentionBackend.get_supported_head_sizes()
+        if head_size not in support_head_sizes:
+            raise ValueError(
+                f"Head size {head_size} is not supported by FlashAttention. "
+                f"Supported head sizes are: {support_head_sizes}.")
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: FlashAttentionMetadata,
+        k_scale: float = 1.0,
+        v_scale: float = 1.0,
+        attn_type: AttentionType = AttentionType.DECODER,
+    ) -> torch.Tensor:
+        """Forward pass with FlashAttention.
+
+        Args:
+            query: shape = [num_tokens, num_heads * head_size]
+            key: shape = [num_tokens, num_kv_heads * head_size]
+            value: shape = [num_tokens, num_kv_heads * head_size]
+            kv_cache = [2, num_blocks, block_size, num_kv_heads, head_size]
+                NOTE: kv_cache will be an empty tensor with shape [0]
+                for profiling run.
+            attn_metadata: Metadata for attention.
+        Returns:
+            shape = [num_tokens, num_heads * head_size]
+        """
+        # NOTE(woosuk): FlashAttention does not support FP8 KV cache.
+        assert k_scale == 1.0 and v_scale == 1.0, (
+            "key/v_scale is not supported in FlashAttention.")
+
+        if (attn_type == AttentionType.ENCODER
+                and (not attn_metadata.is_all_encoder_attn_metadata_set)):
+            raise AttributeError("Encoder attention requires setting "
+                                 "encoder metadata attributes.")
+        elif (attn_type == AttentionType.ENCODER_DECODER
+              and (not attn_metadata.is_all_cross_attn_metadata_set)):
+            raise AttributeError("Encoder/decoder cross-attention "
+                                 "requires setting cross-attention "
+                                 "metadata attributes.")
+
+        output = torch.ops.vllm.unified_flash_attention(
+            query,
+            key,
+            value,
+            self.num_heads,
+            self.head_size,
+            self.num_kv_heads,
+            kv_cache,
+            self.kv_cache_dtype,
+            k_scale,
+            v_scale,
+            self.scale,
+            attn_type.value,
+            self.sliding_window,
+            self.alibi_slopes,
+            self.logits_soft_cap,
+        )
+
+        return output
+
+
+def _get_query_key_seq_metadata(
+    attn_metadata,
+    is_prompt: bool,
+    attn_type: AttentionType,
+) -> tuple:
+    """
+    Returns sequence metadata for key and query based on the specified 
+    attention type and whether input is a prompt.
+
+    This function computes the starting locations and maximum sequence lengths 
+    for key and query sequences for different attention types.
+
+    Args:
+        attn_metadata: The attention metadata object
+        is_prompt (bool): A flag indicating if the input is a prompt
+        attn_type (AttentionType): The type of attention being used.
+
+    Returns:
+        tuple: A tuple containing four integers:
+            - Starting location for the query sequence.
+            - Maximum sequence length for the query sequence.
+            - Starting location for the key sequence.
+            - Maximum sequence length for the key sequence.
+
+    Raises:
+        AttributeError: If an invalid attention type is provided.
+    """
+    if attn_type == AttentionType.DECODER:
+        # Decoder self-attention
+        # Choose max_seq_len based on whether we are in prompt_run
+        if is_prompt:
+            max_seq_len = attn_metadata.max_prefill_seq_len
+        else:
+            max_seq_len = attn_metadata.max_decode_seq_len
+        return (attn_metadata.seq_start_loc, max_seq_len,
+                attn_metadata.seq_start_loc, max_seq_len)
+
+    elif attn_type == AttentionType.ENCODER_DECODER:
+        # This is cross attention between the where the key
+        # is the precomputed encoder attention and query
+        # is the input sequence.
+        # Choose query max length based on whether it is prompt
+        # or not.
+        if is_prompt:
+            max_seq_len = attn_metadata.max_prefill_seq_len
+        else:
+            max_seq_len = attn_metadata.max_decode_seq_len
+        return (attn_metadata.seq_start_loc, max_seq_len,
+                attn_metadata.encoder_seq_start_loc,
+                attn_metadata.max_encoder_seq_len)
+    elif attn_type == AttentionType.ENCODER:
+        # For encoder attention both the query and the key are same i.e the
+        # encoder sequence.
+        return (attn_metadata.encoder_seq_start_loc,
+                attn_metadata.max_encoder_seq_len,
+                attn_metadata.encoder_seq_start_loc,
+                attn_metadata.max_encoder_seq_len)
+    elif attn_type == AttentionType.ENCODER_ONLY:
+        assert is_prompt, "Should not have decode for encoder only model."
+        return (attn_metadata.seq_start_loc, attn_metadata.max_prefill_seq_len,
+                attn_metadata.seq_start_loc, attn_metadata.max_prefill_seq_len)
+    else:
+        raise AttributeError(f"Invalid attention type {str(attn_type)}")
+
+
+def _get_causal_option(attn_type: AttentionType) -> bool:
+    """
+    Determine whether the given attention type is suitable for causal 
+    attention mechanisms.
+
+    Args:
+        attn_type (AttentionType): The type of attention being evaluated
+
+    Returns:
+        bool: Returns `True` if the attention type is suitable for causal 
+        attention (i.e., not encoder, encoder-only, or encoder-decoder), 
+        otherwise returns `False`.
+    """
+    return not (attn_type == AttentionType.ENCODER
+                or attn_type == AttentionType.ENCODER_ONLY
+                or attn_type == AttentionType.ENCODER_DECODER)
+
+
+def unified_flash_attention(
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    num_heads: int,
+    head_size: int,
+    num_kv_heads: int,
+    kv_cache: torch.Tensor,
+    kv_cache_dtype: str,
+    k_scale: float,
+    v_scale: float,
+    softmax_scale: float,
+    attn_type_int_val: int,
+    window_size: Optional[List[int]] = None,
+    alibi_slopes: Optional[torch.Tensor] = None,
+    logits_soft_cap: Optional[float] = None,
+) -> torch.Tensor:
+
+    # Convert integer attn_type to enum
+    try:
+        attn_type = AttentionType(attn_type_int_val)
+    except ValueError as err:
+        raise AttributeError(
+            f"Invalid attention type {str(attn_type_int_val)}") from err
+
+    current_metadata = get_forward_context()
+    assert current_metadata is not None
+    assert isinstance(current_metadata, FlashAttentionMetadata)
+    attn_metadata: FlashAttentionMetadata = current_metadata
+
+    num_tokens, hidden_size = query.shape
+
+    # Reshape the query, key, and value tensors.
+    query = query.view(-1, num_heads, head_size)
+    if (key is not None) and (value is not None):
+        key = key.view(-1, num_kv_heads, head_size)
+        value = value.view(-1, num_kv_heads, head_size)
+
+    if kv_cache.numel() > 0:
+        key_cache = kv_cache[0]
+        value_cache = kv_cache[1]
+        # We skip updating the KV cache under two conditions:
+        #  a. When the Attention Type is ENCODER. In this phase, we compute
+        #     only the encoder attention without updating the cache.
+        #  b. When both Key and Value are None. This occurs during
+        #     cross-attention computation in the decoding phase, where the KV
+        #     cache is already populated with the cross-attention tensor.
+        #     Thus, we skip cache updates during this time.
+        if (attn_type != AttentionType.ENCODER) and (key is not None) and (
+                value is not None):
+            if attn_type == AttentionType.ENCODER_DECODER:
+                # Update cross-attention KV cache (prefill-only)
+                updated_slot_mapping = attn_metadata.cross_slot_mapping
+            else:
+                # Update self-attention KV cache (prefill/decode)
+                updated_slot_mapping = attn_metadata.slot_mapping
+
+            # Reshape the input keys and values and store them in the cache.
+            # If kv_cache is not provided, the new key and value tensors are
+            # not cached. This happens during the initial memory profiling run.
+            torch.ops._C_cache_ops.reshape_and_cache_flash(
+                key,
+                value,
+                kv_cache[0],
+                kv_cache[1],
+                updated_slot_mapping.flatten(),  # type: ignore[union-attr]
+                kv_cache_dtype,
+                k_scale,
+                v_scale,
+            )
+
+    (num_prefill_query_tokens, num_prefill_kv_tokens,
+    num_decode_query_tokens) = \
+        get_num_prefill_decode_query_kv_tokens(attn_metadata, attn_type)
+    decode_query = query[num_prefill_query_tokens:]
+    # QKV for prefill.
+    query = query[:num_prefill_query_tokens]
+    assert query.shape[0] == num_prefill_query_tokens
+    assert decode_query.shape[0] == num_decode_query_tokens
+
+    prefill_output: Optional[torch.Tensor] = None
+    decode_output: Optional[torch.Tensor] = None
+    if prefill_meta := attn_metadata.prefill_metadata:
+        # Prompt run.
+        if (kv_cache.numel() == 0 or prefill_meta.block_tables is None
+                or prefill_meta.block_tables.numel() == 0):
+            # normal attention
+            # When block_tables are not filled, it means q and k are the
+            # prompt, and they have the same length.
+            q_seq_start_loc, q_seq_len, k_seq_start_loc, k_seq_len = \
+                _get_query_key_seq_metadata(prefill_meta, True, attn_type)
+
+            key = key[:num_prefill_kv_tokens]
+            value = value[:num_prefill_kv_tokens]
+
+            prefill_output = flash_attn_varlen_func(
+                q=query,
+                k=key,
+                v=value,
+                cu_seqlens_q=q_seq_start_loc,
+                cu_seqlens_k=k_seq_start_loc,
+                max_seqlen_q=q_seq_len,
+                max_seqlen_k=k_seq_len,
+                softmax_scale=softmax_scale,
+                causal=_get_causal_option(attn_type),
+                window_size=window_size,
+                alibi_slopes=alibi_slopes,
+                softcap=logits_soft_cap,
+            )
+        else:
+            # prefix-enabled attention
+            assert attn_type == AttentionType.DECODER, (
+                "Only decoder-only models support prefix caching")
+            assert prefill_meta.seq_lens is not None
+            max_seq_len = max(prefill_meta.seq_lens)
+            prefill_output = flash_attn_varlen_func(  # noqa
+                q=query,
+                k=key_cache,
+                v=value_cache,
+                cu_seqlens_q=prefill_meta.query_start_loc,
+                max_seqlen_q=prefill_meta.max_query_len,
+                cu_seqlens_k=prefill_meta.seq_start_loc,
+                max_seqlen_k=max_seq_len,
+                softmax_scale=softmax_scale,
+                causal=True,
+                window_size=window_size,
+                alibi_slopes=alibi_slopes,
+                block_table=prefill_meta.block_tables,
+                softcap=logits_soft_cap,
+            )
+
+    if decode_meta := attn_metadata.decode_metadata:
+        # Decoding run.
+        # Use flash_attn_varlen_func kernel for speculative decoding
+        # because different queries might have different lengths.
+
+        assert decode_meta.max_decode_query_len is not None
+        # use only for actual varlen decoding
+        if decode_meta.max_decode_query_len > 1:
+            assert attn_type == AttentionType.DECODER, (
+                "Only decoder-only models support max_decode_query_len > 1")
+            decode_output = flash_attn_varlen_func(
+                q=decode_query,
+                k=key_cache,
+                v=value_cache,
+                cu_seqlens_q=decode_meta.query_start_loc,
+                max_seqlen_q=decode_meta.max_decode_query_len,
+                cu_seqlens_k=decode_meta.seq_start_loc,
+                max_seqlen_k=decode_meta.max_decode_seq_len,
+                softmax_scale=softmax_scale,
+                causal=True,
+                window_size=window_size,
+                alibi_slopes=alibi_slopes,
+                softcap=logits_soft_cap,
+                block_table=decode_meta.block_tables,
+            )
+        else:
+            # Use flash_attn_with_kvcache for normal decoding.
+            (
+                seq_lens_arg,
+                _,
+                block_tables_arg,
+            ) = get_seq_len_block_table_args(decode_meta, False, attn_type)
+            decode_output = flash_attn_with_kvcache(
+                q=decode_query.unsqueeze(1),
+                k_cache=key_cache,
+                v_cache=value_cache,
+                block_table=block_tables_arg,
+                cache_seqlens=seq_lens_arg,
+                softmax_scale=softmax_scale,
+                causal=True,
+                window_size=window_size,
+                alibi_slopes=alibi_slopes,
+                softcap=logits_soft_cap,
+            ).squeeze(1)
+
+    if prefill_output is None:
+        assert decode_output is not None
+        return decode_output.view(num_decode_query_tokens, hidden_size)
+    if decode_output is None:
+        assert prefill_output is not None
+        return prefill_output.view(num_prefill_query_tokens, hidden_size)
+
+    assert decode_meta is not None
+    decode_output = decode_output.squeeze(1)
+    output = torch.cat([prefill_output, decode_output], dim=0)
+    return output.view(num_tokens, hidden_size)
+
+
+def unified_flash_attention_fake(
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    num_heads: int,
+    head_size: int,
+    num_kv_heads: int,
+    kv_cache: torch.Tensor,
+    kv_cache_dtype: str,
+    k_scale: float,
+    v_scale: float,
+    softmax_scale: float,
+    attn_type_int_val: int,
+    window_size: Optional[List[int]] = None,
+    alibi_slopes: Optional[torch.Tensor] = None,
+    logits_soft_cap: Optional[float] = None,
+) -> torch.Tensor:
+    return torch.empty_like(query)
+
+
+direct_register_custom_op(
+    op_name="unified_flash_attention",
+    op_func=unified_flash_attention,
+    mutates_args=["kv_cache"],
+    fake_impl=unified_flash_attention_fake,
+)

vllm-v0.6.2/vllm/attention/backends/flashinfer.py

@@ -0,0 +1,952 @@
+from collections import defaultdict
+from contextlib import contextmanager
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Set, Tuple, Type
+
+from vllm.multimodal import MultiModalPlaceholderMap
+
+try:
+    from flashinfer import BatchDecodeWithPagedKVCacheWrapper
+    from flashinfer.decode import CUDAGraphBatchDecodeWithPagedKVCacheWrapper
+    from flashinfer.prefill import BatchPrefillWithPagedKVCacheWrapper
+
+    from vllm.vllm_flash_attn import flash_attn_varlen_func
+    FLASHINFER_WORKSPACE_BUFFER_SIZE = 256 * 1024 * 1024
+except ImportError:
+    BatchDecodeWithPagedKVCacheWrapper = None
+    CUDAGraphBatchDecodeWithPagedKVCacheWrapper = None
+    BatchPrefillWithPagedKVCacheWrapper = None
+    FLASHINFER_WORKSPACE_BUFFER_SIZE = 0
+
+import torch
+
+import vllm.envs as envs
+from vllm import _custom_ops as ops
+from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
+                                              AttentionMetadata,
+                                              AttentionMetadataBuilder,
+                                              AttentionState, AttentionType)
+from vllm.attention.backends.utils import (PAD_SLOT_ID, compute_slot_mapping,
+                                           compute_slot_mapping_start_idx,
+                                           is_block_tables_empty)
+from vllm.attention.ops.paged_attn import PagedAttention
+from vllm.forward_context import get_forward_context
+from vllm.utils import (async_tensor_h2d, direct_register_custom_op,
+                        get_kv_cache_torch_dtype, make_tensor_with_pad)
+
+if TYPE_CHECKING:
+    from vllm.worker.model_runner import (ModelInputForGPUBuilder,
+                                          ModelInputForGPUWithSamplingMetadata)
+
+
+class FlashInferBackend(AttentionBackend):
+
+    @staticmethod
+    def get_name() -> str:
+        return "FLASHINFER"
+
+    @staticmethod
+    def get_impl_cls() -> Type["FlashInferImpl"]:
+        return FlashInferImpl
+
+    @staticmethod
+    def get_metadata_cls() -> Type["AttentionMetadata"]:
+        return FlashInferMetadata
+
+    @staticmethod
+    def get_builder_cls() -> Type["FlashInferMetadataBuilder"]:
+        return FlashInferMetadataBuilder
+
+    @staticmethod
+    def get_state_cls() -> Type["FlashInferState"]:
+        return FlashInferState
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        return (num_blocks, 2, block_size, num_kv_heads, head_size)
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: torch.Tensor,
+    ) -> None:
+        PagedAttention.swap_blocks(src_kv_cache, dst_kv_cache, src_to_dst)
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: torch.Tensor,
+    ) -> None:
+        PagedAttention.copy_blocks(kv_caches, src_to_dists)
+
+    @staticmethod
+    def get_supported_head_sizes() -> List[int]:
+        return [64, 128, 256]
+
+    @staticmethod
+    def get_fp8_dtype_for_flashinfer(kv_cache_dtype: str) -> torch.dtype:
+        if kv_cache_dtype in ("fp8", "fp8_e4m3"):
+            return torch.float8_e4m3fn
+        elif kv_cache_dtype == "fp8_e5m2":
+            return torch.float8_e5m2
+        else:
+            raise ValueError(f"Unrecognized FP8 dtype: {kv_cache_dtype}")
+
+
+class FlashInferState(AttentionState):
+
+    def __init__(self, runner):
+        self.runner = runner
+        self._is_graph_capturing = False
+        self._workspace_buffer = None
+        self._decode_wrapper = None
+        self._prefill_wrapper = None
+
+    def _get_workspace_buffer(self):
+        if self._workspace_buffer is None:
+            self._workspace_buffer = torch.empty(
+                FLASHINFER_WORKSPACE_BUFFER_SIZE,
+                dtype=torch.uint8,
+                device=self.runner.device)
+        return self._workspace_buffer
+
+    def _get_prefill_wrapper(self):
+        if self._prefill_wrapper is None:
+            self._prefill_wrapper = BatchPrefillWithPagedKVCacheWrapper(
+                self._get_workspace_buffer(), "NHD")
+        return self._prefill_wrapper
+
+    def _get_decode_wrapper(self):
+        if self._decode_wrapper is None:
+            num_qo_heads = (self.runner.model_config.get_num_attention_heads(
+                self.runner.parallel_config))
+            num_kv_heads = self.runner.model_config.get_num_kv_heads(
+                self.runner.parallel_config)
+            use_tensor_cores = envs.VLLM_FLASHINFER_FORCE_TENSOR_CORES or (
+                num_qo_heads // num_kv_heads > 4)
+            self._decode_wrapper = BatchDecodeWithPagedKVCacheWrapper(
+                self._get_workspace_buffer(),
+                "NHD",
+                use_tensor_cores=use_tensor_cores)
+        return self._decode_wrapper
+
+    @contextmanager
+    def graph_capture(self, max_batch_size: int):
+        self._is_graph_capturing = True
+        self._graph_decode_wrapper = None
+        self._graph_slot_mapping = torch.full((max_batch_size, ),
+                                              PAD_SLOT_ID,
+                                              dtype=torch.long,
+                                              device=self.runner.device)
+        self._graph_seq_lens = torch.ones(max_batch_size,
+                                          dtype=torch.int32,
+                                          device=self.runner.device)
+        self._graph_block_tables = torch.from_numpy(
+            self.runner.graph_block_tables).to(device=self.runner.device)
+        self._graph_decode_workspace_buffer = self._get_workspace_buffer()
+        self._graph_indices_buffer = torch.empty(
+            max_batch_size * self.runner.cache_config.num_gpu_blocks,
+            dtype=torch.int32,
+            device=self.runner.device)
+        self._graph_indptr_buffer = torch.empty(max_batch_size + 1,
+                                                dtype=torch.int32,
+                                                device=self.runner.device)
+        self._graph_last_page_len_buffer = torch.empty(
+            max_batch_size, dtype=torch.int32, device=self.runner.device)
+        yield
+        self._is_graph_capturing = False
+        del self._graph_slot_mapping
+        del self._graph_seq_lens
+        del self._graph_block_tables
+        del self._graph_decode_workspace_buffer
+        del self._graph_indices_buffer
+        del self._graph_indptr_buffer
+        del self._graph_last_page_len_buffer
+        del self._graph_decode_wrapper
+
+    def graph_clone(self, batch_size: int):
+        assert self._is_graph_capturing
+        state = self.__class__(self.runner)
+        state._workspace_buffer = self._graph_decode_workspace_buffer
+        state._decode_wrapper = self._graph_decode_wrapper
+        state._prefill_wrapper = self._get_prefill_wrapper()
+        return state
+
+    def graph_capture_get_metadata_for_batch(
+            self, batch_size: int, is_encoder_decoder_model: bool = False):
+        assert self._is_graph_capturing
+        _indptr_buffer = self._graph_indptr_buffer[:batch_size + 1]
+        _last_page_len_buffer = self._graph_last_page_len_buffer[:batch_size]
+
+        num_qo_heads = (self.runner.model_config.get_num_attention_heads(
+            self.runner.parallel_config))
+        num_kv_heads = self.runner.model_config.get_num_kv_heads(
+            self.runner.parallel_config)
+        use_tensor_cores = envs.VLLM_FLASHINFER_FORCE_TENSOR_CORES or (
+            num_qo_heads // num_kv_heads > 4)
+        self._graph_decode_wrapper = \
+            CUDAGraphBatchDecodeWithPagedKVCacheWrapper(
+            self._graph_decode_workspace_buffer, _indptr_buffer,
+            self._graph_indices_buffer, _last_page_len_buffer, "NHD",
+            use_tensor_cores)
+        if self.runner.kv_cache_dtype.startswith("fp8"):
+            kv_cache_dtype = FlashInferBackend.get_fp8_dtype_for_flashinfer(
+                self.runner.kv_cache_dtype)
+        else:
+            kv_cache_dtype = get_kv_cache_torch_dtype(
+                self.runner.kv_cache_dtype, self.runner.model_config.dtype)
+
+        paged_kv_indptr_tensor_host = torch.arange(0,
+                                                   batch_size + 1,
+                                                   dtype=torch.int32)
+        paged_kv_indices_tensor_host = torch.arange(0,
+                                                    batch_size,
+                                                    dtype=torch.int32)
+        paged_kv_last_page_len_tensor_host = torch.full((batch_size, ),
+                                                        self.runner.block_size,
+                                                        dtype=torch.int32)
+        query_start_loc_host = torch.arange(0,
+                                            batch_size + 1,
+                                            dtype=torch.int32)
+
+        attn_metadata = self.runner.attn_backend.make_metadata(
+            num_prefills=0,
+            slot_mapping=self._graph_slot_mapping[:batch_size],
+            multi_modal_placeholder_index_maps=None,
+            num_prefill_tokens=0,
+            num_decode_tokens=batch_size,
+            max_prefill_seq_len=0,
+            block_tables=self._graph_block_tables,
+            paged_kv_indptr=paged_kv_indptr_tensor_host,
+            paged_kv_indices=paged_kv_indices_tensor_host,
+            paged_kv_last_page_len=paged_kv_last_page_len_tensor_host,
+            num_qo_heads=num_qo_heads,
+            num_kv_heads=num_kv_heads,
+            head_dim=self.runner.model_config.get_head_size(),
+            page_size=self.runner.block_size,
+            seq_start_loc=None,
+            query_start_loc=query_start_loc_host,
+            device=self.runner.device,
+            data_type=kv_cache_dtype,
+            q_data_type=self.runner.model_config.dtype,
+            use_cuda_graph=True,
+            decode_wrapper=self._graph_decode_wrapper,
+            prefill_wrapper=None)
+        attn_metadata.begin_forward()
+        return attn_metadata
+
+    def get_graph_input_buffers(self,
+                                attn_metadata,
+                                is_encoder_decoder_model: bool = False):
+        return {
+            "slot_mapping": attn_metadata.slot_mapping,
+        }
+
+    def prepare_graph_input_buffers(self,
+                                    input_buffers,
+                                    attn_metadata,
+                                    is_encoder_decoder_model: bool = False):
+        return
+
+    def begin_forward(self, model_input):
+        assert not self._is_graph_capturing
+        state = self
+        if model_input.attn_metadata.use_cuda_graph:
+            batch_size = model_input.input_tokens.shape[0]
+            state = (self.runner.graph_runners[model_input.virtual_engine]
+                     [batch_size].attn_state)
+        model_input.attn_metadata.prefill_wrapper = state._get_prefill_wrapper(
+        )
+        model_input.attn_metadata.decode_wrapper = state._get_decode_wrapper()
+        model_input.attn_metadata.begin_forward()
+
+
+@dataclass
+class FlashInferMetadata(AttentionMetadata):
+    # Maximum sequence length among prefill batch. 0 if there are decoding
+    # requests only.
+    max_prefill_seq_len: int
+    # Number of query tokens for each request in the batch.
+    # Currently, we require that all requests have the same number of query
+    # tokens during the decoding phase. When speculavie decoding is enabled,
+    # decode_query_len might be greater than 1. In all other cases, it is 1.
+    decode_query_len: Optional[int] = 1
+
+    use_cuda_graph: bool = True
+
+    prefill_wrapper: Optional[BatchPrefillWithPagedKVCacheWrapper] = None
+    decode_wrapper: Optional[BatchDecodeWithPagedKVCacheWrapper] = None
+
+    # Metadata for the prefill stage
+    seq_start_loc: Optional[torch.Tensor] = None
+    query_start_loc: Optional[torch.Tensor] = None
+    block_tables: Optional[torch.Tensor] = None
+
+    # used for GPU in-place advance_step
+    seq_lens_tensor: Optional[torch.Tensor] = None
+    block_table_bound: Optional[torch.Tensor] = None
+
+    # An example for paged_kv_indices, paged_kv_indptr:
+    # request 1, page indices [0, 5, 8]
+    # request 2, page indices [1, 6, 7]
+    # request 3, page indices [3, 4]
+    # paged_kv_indices is a concatenation of page indices of all requests:
+    # [0, 5, 8, 1, 6, 7, 3, 4]
+    # paged_kv_indptr is used to index into paged_kv_indices:
+    # [0, 3, 6, 8]
+    # The indptr of the paged kv cache, shape: [batch_size + 1]
+    paged_kv_indptr: Optional[torch.Tensor] = None
+    # The page indices of the paged kv cache
+    paged_kv_indices: Optional[torch.Tensor] = None
+    # The number of entries in the last page of each request in
+    # the paged kv cache, shape: [batch_size]
+    paged_kv_last_page_len: Optional[torch.Tensor] = None
+    # The number of query/output heads
+    num_qo_heads: Optional[int] = None
+    # The number of key/value heads
+    num_kv_heads: Optional[int] = None
+    # The dimension of the attention heads
+    head_dim: Optional[int] = None
+    # Block size of vllm
+    page_size: Optional[int] = None
+    # The data type of the paged kv cache
+    data_type: torch.dtype = None
+    # The data type of the query
+    q_data_type: torch.dtype = None
+    device: torch.device = torch.device("cuda")
+    is_profile_run: bool = False
+
+    def __post_init__(self):
+        # Refer to
+        # https://github.com/flashinfer-ai/flashinfer/blob/3d55c71a62052c590c130897d3a3db49b14fcc34/include/flashinfer/utils.cuh#L157
+        supported_head_sizes = FlashInferBackend.get_supported_head_sizes()
+        if self.head_dim is not None and self.head_dim \
+                not in supported_head_sizes:
+            raise ValueError(
+                f"Only {supported_head_sizes} are supported for head_dim,",
+                f"received {self.head_dim}.")
+
+    def begin_forward(self):
+        if self.num_prefill_tokens > 0:
+            if self.paged_kv_indices is None:
+                return
+
+            assert self.prefill_wrapper is not None
+            assert self.query_start_loc is not None
+            assert self.paged_kv_indices is not None
+            assert self.paged_kv_indptr is not None
+            assert self.paged_kv_last_page_len is not None
+            assert self.block_table_bound is not None
+            assert self.seq_lens_tensor is not None
+            self.query_start_loc = self.query_start_loc[:self.num_prefills + 1]
+            batch_size = self.query_start_loc.shape[0] - 1
+            assert batch_size >= 0
+            # We will use flash attention for profiling to
+            # determine the number of blocks. Therefore,
+            # we don't need to prepare the input for flashinfer for profile run.
+            if not self.is_profile_run:
+                self.paged_kv_indptr = self.paged_kv_indptr.to(self.device)
+                self.paged_kv_last_page_len = self.paged_kv_last_page_len.to(
+                    self.device)
+                self.block_table_bound = self.block_table_bound.to(self.device)
+                self.seq_lens_tensor = self.seq_lens_tensor.to(self.device)
+                self.paged_kv_indices = self.paged_kv_indices.to(self.device)
+                self.prefill_wrapper.end_forward()
+                self.prefill_wrapper.begin_forward(
+                    self.query_start_loc,
+                    self.paged_kv_indptr[:self.num_prefills + 1],
+                    self.paged_kv_indices,
+                    self.paged_kv_last_page_len[:self.num_prefills],
+                    self.num_qo_heads, self.num_kv_heads, self.head_dim,
+                    self.page_size)
+        if self.num_decode_tokens > 0:
+            assert self.paged_kv_indices is not None
+            assert self.paged_kv_indptr is not None
+            assert self.paged_kv_last_page_len is not None
+            self.paged_kv_indices = self.paged_kv_indices.to(self.device)
+            self.paged_kv_indptr = self.paged_kv_indptr.to(self.device)
+            self.paged_kv_last_page_len = self.paged_kv_last_page_len.to(
+                self.device)
+            # handle model warmup path
+            if self.block_table_bound is not None:
+                self.block_table_bound = self.block_table_bound.to(self.device)
+            if self.seq_lens_tensor is not None:
+                self.seq_lens_tensor = self.seq_lens_tensor.to(self.device)
+
+            assert self.decode_wrapper is not None
+            self.decode_wrapper.end_forward()
+            self.decode_wrapper.begin_forward(
+                self.paged_kv_indptr[self.num_prefills:],
+                self.paged_kv_indices,
+                self.paged_kv_last_page_len[self.num_prefills:],
+                self.num_qo_heads,
+                self.num_kv_heads,
+                self.head_dim,
+                self.page_size,
+                # Disable flashinfer's pos encoding and use vllm's rope.
+                pos_encoding_mode="NONE",
+                # kv-cache data type.
+                data_type=self.data_type,
+                # query data type.
+                q_data_type=self.q_data_type)
+
+    def asdict_zerocopy(self,
+                        skip_fields: Optional[Set[str]] = None
+                        ) -> Dict[str, Any]:
+        if skip_fields is None:
+            skip_fields = set()
+        # We need to skip the prefill/decode_wrapper field since it cannot be
+        # broadcasted with nccl when TP is enabled.
+        skip_fields.add('prefill_wrapper')
+        skip_fields.add('decode_wrapper')
+        return super().asdict_zerocopy(skip_fields)
+
+    @property
+    def prefill_metadata(self) -> Optional["FlashInferMetadata"]:
+        if self.num_prefills == 0:
+            return None
+        return self
+
+    @property
+    def decode_metadata(self) -> Optional["FlashInferMetadata"]:
+        if self.num_decode_tokens == 0:
+            return None
+        return self
+
+    def advance_step(self,
+                     model_input: "ModelInputForGPUWithSamplingMetadata",
+                     sampled_token_ids: Optional[torch.Tensor],
+                     block_size: int,
+                     num_seqs: int,
+                     num_queries: int,
+                     turn_prefills_into_decodes: bool = False):
+        """
+        Update metadata in-place to advance one decode step.
+        """
+
+        assert not turn_prefills_into_decodes, \
+            ("Chunked prefill is not supported with flashinfer yet."
+             "turn_prefills_into_decodes is a Multi-Step + Chunked-Prefill "
+             "specific parameter.")
+
+        assert num_seqs > 0
+        assert num_queries > 0
+        assert model_input.attn_metadata is not None
+        assert sampled_token_ids is not None
+
+        # When using cudagraph, the num_seqs is padded to the next captured
+        # batch sized, but num_queries tracks the actual number of requests in
+        # the batch. For --enforce-eager mode, num_seqs == num_queries
+        if num_seqs != num_queries:
+            assert num_seqs > num_queries
+            assert self.use_cuda_graph
+
+        model_input.input_tokens[:num_queries] = sampled_token_ids.flatten()
+
+        # Update GPU tensors
+        ops.advance_step_flashinfer(
+            num_seqs=num_seqs,
+            num_queries=num_queries,
+            block_size=block_size,
+            input_tokens=model_input.input_tokens,
+            sampled_token_ids=model_input.input_tokens,
+            input_positions=model_input.input_positions,
+            seq_lens=self.seq_lens_tensor,
+            slot_mapping=self.slot_mapping,
+            block_tables=self.block_tables,
+            paged_kv_indices=self.paged_kv_indices,
+            paged_kv_indptr=self.paged_kv_indptr,
+            paged_kv_last_page_len=self.paged_kv_last_page_len,
+            block_table_bound=self.block_table_bound)
+
+
+class FlashInferMetadataBuilder(AttentionMetadataBuilder[FlashInferMetadata]):
+
+    def __init__(self, input_builder: "ModelInputForGPUBuilder"):
+        self.slot_mapping: List[int] = []
+        self.prefill_seq_lens: List[int] = []
+        self.context_lens: List[int] = []
+        self.block_tables: List[List[int]] = []
+        self.curr_seq_lens: List[int] = []
+        self.multimodal_placeholder_maps: Dict[
+            str,
+            MultiModalPlaceholderMap] = defaultdict(MultiModalPlaceholderMap)
+        self.num_prefills = 0
+        self.num_prefill_tokens = 0
+        self.num_decode_tokens = 0
+
+        self.input_builder = input_builder
+        self.runner = input_builder.runner
+
+        self.sliding_window = input_builder.sliding_window
+        self.block_size = input_builder.block_size
+
+        # Please follow https://docs.flashinfer.ai/tutorials/kv_layout.html#page-layout
+        # for the precise definition of the following fields.
+        # An example:
+        # request 1, page indices [0, 5, 8]
+        # request 2, page indices [1, 6, 7]
+        # request 3, page indices [3, 4]
+        # paged_kv_indices is a concatenation of page indices of all requests:
+        # [0, 5, 8, 1, 6, 7, 3, 4]
+        # paged_kv_indptr is used to index into paged_kv_indices:
+        # [0, 3, 6, 8]
+        self.paged_kv_indices: List[int] = []
+        # 0 at the beginning of paged_kv_indptr indicates the start of the
+        # first request’s page indices in the paged_kv_indices list.
+        self.paged_kv_indptr: List[int] = [0]
+        # paged_kv_last_page_len is the length of the last page of each request
+        self.paged_kv_last_page_len: List[int] = []
+        self.total_blocks = 0
+        self.is_profile_run: bool = False
+
+    def _add_seq_group(
+            self, inter_data: "ModelInputForGPUBuilder.InterDataForSeqGroup",
+            chunked_prefill_enabled: bool):
+        """Add a sequence group to the metadata. Specifically update/append
+        1. context length.
+        2. block table.
+        3. slot mapping.
+        """
+        is_prompt = inter_data.is_prompt
+        block_tables = inter_data.block_tables
+        computed_block_nums = inter_data.computed_block_nums
+
+        for (seq_id, token_len, seq_len, curr_seq_len, query_len, context_len,
+             curr_sliding_window_block) in zip(
+                 inter_data.seq_ids, [len(t) for t in inter_data.input_tokens],
+                 inter_data.orig_seq_lens, inter_data.seq_lens,
+                 inter_data.query_lens, inter_data.context_lens,
+                 inter_data.curr_sliding_window_blocks):
+            self.context_lens.append(context_len)
+            if is_prompt:
+                mm_maps = inter_data.multi_modal_placeholder_maps
+                if mm_maps:
+                    for modality, placeholders in mm_maps.items():
+                        self.multimodal_placeholder_maps[modality].extend(
+                            placeholders)
+                self.num_prefills += 1
+                self.num_prefill_tokens += token_len
+                self.prefill_seq_lens.append(seq_len)
+            else:
+                assert query_len == 1, (
+                    "seq_len: {}, context_len: {}, query_len: {}".format(
+                        seq_len, context_len, query_len))
+                self.num_decode_tokens += query_len
+                self.curr_seq_lens.append(curr_seq_len)
+
+            # Compute block table.
+            # TODO(sang): Combine chunked prefill and prefix caching by
+            # only allowing multiple of block_size chunk size.
+            # NOTE: This only works for oooooooxxx style attention.
+            block_table = []
+            if inter_data.prefix_cache_hit:
+                block_table = computed_block_nums
+            elif ((chunked_prefill_enabled or not is_prompt)
+                  and block_tables is not None):
+                block_table = block_tables[seq_id][-curr_sliding_window_block:]
+            self.block_tables.append(block_table)
+
+            is_profile_run = is_block_tables_empty(block_tables)
+
+            # Compute slot mapping.
+            start_idx = compute_slot_mapping_start_idx(is_prompt, query_len,
+                                                       context_len,
+                                                       self.sliding_window)
+            compute_slot_mapping(is_profile_run, self.slot_mapping, seq_id,
+                                 seq_len, context_len, start_idx,
+                                 self.block_size, inter_data.block_tables)
+
+            # It is not necessary to add paged_kv_indices, paged_kv_indptr,
+            # and paged_kv_last_page_len for profile run because we will
+            # create dummy inputs.
+            if is_profile_run:
+                self.is_profile_run = is_profile_run
+                return
+
+            block_table = block_tables[seq_id]
+            self._update_paged_kv_tensors(block_table, seq_len)
+
+    def _update_paged_kv_tensors(self, block_table: List[int], seq_len: int):
+        # Get the number of valid blocks based on sequence length.
+        # If seq_len = 16, block_size = 16,
+        # block_table_bound is 1 with 1 valid block.
+        # If seq_len = 15, block_size = 16,
+        # block_table_bound is 0 + 1 with 1 valid block.
+        self.total_blocks += len(block_table)
+        block_table_bound = seq_len // self.block_size + 1 \
+                            if seq_len % self.block_size != 0 \
+                            else seq_len // self.block_size
+        self.paged_kv_indices.extend(block_table[:block_table_bound])
+        self.paged_kv_indptr.append(self.paged_kv_indptr[-1] +
+                                    block_table_bound)
+
+        last_page_len = seq_len % self.block_size
+        if last_page_len == 0:
+            last_page_len = self.block_size
+        self.paged_kv_last_page_len.append(last_page_len)
+
+    def build(self, seq_lens: List[int], query_lens: List[int],
+              cuda_graph_pad_size: int, batch_size: int):
+        """Build attention metadata with on-device tensors.
+
+        Args:
+            seq_lens: The maybe padded sequence lengths of the input sequences.
+            query_lens: The query lengths of the input sequences.
+            cuda_graph_pad_size: The padding size for cuda graph.
+                                 -1 if cuda graph is not used.
+            batch_size: The maybe padded batch size.
+        """
+        for inter_data in self.input_builder.inter_data_list:
+            self._add_seq_group(inter_data,
+                                self.input_builder.chunked_prefill_enabled)
+
+        device = self.runner.device
+        use_captured_graph = cuda_graph_pad_size != -1
+
+        max_prefill_seq_len = max(self.prefill_seq_lens, default=0)
+        num_decode_tokens = self.num_decode_tokens
+        decode_query_len = max(query_lens[self.num_prefills:], default=1)
+
+        if use_captured_graph:
+            self.slot_mapping.extend([PAD_SLOT_ID] * cuda_graph_pad_size)
+            self.block_tables.extend([] * cuda_graph_pad_size)
+            num_decode_tokens = batch_size - self.num_prefill_tokens
+
+            # The shape of graph_block_tables is
+            # [max batch size, max context len // block size].
+            input_block_tables = self.runner.graph_block_tables[:batch_size]
+            max_blocks = input_block_tables.shape[1]
+            for i, block_table in enumerate(self.block_tables):
+                if block_table:
+                    num_blocks = len(block_table)
+                    if num_blocks <= max_blocks:
+                        input_block_tables[i, :num_blocks] = block_table
+                    else:
+                        # It may be possible to have more blocks allocated due
+                        # to lookahead slots of multi-step, however, they are
+                        # not used anyway, so can be safely ignored.
+                        input_block_tables[
+                            i, :max_blocks] = block_table[:max_blocks]
+
+            block_tables = torch.from_numpy(input_block_tables).to(
+                device, non_blocking=True)
+
+            last_paged_kv_indptr = self.paged_kv_indptr[-1]
+            self.paged_kv_indptr.extend([last_paged_kv_indptr] *
+                                        cuda_graph_pad_size)
+            self.paged_kv_last_page_len.extend([0] * cuda_graph_pad_size)
+        else:
+            block_tables = make_tensor_with_pad(
+                self.block_tables,
+                pad=0,
+                dtype=torch.int,
+                device=device,
+            )
+
+        assert device is not None
+        seq_lens_tensor = async_tensor_h2d(seq_lens, torch.int, device,
+                                           self.runner.pin_memory)
+        query_lens_tensor = async_tensor_h2d(query_lens, torch.long, device,
+                                             self.runner.pin_memory)
+        slot_mapping_tensor = async_tensor_h2d(self.slot_mapping, torch.long,
+                                               device, self.runner.pin_memory)
+        query_start_loc = torch.zeros(query_lens_tensor.shape[0] + 1,
+                                      dtype=torch.int32,
+                                      device=device)
+        seq_start_loc = torch.zeros(seq_lens_tensor.shape[0] + 1,
+                                    dtype=torch.int32,
+                                    device=device)
+        placeholder_index_maps = {
+            modality: placeholder_map.index_map()
+            for modality, placeholder_map in
+            self.multimodal_placeholder_maps.items()
+        }
+        torch.cumsum(seq_lens_tensor,
+                     dim=0,
+                     dtype=seq_start_loc.dtype,
+                     out=seq_start_loc[1:])
+        torch.cumsum(query_lens_tensor,
+                     dim=0,
+                     dtype=query_start_loc.dtype,
+                     out=query_start_loc[1:])
+
+        if len(self.paged_kv_indptr) > 0:
+            # extend to the maximum number of blocks as returned by the
+            # scheduler
+            self.paged_kv_indices.extend(
+                [0] * (self.total_blocks - len(self.paged_kv_indices)))
+            paged_kv_indices_tensor = torch.tensor(self.paged_kv_indices,
+                                                   device="cpu",
+                                                   dtype=torch.int)
+            paged_kv_indptr_tensor = torch.tensor(self.paged_kv_indptr,
+                                                  device="cpu",
+                                                  dtype=torch.int)
+            paged_kv_last_page_len_tensor = torch.tensor(
+                self.paged_kv_last_page_len, device="cpu", dtype=torch.int)
+            block_table_bound_tensor = torch.zeros(len(self.paged_kv_indptr) -
+                                                   1,
+                                                   device="cpu",
+                                                   dtype=torch.int)
+        else:
+            paged_kv_indices_tensor = None
+            paged_kv_indptr_tensor = None
+            paged_kv_last_page_len_tensor = None
+            block_table_bound_tensor = None
+
+        if self.runner.kv_cache_dtype.startswith("fp8"):
+            kv_cache_dtype = FlashInferBackend.get_fp8_dtype_for_flashinfer(
+                self.runner.kv_cache_dtype)
+        else:
+            kv_cache_dtype = get_kv_cache_torch_dtype(
+                self.runner.kv_cache_dtype, self.runner.model_config.dtype)
+
+        return FlashInferMetadata(
+            decode_query_len=decode_query_len,
+            num_prefills=self.num_prefills,
+            slot_mapping=slot_mapping_tensor,
+            multi_modal_placeholder_index_maps=placeholder_index_maps,
+            num_prefill_tokens=self.num_prefill_tokens,
+            num_decode_tokens=num_decode_tokens,
+            max_prefill_seq_len=max_prefill_seq_len,
+            block_tables=block_tables,
+            paged_kv_indptr=paged_kv_indptr_tensor,
+            paged_kv_indices=paged_kv_indices_tensor,
+            paged_kv_last_page_len=paged_kv_last_page_len_tensor,
+            block_table_bound=block_table_bound_tensor,
+            seq_lens_tensor=seq_lens_tensor,
+            num_qo_heads=self.runner.model_config.get_num_attention_heads(
+                self.runner.parallel_config),
+            num_kv_heads=self.runner.model_config.get_num_kv_heads(
+                self.runner.parallel_config),
+            head_dim=self.runner.model_config.get_head_size(),
+            page_size=self.block_size,
+            seq_start_loc=seq_start_loc,
+            query_start_loc=query_start_loc,
+            device=device,
+            data_type=kv_cache_dtype,
+            q_data_type=self.runner.model_config.dtype,
+            use_cuda_graph=use_captured_graph,
+            is_profile_run=self.is_profile_run)
+
+
+class FlashInferImpl(AttentionImpl):
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: Optional[List[float]],
+        sliding_window: Optional[int],
+        kv_cache_dtype: str,
+        blocksparse_params: Optional[Dict[str, Any]] = None,
+        logits_soft_cap: Optional[float] = None,
+    ) -> None:
+        self.num_heads = num_heads
+        self.head_size = head_size
+        self.scale = float(scale)
+        self.num_kv_heads = num_kv_heads
+        if alibi_slopes is not None:
+            alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
+        self.alibi_slopes = alibi_slopes
+        if sliding_window is not None:
+            raise ValueError("Sliding window is not supported in FlashInfer.")
+        self.sliding_window = (-1, -1)
+        self.kv_cache_dtype = kv_cache_dtype
+        self.logits_soft_cap = logits_soft_cap
+
+        assert self.num_heads % self.num_kv_heads == 0
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: FlashInferMetadata,
+        k_scale: float = 1.0,
+        v_scale: float = 1.0,
+        attn_type: AttentionType = AttentionType.DECODER,
+    ) -> torch.Tensor:
+        if attn_type != AttentionType.DECODER:
+            raise NotImplementedError("Encoder self-attention and "
+                                      "encoder/decoder cross-attention "
+                                      "are not implemented for "
+                                      "FlashInferImpl")
+
+        return torch.ops.vllm.unified_flash_infer(
+            query,
+            key,
+            value,
+            self.num_heads,
+            self.head_size,
+            self.num_kv_heads,
+            kv_cache,
+            self.kv_cache_dtype,
+            k_scale,
+            v_scale,
+            self.scale,
+            self.sliding_window,
+            self.alibi_slopes,
+            self.logits_soft_cap,
+        )
+
+
+def unified_flash_infer(
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    num_heads: int,
+    head_size: int,
+    num_kv_heads: int,
+    kv_cache: torch.Tensor,
+    kv_cache_dtype: str,
+    k_scale: float,
+    v_scale: float,
+    softmax_scale: float,
+    window_size: Optional[List[int]] = None,
+    alibi_slopes: Optional[torch.Tensor] = None,
+    logits_soft_cap: Optional[float] = None,
+) -> torch.Tensor:
+
+    current_metadata = get_forward_context()
+    assert current_metadata is not None
+    assert isinstance(current_metadata, FlashInferMetadata)
+    attn_metadata: FlashInferMetadata = current_metadata
+
+    num_tokens, hidden_size = query.shape
+    query = query.view(-1, num_heads, head_size)
+    key = key.view(-1, num_kv_heads, head_size)
+    value = value.view(-1, num_kv_heads, head_size)
+
+    if kv_cache.numel() > 0:
+        # Use the same reshape and cache kernel as flash attention.
+        ops.reshape_and_cache_flash(
+            key,
+            value,
+            kv_cache[:, 0],
+            kv_cache[:, 1],
+            attn_metadata.slot_mapping.flatten(),
+            kv_cache_dtype,
+            k_scale,
+            v_scale,
+        )
+        # The FlashInfer api requires data to be in fp8_e4m3 or fp8_e5m2
+        # to process the cache when the kv_cache_dtype is fp8
+        if kv_cache_dtype.startswith("fp8"):
+            torch_dtype = FlashInferBackend.get_fp8_dtype_for_flashinfer(
+                kv_cache_dtype)
+            kv_cache = kv_cache.view(torch_dtype)
+
+    num_prefill_tokens = attn_metadata.num_prefill_tokens
+    num_decode_tokens = attn_metadata.num_decode_tokens
+    assert key.shape[0] == num_prefill_tokens + num_decode_tokens, \
+                f"key : {key.shape} : #prefill tokens {num_prefill_tokens} : #decode tokens {num_decode_tokens}" # noqa
+    assert value.shape[0] == num_prefill_tokens + num_decode_tokens, \
+                f"value : {value.shape} : #prefill toks {num_prefill_tokens} : #decode toks {num_decode_tokens}" # noqa
+    query = query.contiguous()  # Flashinfer requires query to be contiguous
+    # Query for decode. KV is not needed because it is already cached.
+    # QKV for prefill.
+    decode_query = query[num_prefill_tokens:]
+    query = query[:num_prefill_tokens]
+
+    key = key[:num_prefill_tokens]
+    value = value[:num_prefill_tokens]
+
+    assert query.shape[0] == num_prefill_tokens
+    assert decode_query.shape[0] == num_decode_tokens
+
+    prefill_output: Optional[torch.Tensor] = None
+    decode_output: Optional[torch.Tensor] = None
+    if prefill_meta := attn_metadata.prefill_metadata:
+        # We will use flash attention for prefill
+        # when kv_cache is not provided.
+        # This happens when vllm runs the profiling to
+        # determine the number of blocks.
+        if kv_cache.numel() == 0:
+            prefill_output = flash_attn_varlen_func(
+                q=query,
+                k=key,
+                v=value,
+                cu_seqlens_q=prefill_meta.seq_start_loc,
+                cu_seqlens_k=prefill_meta.seq_start_loc,
+                max_seqlen_q=prefill_meta.max_prefill_seq_len,
+                max_seqlen_k=prefill_meta.max_prefill_seq_len,
+                softmax_scale=softmax_scale,
+                causal=True,
+                window_size=window_size,
+                alibi_slopes=alibi_slopes,
+            )
+        else:
+            assert prefill_meta is not None
+            assert prefill_meta.prefill_wrapper is not None
+            prefill_output = prefill_meta.prefill_wrapper.forward(
+                query,
+                kv_cache,
+                logits_soft_cap=logits_soft_cap,
+                causal=True,
+                k_scale=k_scale,
+                v_scale=v_scale)
+    if decode_meta := attn_metadata.decode_metadata:
+        assert attn_metadata.decode_metadata is not None
+        assert attn_metadata.decode_metadata.decode_wrapper is not None
+        decode_output = attn_metadata.decode_metadata.decode_wrapper.forward(
+            decode_query,
+            kv_cache,
+            sm_scale=softmax_scale,
+            logits_soft_cap=logits_soft_cap,
+            k_scale=k_scale,
+            v_scale=v_scale)
+
+    if prefill_output is None and decode_output is not None:
+        # Decode only batch.
+        output, num_tokens = decode_output, num_decode_tokens
+    elif decode_output is None and prefill_output is not None:
+        # Prefill only batch.
+        output, num_tokens = prefill_output, num_prefill_tokens
+    else:
+        # Chunked prefill batch does not work with speculative decoding in
+        # FlashInfer backend, so the query length for decode should be 1.
+        assert prefill_output is not None
+        assert decode_output is not None
+        assert decode_meta is not None
+        assert decode_meta.decode_query_len == 1
+        decode_output = decode_output.squeeze(1)
+        output = torch.cat([prefill_output, decode_output], dim=0)
+    return output.view(num_tokens, hidden_size)
+
+
+def unified_flash_infer_fake(
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    num_heads: int,
+    head_size: int,
+    num_kv_heads: int,
+    kv_cache: torch.Tensor,
+    kv_cache_dtype: str,
+    k_scale: float,
+    v_scale: float,
+    softmax_scale: float,
+    window_size: Optional[List[int]] = None,
+    alibi_slopes: Optional[torch.Tensor] = None,
+    logits_soft_cap: Optional[float] = None,
+) -> torch.Tensor:
+    return torch.empty_like(query).contiguous()
+
+
+direct_register_custom_op(
+    op_name="unified_flash_infer",
+    op_func=unified_flash_infer,
+    mutates_args=["kv_cache"],
+    fake_impl=unified_flash_infer_fake,
+)

vllm-v0.6.2/vllm/attention/backends/hpu_attn.py

@@ -0,0 +1,264 @@
+###############################################################################
+# Copyright (C) 2024 Habana Labs, Ltd. an Intel Company
+###############################################################################
+
+import os
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Type
+
+import torch
+import vllm_hpu_extension.ops as ops
+from vllm_hpu_extension.utils import Matmul, Softmax, VLLMKVCache
+
+from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
+                                              AttentionMetadata, AttentionType)
+from vllm.attention.backends.utils import CommonAttentionState
+from vllm.attention.ops.hpu_paged_attn import (HPUPagedAttention,
+                                               HPUPagedAttentionMetadata)
+from vllm.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+class HPUAttentionBackend(AttentionBackend):
+
+    @staticmethod
+    def get_impl_cls() -> Type["HPUAttentionImpl"]:
+        return HPUAttentionImpl
+
+    @staticmethod
+    def get_metadata_cls() -> Type["AttentionMetadata"]:
+        return HPUAttentionMetadata
+
+    @staticmethod
+    def get_state_cls() -> Type["CommonAttentionState"]:
+        return CommonAttentionState
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        return HPUPagedAttention.get_kv_cache_shape(num_blocks, block_size,
+                                                    num_kv_heads, head_size)
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: Dict[int, int],
+    ) -> None:
+        HPUPagedAttention.swap_blocks(src_kv_cache, dst_kv_cache, src_to_dst)
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: Dict[int, List[int]],
+    ) -> None:
+        HPUPagedAttention.copy_blocks(kv_caches, src_to_dists)
+
+
+@dataclass
+class HPUAttentionMetadata(HPUPagedAttentionMetadata, AttentionMetadata):
+    """Metadata for HPUAttentionbackend."""
+    # Currently, input sequences can only contain all prompts
+    # or all decoding. True if all sequences are prompts.
+    is_prompt: bool
+    attn_bias: Optional[torch.Tensor]
+    seq_lens_tensor: Optional[torch.Tensor]
+
+
+class HPUAttentionImpl(AttentionImpl, torch.nn.Module):
+    """
+    If the input tensors contain prompt tokens, the layout is as follows:
+    |<--------------- num_prefill_tokens ----------------->|
+    |<--prefill_0-->|<--prefill_1-->|...|<--prefill_N-1--->|
+
+    Otherwise, the layout is as follows:
+    |<----------------- num_decode_tokens ------------------>|
+    |<--decode_0-->|..........|<--decode_M-1-->|<--padding-->|
+
+    Generation tokens can contain padding when cuda-graph is used.
+    Currently, prompt tokens don't contain any padding.
+
+    The prompts might have different lengths, while the generation tokens
+    always have length 1.
+    """
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: Optional[List[float]],
+        sliding_window: Optional[int],
+        kv_cache_dtype: str,
+        blocksparse_params: Optional[Dict[str, Any]] = None,
+        max_seq_len: int = 4096,
+    ) -> None:
+        super(AttentionImpl, self).__init__()
+        self.kv_cache_dtype = kv_cache_dtype
+        self.num_heads = num_heads
+        self.head_size = head_size
+        self.scale = float(scale)
+        self.matmul_qk = Matmul()
+        self.softmax = Softmax()
+        self.matmul_av = Matmul()
+        self.k_cache = VLLMKVCache()
+        self.v_cache = VLLMKVCache()
+        self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
+        self.sliding_window = sliding_window
+        self.alibi_slopes = alibi_slopes
+        if alibi_slopes is not None:
+            alibi_slopes_tensor = torch.tensor(alibi_slopes,
+                                               dtype=torch.bfloat16)
+            self.alibi_slopes = alibi_slopes_tensor
+        assert self.num_heads % self.num_kv_heads == 0
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+
+        self.prefill_usefusedsdpa = os.getenv('VLLM_PROMPT_USE_FUSEDSDPA',
+                                              '0').lower() in ['1', 'true']
+        if self.prefill_usefusedsdpa:
+            assert alibi_slopes is None, \
+                'Prefill with FusedSDPA not supported with alibi slopes!'
+
+        suppored_head_sizes = HPUPagedAttention.get_supported_head_sizes()
+        if head_size not in suppored_head_sizes:
+            raise ValueError(
+                f"Head size {head_size} is not supported by PagedAttention. "
+                f"Supported head sizes are: {suppored_head_sizes}.")
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: HPUAttentionMetadata,
+        k_scale: float = 1.0,
+        v_scale: float = 1.0,
+        attn_type: AttentionType = AttentionType.DECODER,
+    ) -> torch.Tensor:
+        """Forward pass with xFormers and PagedAttention.
+
+        Args:
+            query: shape = [num_tokens, num_heads * head_size]
+            key: shape = [num_tokens, num_kv_heads * head_size]
+            value: shape = [num_tokens, num_kv_heads * head_size]
+            kv_cache = [2, num_blocks, block_size * num_kv_heads * head_size]
+            attn_metadata: Metadata for attention.
+        Returns:
+            shape = [num_tokens, num_heads * head_size]
+        """
+        if attn_type != AttentionType.DECODER:
+            raise NotImplementedError("Encoder self-attention and "
+                                      "encoder/decoder cross-attention "
+                                      "are not implemented for "
+                                      "HPUAttentionImpl")
+        batch_size, seq_len, hidden_size = query.shape
+        _, seq_len_kv, _ = key.shape
+
+        query = query.view(-1, self.num_heads, self.head_size)
+        key = key.view(-1, self.num_kv_heads, self.head_size)
+        value = value.view(-1, self.num_kv_heads, self.head_size)
+        block_indices = attn_metadata.block_indices
+        block_offsets = attn_metadata.block_offsets
+        if attn_metadata.is_prompt:
+            key = key.unflatten(0, (block_indices.size(0), -1))
+            value = value.unflatten(0, (block_indices.size(0), -1))
+        if kv_cache is not None:
+            key_cache, value_cache = HPUPagedAttention.split_kv_cache(
+                kv_cache, self.num_kv_heads, self.head_size)
+
+            # Reshape the input keys and values and store them in the cache.
+            # If kv_cache is not provided, the new key and value tensors are
+            # not cached. This happens during the initial memory profiling run.
+            key_cache = self.k_cache(key, key_cache, block_indices,
+                                     block_offsets)
+            value_cache = self.v_cache(value, value_cache, block_indices,
+                                       block_offsets)
+
+        if attn_metadata.is_prompt:
+            # Prompt run.
+            if not self.prefill_usefusedsdpa:
+                # TODO: move this outside of model
+                assert attn_metadata.attn_bias is not None, \
+                        'attn_bias must be set before calling model.forward!'
+                attn_bias = attn_metadata.attn_bias
+                if self.alibi_slopes is not None:
+                    position_bias = _make_alibi_bias(self.alibi_slopes,
+                                                     self.num_kv_heads,
+                                                     attn_bias.dtype,
+                                                     attn_bias.shape[-1])
+                    attn_bias = attn_bias.tile((1, self.num_kv_heads, 1, 1))
+                    attn_bias.add_(position_bias)
+            else:
+                attn_bias = None
+
+            query_shape = (batch_size, seq_len, self.num_heads, self.head_size)
+            kv_shape = (batch_size, seq_len_kv, self.num_kv_heads,
+                        self.head_size)
+            out = ops.prompt_attention(
+                query.view(query_shape),
+                key.view(kv_shape),
+                value.view(kv_shape),
+                attn_bias=attn_bias,
+                p=0.0,
+                scale=self.scale,
+                matmul_qk_op=self.matmul_qk,
+                softmax_op=self.softmax,
+                matmul_av_op=self.matmul_av,
+            )
+            output = out.reshape(batch_size, seq_len, hidden_size)
+        else:
+            # Decoding run.
+            output = HPUPagedAttention.forward_decode(
+                query=query,
+                key_cache=key_cache,
+                value_cache=value_cache,
+                block_list=attn_metadata.block_list,
+                block_mapping=attn_metadata.block_mapping,
+                block_bias=attn_metadata.attn_bias,
+                block_scales=attn_metadata.block_scales,
+                scale=self.scale,
+                matmul_qk_op=self.matmul_qk,
+                matmul_av_op=self.matmul_av,
+                keys_fetch_func=self.k_cache.fetch_from_cache,
+                values_fetch_func=self.v_cache.fetch_from_cache)
+        # Reshape the output tensor.
+        return output.view(batch_size, seq_len, hidden_size)
+
+
+def _make_alibi_bias(
+    alibi_slopes: torch.Tensor,
+    num_kv_heads: int,
+    dtype: torch.dtype,
+    seq_len: int,
+) -> torch.Tensor:
+    bias = torch.arange(seq_len, dtype=dtype)
+    # NOTE(zhuohan): HF uses
+    #     `bias = bias[None, :].repeat(seq_len, 1)`
+    # here. We find that both biases give the same results, but
+    # the bias below more accurately follows the original ALiBi
+    # paper.
+    # Calculate a matrix where each element represents ith element- jth
+    # element.
+    bias = bias[None, :] - bias[:, None]
+
+    padded_len = (seq_len + 7) // 8 * 8
+    num_heads = alibi_slopes.shape[0]
+    bias = torch.empty(
+        1,  # batch size
+        num_heads,
+        seq_len,
+        padded_len,
+        device=alibi_slopes.device,
+        dtype=dtype,
+    )[:, :, :, :seq_len].copy_(bias)
+    bias.mul_(alibi_slopes[:, None, None])
+    if num_heads != num_kv_heads:
+        bias = bias.unflatten(1, (num_kv_heads, num_heads // num_kv_heads))
+    return bias

vllm-v0.6.2/vllm/attention/backends/ipex_attn.py

@@ -0,0 +1,385 @@
+""" Attention layer with torch scaled_dot_product_attention
+    and PagedAttention."""
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Type
+
+import torch
+
+from vllm._ipex_ops import ipex_ops
+from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
+                                              AttentionMetadata, AttentionType)
+from vllm.attention.backends.utils import CommonAttentionState
+from vllm.attention.ops.paged_attn import (PagedAttention,
+                                           PagedAttentionMetadata)
+
+_PARTITION_SIZE = 512
+
+
+class IpexAttnBackend(AttentionBackend):
+
+    @staticmethod
+    def get_name() -> str:
+        return "IPEX"
+
+    @staticmethod
+    def get_impl_cls() -> Type["IpexAttnBackendImpl"]:
+        return IpexAttnBackendImpl
+
+    @staticmethod
+    def get_metadata_cls() -> Type["IpexAttnMetadata"]:
+        return IpexAttnMetadata
+
+    @staticmethod
+    def get_state_cls() -> Type["CommonAttentionState"]:
+        return CommonAttentionState
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        return PagedAttention.get_kv_cache_shape(num_blocks, block_size,
+                                                 num_kv_heads, head_size)
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: torch.Tensor,
+    ) -> None:
+        from vllm._ipex_ops import ipex_ops as ops
+        ops.swap_blocks(src_kv_cache, dst_kv_cache, src_to_dst)
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: torch.Tensor,
+    ) -> None:
+        from vllm._ipex_ops import ipex_ops as ops
+        key_caches = [kv_cache[0] for kv_cache in kv_caches]
+        value_caches = [kv_cache[1] for kv_cache in kv_caches]
+        ops.copy_blocks(key_caches, value_caches, src_to_dists)
+
+
+@dataclass
+class IpexAttnMetadata(AttentionMetadata, PagedAttentionMetadata):
+    """Metadata for IpexAttnBackend.
+    """
+    # Currently, input sequences can only contain all prompts
+    # or all decoding. True if all sequences are prompts.
+    is_prompt: bool
+    slot_mapping: torch.Tensor
+    seq_lens: Optional[List[int]]
+    seqlen_q: Optional[torch.Tensor]
+    max_seqlen: Optional[int]
+
+    def __post_init__(self):
+        # Set during the execution of the first attention op.
+        # It is a list because it is needed to set per prompt
+        # when alibi slopes is used. It is because of the limitation
+        # from xformer API.
+        # will not appear in the __repr__ and __init__
+        self.attn_bias: Optional[List[torch.Tensor]] = None
+
+    @property
+    def prefill_metadata(self) -> Optional["IpexAttnMetadata"]:
+        # Currently chunked prefill is not supported
+        if self.num_decode_tokens == 0:
+            assert self.num_prefills > 0
+            return self
+
+        return None
+
+    @property
+    def decode_metadata(self) -> Optional["IpexAttnMetadata"]:
+        # Currently chunked prefill is not supported
+        if self.num_prefills > 0:
+            assert self.num_decode_tokens == 0
+            return None
+
+        return self
+
+
+class IpexAttnBackendImpl(AttentionImpl[IpexAttnMetadata]):
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: Optional[List[float]],
+        sliding_window: Optional[int],
+        kv_cache_dtype: str,
+        blocksparse_params: Optional[Dict[str, Any]] = None,
+        logits_soft_cap: Optional[float] = None,
+    ) -> None:
+        if blocksparse_params is not None:
+            raise ValueError(
+                "IPEX backend does not support block-sparse attention.")
+        self.num_heads = num_heads
+        self.head_size = head_size
+        self.scale = float(scale)
+        self.num_kv_heads = num_kv_heads
+        if alibi_slopes is not None:
+            alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
+        self.alibi_slopes = alibi_slopes
+        self.sliding_window = sliding_window
+        self.kv_cache_dtype = kv_cache_dtype
+
+        assert self.num_heads % self.num_kv_heads == 0
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+        self.need_mask = (self.alibi_slopes is not None
+                          or self.sliding_window is not None)
+        if logits_soft_cap is None:
+            logits_soft_cap = 0
+        self.logits_soft_cap = logits_soft_cap
+
+        supported_head_sizes = PagedAttention.get_supported_head_sizes()
+        if head_size not in supported_head_sizes:
+            raise ValueError(
+                f"Head size {head_size} is not supported by PagedAttention. "
+                f"Supported head sizes are: {supported_head_sizes}.")
+        if kv_cache_dtype != "auto":
+            raise NotImplementedError(
+                "IPEX backend does not support FP8 KV cache. "
+                "Please use xFormers backend instead.")
+
+    def split_kv_cache(
+        self,
+        kv_cache: torch.Tensor,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        x = 1
+        num_blocks = kv_cache.shape[1]
+
+        key_cache = kv_cache[0]
+        key_cache = key_cache.view(num_blocks, num_kv_heads, head_size // x,
+                                   -1, x)
+        value_cache = kv_cache[1]
+        value_cache = value_cache.view(num_blocks, num_kv_heads, head_size, -1)
+        return key_cache, value_cache
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: IpexAttnMetadata,  # type: ignore
+        k_scale: float = 1.0,
+        v_scale: float = 1.0,
+        attn_type: AttentionType = AttentionType.DECODER,
+    ) -> torch.Tensor:
+        """Forward pass with IPEX varlen_attention and PagedAttention.
+
+        Args:
+            query: shape = [num_tokens, num_heads * head_size]
+            key: shape = [num_tokens, num_kv_heads * head_size]
+            value: shape = [num_tokens, num_kv_heads * head_size]
+            kv_cache = [2, num_blocks, block_size * num_kv_heads * head_size]
+                NOTE: kv_cache will be an empty tensor with shape [0]
+                for profiling run.
+            attn_metadata: Metadata for attention.
+        Returns:
+            shape = [num_tokens, num_heads * head_size]
+        """
+        assert k_scale == 1.0 and v_scale == 1.0
+        if attn_type != AttentionType.DECODER:
+            raise NotImplementedError("Encoder self-attention and "
+                                      "encoder/decoder cross-attention "
+                                      "are not implemented for "
+                                      "IpexAttnBackendImpl")
+        num_tokens, hidden_size = query.shape
+        # Reshape the query, key, and value tensors.
+        query = query.view(-1, self.num_heads, self.head_size)
+        key = key.view(-1, self.num_kv_heads, self.head_size)
+        value = value.view(-1, self.num_kv_heads, self.head_size)
+
+        if kv_cache.numel() > 0:
+            key_cache, value_cache = self.split_kv_cache(
+                kv_cache, self.num_kv_heads, self.head_size)
+            ipex_ops.reshape_and_cache(
+                key,
+                value,
+                key_cache,
+                value_cache,
+                attn_metadata.slot_mapping.flatten(),
+                self.kv_cache_dtype,
+                k_scale,
+                v_scale,
+            )
+
+        if attn_metadata.is_prompt:
+            assert attn_metadata.seq_lens is not None
+            if (kv_cache.numel() == 0
+                    or attn_metadata.block_tables.numel() == 0):
+                if self.num_kv_heads != self.num_heads:
+                    key = key.repeat_interleave(self.num_queries_per_kv, dim=1)
+                    value = value.repeat_interleave(self.num_queries_per_kv,
+                                                    dim=1)
+
+                if attn_metadata.attn_bias is None:
+                    if self.alibi_slopes is not None:
+                        att_masks = _make_alibi_bias(
+                            self.alibi_slopes, query.dtype,
+                            attn_metadata.seq_lens)  # type: ignore
+                    elif self.sliding_window is not None:
+                        att_masks = _make_sliding_window_bias(
+                            attn_metadata.seq_lens, self.sliding_window,
+                            query.dtype)  # type: ignore
+                    else:
+                        att_masks = _make_sliding_window_bias(
+                            attn_metadata.seq_lens, None, dtype=query.dtype)
+                    attn_metadata.attn_bias = att_masks
+
+                output = torch.empty(
+                    (num_tokens, self.num_heads, self.head_size),
+                    dtype=query.dtype,
+                    device=query.device)
+                ipex_ops.varlen_attention(
+                    query,
+                    key,
+                    value,
+                    output,
+                    attn_metadata.seqlen_q,
+                    attn_metadata.seqlen_q,
+                    attn_metadata.max_seqlen,
+                    attn_metadata.max_seqlen,
+                    pdropout=0.0,
+                    softmax_scale=self.scale,
+                    zero_tensors=False,
+                    is_causal=True,
+                    return_softmax=False,
+                    gen_=None,
+                    logits_soft_cap=self.logits_soft_cap,
+                )
+            else:
+                # prefix-enabled attention
+                raise RuntimeError(
+                    "IPEX backend doesn't support prefix decoding.")
+
+        else:
+            # Decoding run.
+            max_seq_len = attn_metadata.max_decode_seq_len
+            output = torch.empty_like(query)
+            block_size = value_cache.shape[3]
+            num_seqs, num_heads, head_size = query.shape
+            max_num_partitions = ((max_seq_len + _PARTITION_SIZE - 1) //
+                                  _PARTITION_SIZE)
+            # NOTE(woosuk): We use a simple heuristic to decide whether to use
+            # PagedAttention V1 or V2. If the number of partitions is 1, we use
+            # V1 to avoid the overhead of reduction. Also, if the number of
+            # sequences or heads is large, we use V1 since there is enough work
+            # to parallelize.
+            # TODO(woosuk): Tune this heuristic.
+            # For context len > 8192, use V2 kernel to avoid shared memory
+            # shortage.
+            use_v1 = (max_seq_len <= 8192 and
+                      (max_num_partitions == 1 or num_seqs * num_heads > 512))
+            if use_v1:
+                # Run PagedAttention V1.
+                ipex_ops.paged_attention_v1(
+                    output,
+                    query,
+                    key_cache,
+                    value_cache,
+                    self.num_kv_heads,
+                    self.scale,
+                    attn_metadata.block_tables,
+                    attn_metadata.seq_lens_tensor,
+                    block_size,
+                    max_seq_len,
+                    self.alibi_slopes,
+                    self.kv_cache_dtype,
+                    k_scale,
+                    v_scale,
+                )
+            else:
+                # Run PagedAttention V2.
+                assert _PARTITION_SIZE % block_size == 0
+                tmp_output = torch.empty(
+                    size=(num_seqs, num_heads, max_num_partitions, head_size),
+                    dtype=output.dtype,
+                    device=output.device,
+                )
+                exp_sums = torch.empty(
+                    size=(num_seqs, num_heads, max_num_partitions),
+                    dtype=torch.float32,
+                    device=output.device,
+                )
+                max_logits = torch.empty_like(exp_sums)
+                ipex_ops.paged_attention_v2(
+                    output,
+                    exp_sums,
+                    max_logits,
+                    tmp_output,
+                    query,
+                    key_cache,
+                    value_cache,
+                    self.num_kv_heads,
+                    self.scale,
+                    attn_metadata.block_tables,
+                    attn_metadata.seq_lens_tensor,
+                    block_size,
+                    max_seq_len,
+                    self.alibi_slopes,
+                    self.kv_cache_dtype,
+                    k_scale,
+                    v_scale,
+                )
+
+            # Reshape the output tensor.
+        return output.view(-1, self.num_heads * self.head_size)
+
+
+def _make_alibi_bias(
+    alibi_slopes: torch.Tensor,
+    dtype: torch.dtype,
+    seq_lens: List[int],
+) -> List[torch.Tensor]:
+    attn_biases = []
+    for seq_len in seq_lens:
+        bias = torch.arange(seq_len, dtype=dtype, device=alibi_slopes.device)
+        # NOTE(zhuohan): HF uses
+        #     `bias = bias[None, :].repeat(seq_len, 1)`
+        # here. We find that both biases give the same results, but
+        # the bias below more accurately follows the original ALiBi
+        # paper.
+        bias = bias[None, :] - bias[:, None]
+
+        num_heads = alibi_slopes.shape[0]
+        bias = bias[None, :].repeat((num_heads, 1, 1))
+        bias.mul_(alibi_slopes[:, None, None])
+        inf_mask = torch.empty(
+            (1, seq_len, seq_len),
+            dtype=bias.dtype,
+            device=alibi_slopes.device).fill_(-torch.inf).triu_(diagonal=1)
+        attn_biases.append((bias + inf_mask).to(dtype))
+
+    return attn_biases
+
+
+def _make_sliding_window_bias(
+    seq_lens: List[int],
+    window_size: Optional[int],
+    dtype: torch.dtype,
+) -> List[torch.Tensor]:
+    attn_biases = []
+    for seq_len in seq_lens:
+        tensor = torch.full(
+            (1, seq_len, seq_len),
+            dtype=dtype,
+            fill_value=1,
+        )
+        shift = 0
+        mask = torch.tril(tensor, diagonal=shift).to(dtype)  # type: ignore
+        if window_size is not None:
+            mask = torch.triu(mask, diagonal=shift - window_size + 1)
+        mask = torch.log(mask)
+        attn_biases.append(mask.to(dtype))
+
+    return attn_biases

vllm-v0.6.2/vllm/attention/backends/openvino.py

@@ -0,0 +1,140 @@
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Tuple, Type
+
+import openvino as ov
+import torch
+
+from vllm.attention.backends.abstract import (AttentionBackend,
+                                              AttentionMetadata)
+from vllm.attention.backends.utils import CommonAttentionState
+from vllm.multimodal import MultiModalPlaceholderMap
+
+
+def copy_cache_block(src_tensor: ov.Tensor, dst_tensor: ov.Tensor,
+                     src_offset: int, dst_offset: int) -> None:
+
+    def create_roi_tensor(
+        tensor: ov.Tensor,
+        block_number: int,
+    ) -> ov.Tensor:
+        roi_begin = ov.runtime.Coordinate([0, 0, 0, 0])
+        roi_end = ov.runtime.Coordinate(tensor.get_shape())
+
+        roi_begin[0] = block_number
+        roi_end[0] = block_number + 1
+
+        if isinstance(tensor, ov.Tensor):
+            return ov.Tensor(tensor, roi_begin, roi_end)
+        else:
+            return ov.RemoteTensor(tensor, roi_begin, roi_end)
+
+    src_roi_tensor = \
+        create_roi_tensor(src_tensor, src_offset)
+    dst_roi_tensor = \
+        create_roi_tensor(dst_tensor, dst_offset)
+    src_roi_tensor.copy_to(dst_roi_tensor)
+
+
+class OpenVINOAttentionBackend(AttentionBackend):
+
+    @staticmethod
+    def get_name() -> str:
+        return "OPENVINO"
+
+    @staticmethod
+    def get_impl_cls():
+        # OpenVINO implements PagedAttention as part of the Optimum
+        # exported model
+        raise NotImplementedError
+
+    @staticmethod
+    def make_metadata(*args, **kwargs) -> "AttentionMetadata":
+        raise NotImplementedError
+
+    @staticmethod
+    def get_state_cls() -> Type["CommonAttentionState"]:
+        return CommonAttentionState
+
+    @staticmethod
+    def make_openvino_metadata(*args, **kwargs) -> "OpenVINOAttentionMetadata":
+        return OpenVINOAttentionMetadata(*args, **kwargs)
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        return (2, num_blocks, num_kv_heads, block_size, head_size)
+
+    @staticmethod
+    def swap_blocks(
+        src_tensor: ov.Tensor,
+        dst_tensor: ov.Tensor,
+        src_to_dists: List[Tuple[int, int]],
+    ) -> None:
+        for src, dst in src_to_dists:
+            copy_cache_block(src_tensor, dst_tensor, src, dst)
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[Tuple[ov.Tensor, ov.Tensor]],
+        src_to_dists: List[Tuple[int, int]],
+    ) -> None:
+        for src, dst in src_to_dists:
+            for key_cache, value_cache in kv_caches:
+                copy_cache_block(key_cache, key_cache, src, dst)
+                copy_cache_block(value_cache, value_cache, src, dst)
+
+
+@dataclass
+class OpenVINOAttentionMetadata:
+    """Metadata for OpenVINOAttentionBackend.
+
+    Basic terms used below:
+    - batch_size_in_sequences - total number of sequences to execute​
+    - prompt_lens – per sequence size number of scheduled tokens​
+    - batch_size_in_tokens = sum(prompt_lens)​
+    - max_context_len = max(context_lens)​
+    - max_num_blocks = div_up(max_context_len / BLOCK_SIZE)​
+    - num_blocks – total number of blocks in block_indices​
+    """
+
+    # Describes past KV cache size for each sequence within a batch
+    # Shape: [batch_size_in_sequences]
+    # Type: i32​
+    past_lens: torch.Tensor
+
+    # Describes start indices of input / speculative tokens from
+    # current sequences within a batch sequence​
+    # Shape: [batch_size_in_sequences + 1]​
+    # Type: i32
+    subsequence_begins: torch.Tensor
+
+    # Describes block tables for each sequence within a batch​ -
+    # indices along 0th dimension in key_cache and value_cache inputs​
+    # Shape: [num_blocks]
+    # Type: i32​
+    block_indices: torch.Tensor
+
+    # Describes block tables for each sequence within a batch​ -
+    # for i-th element, it is an index in block_indices with the
+    # first block belonging to i-th sequence​
+    # Shape: [batch_size_in_sequences + 1]
+    # Type: i32​
+    block_indices_begins: torch.Tensor
+
+    # Describes max context length
+    # Shape: scalar
+    # Type: i32
+    max_context_len: torch.Tensor
+
+    # The index maps that relate multi-modal embeddings to the corresponding
+    # placeholders.
+    #
+    # N.B. These aren't really related to attention and don't belong on this
+    # type -- this is just a temporary solution to make them available to
+    # `model_executable`.
+    multi_modal_placeholder_index_maps: Optional[Dict[
+        str, MultiModalPlaceholderMap.IndexMap]]

vllm-v0.6.2/vllm/attention/backends/pallas.py

@@ -0,0 +1,323 @@
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Type
+
+import torch
+import torch_xla.experimental.custom_kernel  # Required to register custom ops.
+
+from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
+                                              AttentionMetadata, AttentionType)
+from vllm.attention.backends.utils import CommonAttentionState
+
+
+class PallasAttentionBackend(AttentionBackend):
+
+    @staticmethod
+    def get_name() -> str:
+        return "PALLAS"
+
+    @staticmethod
+    def get_impl_cls() -> Type["PallasAttentionBackendImpl"]:
+        return PallasAttentionBackendImpl
+
+    @staticmethod
+    def get_metadata_cls() -> Type["PallasMetadata"]:
+        return PallasMetadata
+
+    @staticmethod
+    def get_state_cls() -> Type["CommonAttentionState"]:
+        return CommonAttentionState
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        return (num_kv_heads, num_blocks, block_size, head_size)
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: torch.Tensor,
+    ) -> None:
+        raise RuntimeError("swap_blocks is not used for the TPU backend.")
+
+    @torch.compile(backend="openxla")
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[Tuple[torch.Tensor, torch.Tensor]],
+        src_to_dists: Tuple[torch.Tensor, torch.Tensor],
+    ) -> None:
+        src_indices, dst_indices = src_to_dists
+        for k_cache, v_cache in kv_caches:
+            torch.ops.xla.dynamo_set_buffer_donor_(k_cache, True)
+            k_cache[:, dst_indices] = k_cache[:, src_indices]
+            torch.ops.xla.dynamo_set_buffer_donor_(v_cache, True)
+            v_cache[:, dst_indices] = v_cache[:, src_indices]
+
+
+@dataclass
+class PallasMetadata(AttentionMetadata):
+
+    # Currently, input sequences can only contain all prefills
+    # or all decoding.
+    block_tables: Optional[torch.Tensor] = None
+    context_lens: Optional[torch.Tensor] = None
+
+    @property
+    def prefill_metadata(self) -> Optional["PallasMetadata"]:
+        if self.num_prefills == 0:
+            return None
+
+        assert self.num_decode_tokens == 0
+        assert self.block_tables is None
+        assert self.context_lens is None
+        return self
+
+    @property
+    def decode_metadata(self) -> Optional["PallasMetadata"]:
+        if self.num_decode_tokens == 0:
+            return None
+
+        assert self.num_prefills == 0
+        assert self.num_prefill_tokens == 0
+        assert self.block_tables is not None
+        assert self.context_lens is not None
+        return self
+
+
+class PallasAttentionBackendImpl(AttentionImpl):
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: Optional[List[float]],
+        sliding_window: Optional[int],
+        kv_cache_dtype: str,
+        blocksparse_params: Optional[Dict[str, Any]] = None,
+        logits_soft_cap: Optional[float] = None,
+    ) -> None:
+        self.num_heads = num_heads
+        self.head_size = head_size
+        self.scale = float(scale)
+        self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
+
+        assert self.num_heads % self.num_kv_heads == 0
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+        if head_size % 128 != 0:
+            raise NotImplementedError("Head size must be a multiple of 128.")
+        if alibi_slopes is not None:
+            raise NotImplementedError("Alibi slopes is not supported.")
+        if sliding_window is not None:
+            raise NotImplementedError("Sliding window is not supported.")
+        if kv_cache_dtype != "auto":
+            raise NotImplementedError("FP8 KV cache dtype is not supported.")
+        if blocksparse_params is not None:
+            raise NotImplementedError("Blocksparse is not supported.")
+        if logits_soft_cap is not None:
+            raise NotImplementedError(
+                "Attention logits soft-capping is not supported.")
+
+        if torch_xla.tpu.version() < 4:
+            raise NotImplementedError("TPU version must be 4 or higher.")
+
+        self.megacore_mode = None
+        tpu_env = torch_xla.tpu.get_tpu_env()
+        tpu_type = (tpu_env.get("ACCELERATOR_TYPE", None)
+                    or tpu_env.get("TYPE", None)
+                    or tpu_env.get("TPU_ACCELERATOR_TYPE", None))
+        assert tpu_type is not None
+        tpu_type = tpu_type.lower()
+
+        if (("lite" not in tpu_type) and ("v6" not in tpu_type)):
+            if self.num_kv_heads % 2 == 0:
+                self.megacore_mode = "kv_head"
+            else:
+                # NOTE(woosuk): If the batch size is not a multiple of 2, the
+                # megacore mode will be None.
+                self.megacore_mode = "batch"
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: Tuple[torch.Tensor, torch.Tensor],
+        attn_metadata: PallasMetadata,
+        k_scale: float = 1.0,
+        v_scale: float = 1.0,
+        attn_type: AttentionType = AttentionType.DECODER,
+    ) -> torch.Tensor:
+        """Forward pass with Pallas attention.
+
+        Args:
+            query: shape = [batch_size, seq_len, num_heads * head_size]
+            key: shape = [batch_size, seq_len, num_kv_heads * head_size]
+            value: shape = [batch_size, seq_len, num_kv_heads * head_size]
+            kv_cache[0] = [num_kv_heads, num_blocks, block_size, head_size]
+            kv_cache[1] = [num_kv_heads, num_blocks, block_size, head_size]
+                NOTE: kv_cache[0] and kv_cache[1] will be an empty tensor 
+                with shape [0] for profiling run.
+            attn_metadata: Metadata for attention.
+        Returns:
+            shape = [batch_size, seq_len, num_heads * head_size]
+        """
+        assert k_scale == 1.0 and v_scale == 1.0
+        if attn_type != AttentionType.DECODER:
+            raise NotImplementedError("Encoder self-attention and "
+                                      "encoder/decoder cross-attention "
+                                      "are not implemented for "
+                                      "PallasAttentionBackendImpl")
+        batch_size, seq_len, hidden_size = query.shape
+        query = query.view(batch_size, seq_len, self.num_heads, self.head_size)
+        key = key.view(batch_size, seq_len, self.num_kv_heads, self.head_size)
+        value = value.view(batch_size, seq_len, self.num_kv_heads,
+                           self.head_size)
+
+        if kv_cache[0].numel() > 0:
+            slot_mapping = attn_metadata.slot_mapping
+            key_cache, value_cache = kv_cache
+            write_to_kv_cache(key, value, key_cache, value_cache, slot_mapping)
+
+        query = query * self.scale
+        if attn_metadata.num_prefills > 0:
+            assert seq_len % 16 == 0, (
+                "Pallas FlashAttention kernel requires seq_len to be a "
+                f"multiple of 16 but got {seq_len}")
+
+            # Handle GQA/MQA.
+            if self.num_kv_heads != self.num_heads:
+                key = key.repeat_interleave(self.num_queries_per_kv, dim=-2)
+                key = key.view(batch_size, seq_len, self.num_heads,
+                               self.head_size)
+                value = value.repeat_interleave(self.num_queries_per_kv,
+                                                dim=-2)
+                value = value.view(batch_size, seq_len, self.num_heads,
+                                   self.head_size)
+            # FlashAttention requires [batch_size, num_heads, seq_len, d_model]
+            # while the input is [batch_size, seq_len, num_heads, d_model].
+            # Permute the input to match the required format.
+            output = torch.ops.xla.flash_attention(
+                query.permute(0, 2, 1, 3),
+                key.permute(0, 2, 1, 3),
+                value.permute(0, 2, 1, 3),
+                True,
+            )
+            output = output.permute(0, 2, 1, 3)
+        else:
+            # Decoding run.
+            assert kv_cache[0].numel() > 0
+            query = query.squeeze(dim=1)
+            pages_per_compute_block = 16  # TODO(woosuk): Tune this value.
+
+            assert attn_metadata.block_tables is not None
+            assert attn_metadata.context_lens is not None
+            # NOTE(woosuk): The PagedAttention Pallas kernel stores the entire
+            # block table in SMEM. Therefore, if the block table is too large,
+            # the kernel compilation will fail. To avoid this, we split the
+            # batch dimension into smaller chunks and run the kernel multiple
+            # times.
+            MAX_SMEM_USAGE = 512 * 1024
+            size_per_seq = 4 * attn_metadata.block_tables.shape[1]
+            max_num_seq = MAX_SMEM_USAGE // size_per_seq
+
+            if batch_size <= max_num_seq:
+                output = paged_attention(
+                    query,
+                    key_cache,
+                    value_cache,
+                    attn_metadata.context_lens,
+                    attn_metadata.block_tables,
+                    pages_per_compute_block,
+                    self.megacore_mode,
+                )
+            else:
+                chunk_size = max_num_seq
+                # Make sure the chunk size is a multiple of 2.
+                chunk_size = chunk_size // 2 * 2
+                num_chunks = (batch_size + chunk_size - 1) // chunk_size
+
+                output = torch.empty_like(query)
+                for chunk_idx in range(num_chunks):
+                    chunk_start = chunk_idx * chunk_size
+                    chunk_end = chunk_start + chunk_size
+                    # NOTE(woosuk): We skip this line because it causes Dynamo
+                    # compilation error. Instead, we rely on the slice operation
+                    # to handle the out-of-bound case.
+                    # chunk_end = min(chunk_end, batch_size)
+                    chunk_output = paged_attention(
+                        query[chunk_start:chunk_end],
+                        key_cache,
+                        value_cache,
+                        attn_metadata.context_lens[chunk_start:chunk_end],
+                        attn_metadata.block_tables[chunk_start:chunk_end],
+                        pages_per_compute_block,
+                        self.megacore_mode,
+                    )
+                    output[chunk_start:chunk_end] = chunk_output
+
+        # Reshape the output tensor.
+        return output.reshape(batch_size, seq_len, hidden_size)
+
+
+def write_to_kv_cache(
+    key: torch.Tensor,
+    value: torch.Tensor,
+    key_cache: torch.Tensor,
+    value_cache: torch.Tensor,
+    slot_mapping: torch.Tensor,
+) -> None:
+    torch.ops.xla.dynamo_set_buffer_donor_(key_cache, True)
+    torch.ops.xla.dynamo_set_buffer_donor_(value_cache, True)
+
+    key = key.flatten(0, 2)
+    value = value.flatten(0, 2)
+    key_cache = key_cache.flatten(0, 2)
+    value_cache = value_cache.flatten(0, 2)
+    key_cache.index_copy_(0, slot_mapping, key)
+    value_cache.index_copy_(0, slot_mapping, value)
+
+
+def paged_attention(
+    query: torch.Tensor,
+    key_cache: torch.Tensor,
+    value_cache: torch.Tensor,
+    context_lens: torch.Tensor,
+    block_tables: torch.Tensor,
+    pages_per_compute_block: int,
+    megacore_mode: Optional[str],
+) -> torch.Tensor:
+    batch_size = query.shape[0]
+    if megacore_mode == "batch" and batch_size % 2 != 0:
+        megacore_mode = None
+    else:
+        megacore_mode = megacore_mode
+
+    # NOTE(woosuk): A temporary workaround to avoid the error:
+    # "xla::paged_attention() Expected a value of type 'str' for
+    # argument 'megacore_mode' but instead found type 'NoneType'."
+    if megacore_mode is not None:
+        output = torch.ops.xla.paged_attention(
+            query,
+            key_cache,
+            value_cache,
+            context_lens,
+            block_tables,
+            pages_per_compute_block,
+            megacore_mode=megacore_mode,
+        )
+    else:
+        output = torch.ops.xla.paged_attention(
+            query,
+            key_cache,
+            value_cache,
+            context_lens,
+            block_tables,
+            pages_per_compute_block,
+        )
+    return output

vllm-v0.6.2/vllm/attention/backends/placeholder_attn.py

@@ -0,0 +1,341 @@
+from collections import defaultdict
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Type
+
+import torch
+
+from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
+                                              AttentionMetadata,
+                                              AttentionMetadataBuilder)
+from vllm.attention.backends.utils import CommonAttentionState
+from vllm.multimodal import MultiModalPlaceholderMap
+
+if TYPE_CHECKING:
+    from vllm.worker.model_runner import ModelInputForGPUBuilder
+
+# Placeholder attention backend for models like Mamba and embedding models that
+# lack attention.
+
+
+class PlaceholderAttentionBackend(AttentionBackend):
+    """Placeholder backend for when no attention is needed."""
+
+    @staticmethod
+    def get_name() -> str:
+        return "NO_ATTENTION"
+
+    @staticmethod
+    def get_impl_cls() -> Type["PlaceholderAttentionImpl"]:
+        return PlaceholderAttentionImpl
+
+    @staticmethod
+    def get_builder_cls() -> Type["PlaceholderAttentionMetadataBuilder"]:
+        return PlaceholderAttentionMetadataBuilder
+
+    @staticmethod
+    def get_metadata_cls() -> Type["PlaceholderAttentionMetadata"]:
+        return PlaceholderAttentionMetadata
+
+    @staticmethod
+    def get_state_cls() -> Type["CommonAttentionState"]:
+        return CommonAttentionState
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        return (1, 1, 1, 1, 1)
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: torch.Tensor,
+    ) -> None:
+        return
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: torch.Tensor,
+    ) -> None:
+        return
+
+
+@dataclass
+class PlaceholderAttentionMetadata(AttentionMetadata):
+    """Attention metadata for prefill and decode batched together."""
+    # (batch_size,). The sequence length per sequence. Sequence length means
+    # the computed tokens + new tokens None if it is a decoding.
+    seq_lens: Optional[List[int]]
+    # seq_lens stored as a tensor.
+    seq_lens_tensor: Optional[torch.Tensor]
+
+    # Maximum query length in the batch.
+    max_query_len: Optional[int]
+
+    # Max number of query tokens among request in the batch.
+    max_decode_query_len: Optional[int]
+
+    # Maximum sequence length among prefill batch. 0 if there are decoding
+    # requests only.
+    max_prefill_seq_len: int
+    # Maximum sequence length among decode batch. 0 if there are prefill
+    # requests only.
+    max_decode_seq_len: int
+    # (batch_size + 1,). The cumulative subquery lengths of the sequences in
+    # the batch, used to index into subquery. E.g., if the subquery length
+    # is [4, 6], it is [0, 4, 10].
+    query_start_loc: Optional[torch.Tensor]
+    # (batch_size + 1,). The cumulative sequence lengths of the sequences in
+    # the batch, used to index into sequence. E.g., if the sequence length is
+    # [4, 6], it is [0, 4, 10].
+    seq_start_loc: Optional[torch.Tensor]
+    # (batch_size,) A tensor of context lengths (tokens that are computed
+    # so far).
+    context_lens_tensor: Optional[torch.Tensor]
+
+    # (batch_size, max_blocks_per_seq).
+    # Block addresses per sequence. (Seq id -> list of physical block)
+    # E.g., [0, 1, 2] means tokens are stored in 0th, 1st, and 2nd blocks
+    # in the kv cache. Each block can contain up to block_size tokens.
+    # 2nd dimensions are padded up to max_blocks_per_seq if it is cuda-graph
+    # captured.
+    block_tables: Optional[torch.Tensor]
+
+    # Whether or not if cuda graph is enabled.
+    # Cuda-graph is currently enabled for decoding only.
+    # TODO(woosuk): Move `use_cuda_graph` out since it's unrelated to attention.
+    use_cuda_graph: bool
+
+    _cached_prefill_metadata: Optional["PlaceholderAttentionMetadata"] = None
+    _cached_decode_metadata: Optional["PlaceholderAttentionMetadata"] = None
+
+    @property
+    def prefill_metadata(self) -> Optional["PlaceholderAttentionMetadata"]:
+        if self.num_prefills == 0:
+            return None
+
+        if self._cached_prefill_metadata is not None:
+            return self._cached_prefill_metadata
+
+        assert self.seq_lens is not None
+        assert self.seq_lens_tensor is not None
+        assert self.query_start_loc is not None
+        assert self.context_lens_tensor is not None
+        assert self.seq_start_loc is not None
+
+        # Placeholders
+        slot_mapping = torch.empty(0)
+        block_tables = torch.empty(0)
+
+        self._cached_prefill_metadata = PlaceholderAttentionMetadata(
+            num_prefills=self.num_prefills,
+            num_prefill_tokens=self.num_prefill_tokens,
+            num_decode_tokens=0,
+            slot_mapping=slot_mapping,
+            multi_modal_placeholder_index_maps=self.
+            multi_modal_placeholder_index_maps,
+            seq_lens=self.seq_lens[:self.num_prefills],
+            seq_lens_tensor=self.seq_lens_tensor[:self.num_prefills],
+            max_decode_query_len=0,
+            max_query_len=self.max_query_len,
+            max_prefill_seq_len=self.max_prefill_seq_len,
+            max_decode_seq_len=0,
+            query_start_loc=self.query_start_loc[:self.num_prefills + 1],
+            seq_start_loc=self.seq_start_loc[:self.num_prefills + 1],
+            context_lens_tensor=self.context_lens_tensor[:self.num_prefills],
+            block_tables=block_tables,
+            use_cuda_graph=False,
+        )
+        return self._cached_prefill_metadata
+
+    @property
+    def decode_metadata(self) -> Optional["PlaceholderAttentionMetadata"]:
+        if self.num_decode_tokens == 0:
+            return None
+
+        if self._cached_decode_metadata is not None:
+            return self._cached_decode_metadata
+        assert self.seq_lens_tensor is not None
+
+        # Placeholders
+        slot_mapping = torch.empty(0)
+        block_tables = torch.empty(0)
+
+        self._cached_decode_metadata = PlaceholderAttentionMetadata(
+            num_prefills=0,
+            num_prefill_tokens=0,
+            num_decode_tokens=self.num_decode_tokens,
+            slot_mapping=slot_mapping,
+            multi_modal_placeholder_index_maps=None,
+            seq_lens=None,
+            seq_lens_tensor=self.seq_lens_tensor[self.num_prefills:],
+            max_decode_query_len=self.max_decode_query_len,
+            max_query_len=None,
+            max_prefill_seq_len=0,
+            max_decode_seq_len=self.max_decode_seq_len,
+            query_start_loc=None,
+            seq_start_loc=None,
+            context_lens_tensor=None,
+            block_tables=block_tables,
+            use_cuda_graph=self.use_cuda_graph,
+        )
+        return self._cached_decode_metadata
+
+
+class PlaceholderAttentionMetadataBuilder(
+        AttentionMetadataBuilder[PlaceholderAttentionMetadata]):
+
+    def __init__(self, input_builder: "ModelInputForGPUBuilder"):
+        self.prefill_seq_lens: List[int] = []
+        self.context_lens: List[int] = []
+        self.curr_seq_lens: List[int] = []
+        self.multimodal_placeholder_maps: Dict[
+            str,
+            MultiModalPlaceholderMap] = defaultdict(MultiModalPlaceholderMap)
+        self.num_prefills = 0
+        self.num_prefill_tokens = 0
+        self.num_decode_tokens = 0
+
+        self.input_builder = input_builder
+        self.runner = input_builder.runner
+
+    def _add_seq_group(
+            self, inter_data: "ModelInputForGPUBuilder.InterDataForSeqGroup",
+            chunked_prefill_enabled: bool):
+        """Add a sequence group to the metadata. Specifically update/append
+        1. context length.
+        """
+        is_prompt = inter_data.is_prompt
+
+        for (seq_id, token_len, seq_len, curr_seq_len, query_len, context_len,
+             curr_sliding_window_block) in zip(
+                 inter_data.seq_ids, [len(t) for t in inter_data.input_tokens],
+                 inter_data.orig_seq_lens, inter_data.seq_lens,
+                 inter_data.query_lens, inter_data.context_lens,
+                 inter_data.curr_sliding_window_blocks):
+            self.context_lens.append(context_len)
+
+            if is_prompt:
+                mm_maps = inter_data.multi_modal_placeholder_maps
+                if mm_maps:
+                    for modality, placeholders in mm_maps.items():
+                        self.multimodal_placeholder_maps[modality].extend(
+                            placeholders)
+
+                self.num_prefills += 1
+                self.num_prefill_tokens += token_len
+                self.prefill_seq_lens.append(seq_len)
+            else:
+                assert query_len == 1, (
+                    "seq_len: {}, context_len: {}, query_len: {}".format(
+                        seq_len, context_len, query_len))
+                self.num_decode_tokens += query_len
+                self.curr_seq_lens.append(curr_seq_len)
+
+    def build(self, seq_lens: List[int], query_lens: List[int],
+              cuda_graph_pad_size: int, batch_size: int):
+        """Build attention metadata with on-device tensors.
+
+        Args:
+            seq_lens: The maybe padded sequence lengths of the input sequences.
+            query_lens: The query lengths of the input sequences.
+            cuda_graph_pad_size: The padding size for cuda graph.
+                                 -1 if cuda graph is not used.
+            batch_size: The maybe padded batch size.
+        """
+        for inter_data in self.input_builder.inter_data_list:
+            self._add_seq_group(inter_data,
+                                self.input_builder.chunked_prefill_enabled)
+
+        device = self.runner.device
+        use_captured_graph = cuda_graph_pad_size != -1
+
+        logits_soft_cap = getattr(self.runner.model_config.hf_config,
+                                  "attn_logit_softcapping", None)
+        if logits_soft_cap is not None:
+            raise ValueError(
+                "Please use Flashinfer backend for models with logits_soft_cap"
+                " (i.e., Gemma-2). Otherwise, the output might be wrong."
+                " Set Flashinfer backend by "
+                "export VLLM_ATTENTION_BACKEND=FLASHINFER.")
+
+        max_query_len = max(query_lens)
+        decode_query_lens = query_lens[self.num_prefills:]
+        if len(decode_query_lens) > 0:
+            max_decode_query_len = max(decode_query_lens)
+        else:
+            max_decode_query_len = 1
+        max_prefill_seq_len = max(self.prefill_seq_lens, default=0)
+        max_decode_seq_len = max(self.curr_seq_lens, default=0)
+        num_decode_tokens = self.num_decode_tokens
+
+        if use_captured_graph:
+            num_decode_tokens = batch_size
+
+        assert max_query_len > 0, ("query_lens: {}".format(query_lens))
+
+        context_lens_tensor = torch.tensor(self.context_lens,
+                                           dtype=torch.int,
+                                           device=device)
+        seq_lens_tensor = torch.tensor(seq_lens,
+                                       dtype=torch.int,
+                                       device=device)
+        query_lens_tensor = torch.tensor(query_lens,
+                                         dtype=torch.long,
+                                         device=device)
+        query_start_loc = torch.zeros(query_lens_tensor.shape[0] + 1,
+                                      dtype=torch.int32,
+                                      device=device)
+        seq_start_loc = torch.zeros(seq_lens_tensor.shape[0] + 1,
+                                    dtype=torch.int32,
+                                    device=device)
+        placeholder_index_maps = {
+            modality: placeholder_map.index_map()
+            for modality, placeholder_map in
+            self.multimodal_placeholder_maps.items()
+        }
+        torch.cumsum(seq_lens_tensor,
+                     dim=0,
+                     dtype=seq_start_loc.dtype,
+                     out=seq_start_loc[1:])
+        torch.cumsum(query_lens_tensor,
+                     dim=0,
+                     dtype=query_start_loc.dtype,
+                     out=query_start_loc[1:])
+
+        # Placeholders
+        slot_mapping = torch.empty(0)
+        block_tables = torch.empty(0)
+
+        return PlaceholderAttentionMetadata(
+            num_prefills=self.num_prefills,
+            slot_mapping=slot_mapping,
+            multi_modal_placeholder_index_maps=placeholder_index_maps,
+            num_prefill_tokens=self.num_prefill_tokens,
+            num_decode_tokens=num_decode_tokens,
+            seq_lens=seq_lens,
+            seq_lens_tensor=seq_lens_tensor,
+            max_query_len=max_query_len,
+            max_decode_query_len=max_decode_query_len,
+            max_prefill_seq_len=max_prefill_seq_len,
+            max_decode_seq_len=max_decode_seq_len,
+            query_start_loc=query_start_loc,
+            seq_start_loc=seq_start_loc,
+            context_lens_tensor=context_lens_tensor,
+            block_tables=block_tables,
+            use_cuda_graph=use_captured_graph,
+        )
+
+
+class PlaceholderAttentionImpl(AttentionImpl):
+
+    def __init__(self, *args, **kwargs) -> None:
+        return
+
+    def forward(self, *args, **kwargs) -> torch.Tensor:
+        raise NotImplementedError

vllm-v0.6.2/vllm/attention/backends/rocm_flash_attn.py

@@ -0,0 +1,681 @@
+"""Attention layer ROCm GPUs."""
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Type
+
+import torch
+
+import vllm.envs as envs
+from vllm import _custom_ops as ops
+from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
+                                              AttentionMetadata, AttentionType)
+from vllm.attention.backends.utils import (CommonAttentionState,
+                                           CommonMetadataBuilder)
+from vllm.attention.ops.paged_attn import (PagedAttention,
+                                           PagedAttentionMetadata)
+from vllm.logger import init_logger
+from vllm.platforms import current_platform
+
+if TYPE_CHECKING:
+    from vllm.worker.model_runner import ModelInputForGPUWithSamplingMetadata
+
+logger = init_logger(__name__)
+
+_PARTITION_SIZE_ROCM = 512
+_GPU_ARCH = torch.cuda.get_device_properties("cuda").gcnArchName
+_ON_NAVI = "gfx1" in _GPU_ARCH
+_ON_MI250_MI300 = any(arch in _GPU_ARCH
+                      for arch in ["gfx90a", "gfx940", "gfx941", "gfx942"])
+
+
+class ROCmFlashAttentionBackend(AttentionBackend):
+
+    @staticmethod
+    def get_name() -> str:
+        return "ROCM_FLASH"
+
+    @staticmethod
+    def get_impl_cls() -> Type["ROCmFlashAttentionImpl"]:
+        return ROCmFlashAttentionImpl
+
+    @staticmethod
+    def get_metadata_cls() -> Type["AttentionMetadata"]:
+        return ROCmFlashAttentionMetadata
+
+    @staticmethod
+    def get_builder_cls() -> Type["ROCmFlashAttentionMetadataBuilder"]:
+        return ROCmFlashAttentionMetadataBuilder
+
+    @staticmethod
+    def get_state_cls() -> Type["CommonAttentionState"]:
+        return CommonAttentionState
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        return PagedAttention.get_kv_cache_shape(num_blocks, block_size,
+                                                 num_kv_heads, head_size)
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: torch.Tensor,
+    ) -> None:
+        PagedAttention.swap_blocks(src_kv_cache, dst_kv_cache, src_to_dst)
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: torch.Tensor,
+    ) -> None:
+        PagedAttention.copy_blocks(kv_caches, src_to_dists)
+
+
+@dataclass
+class ROCmFlashAttentionMetadata(AttentionMetadata, PagedAttentionMetadata):
+    """Metadata for FlashAttentionBackend.
+
+    NOTE: Any python object stored here is not updated when it is
+    cuda-graph replayed. If you have values that need to be changed
+    dynamically, it should be stored in tensor. The tensor has to be
+    updated from `CUDAGraphRunner.forward` API.
+    """
+    # (batch_size,). The sequence length per sequence. Sequence length means
+    # the computed tokens + new tokens None if it is a decoding.
+    seq_lens: Optional[List[int]]
+    # seq_lens stored as a tensor.
+    seq_lens_tensor: Optional[torch.Tensor]
+
+    # NOTE(sang): Definition of context_len, query_len, and seq_len.
+    # |---------- N-1 iteration --------|
+    # |---------------- N iteration ---------------------|
+    # |- tokenA -|......................|-- newTokens ---|
+    # |---------- context_len ----------|
+    # |-------------------- seq_len ----------------------|
+    #                                   |-- query_len ---|
+
+    # Maximum query length in the batch. None for decoding.
+    max_query_len: Optional[int]
+    # Maximum sequence length among prefill batch. 0 if there are decoding
+    # requests only.
+    max_prefill_seq_len: int
+    # Maximum sequence length among decode batch. 0 if there are prefill
+    # requests only.
+    max_decode_seq_len: int
+    # (batch_size + 1,). The cumulative subquery lengths of the sequences in
+    # the batch, used to index into subquery. E.g., if the subquery length
+    # is [4, 6], it is [0, 4, 10].
+    query_start_loc: Optional[torch.Tensor]
+    # (batch_size + 1,). The cumulative sequence lengths of the sequences in
+    # the batch, used to index into sequence. E.g., if the sequence length is
+    # [4, 6], it is [0, 4, 10].
+    seq_start_loc: Optional[torch.Tensor]
+
+    # Whether or not if cuda graph is enabled.
+    # Cuda-graph is currently enabled for decoding only.
+    # TODO(woosuk): Move `use_cuda_graph` out since it's unrelated to attention.
+    use_cuda_graph: bool
+
+    # (batch_size,) A tensor of context lengths (tokens that are computed
+    # so far).
+    context_lens_tensor: Optional[torch.Tensor]
+
+    # Max number of query tokens among request in the batch.
+    max_decode_query_len: Optional[int] = None
+
+    _cached_prefill_metadata: Optional["ROCmFlashAttentionMetadata"] = None
+    _cached_decode_metadata: Optional["ROCmFlashAttentionMetadata"] = None
+
+    @property
+    def prefill_metadata(self) -> Optional["ROCmFlashAttentionMetadata"]:
+        if self.num_prefills == 0:
+            return None
+
+        if self._cached_prefill_metadata is not None:
+            return self._cached_prefill_metadata
+
+        assert self.seq_lens is not None
+        assert self.seq_lens_tensor is not None
+        assert self.query_start_loc is not None
+        assert self.context_lens_tensor is not None
+        assert self.block_tables is not None
+        assert self.seq_start_loc is not None
+
+        self._cached_prefill_metadata = ROCmFlashAttentionMetadata(
+            num_prefills=self.num_prefills,
+            num_prefill_tokens=self.num_prefill_tokens,
+            num_decode_tokens=0,
+            slot_mapping=self.slot_mapping[:self.num_prefill_tokens],
+            multi_modal_placeholder_index_maps=self.
+            multi_modal_placeholder_index_maps,
+            seq_lens=self.seq_lens[:self.num_prefills],
+            seq_lens_tensor=self.seq_lens_tensor[:self.num_prefills],
+            max_query_len=self.max_query_len,
+            max_prefill_seq_len=self.max_prefill_seq_len,
+            max_decode_seq_len=0,
+            query_start_loc=self.query_start_loc[:self.num_prefills + 1],
+            seq_start_loc=self.seq_start_loc[:self.num_prefills + 1],
+            context_lens_tensor=self.context_lens_tensor[:self.num_prefills],
+            block_tables=self.block_tables[:self.num_prefills],
+            use_cuda_graph=False,
+        )
+        return self._cached_prefill_metadata
+
+    @property
+    def decode_metadata(self) -> Optional["ROCmFlashAttentionMetadata"]:
+        if self.num_decode_tokens == 0:
+            return None
+
+        if self._cached_decode_metadata is not None:
+            return self._cached_decode_metadata
+        assert self.block_tables is not None
+        assert self.seq_lens_tensor is not None
+
+        self._cached_decode_metadata = ROCmFlashAttentionMetadata(
+            num_prefills=0,
+            num_prefill_tokens=0,
+            num_decode_tokens=self.num_decode_tokens,
+            slot_mapping=self.slot_mapping[self.num_prefill_tokens:],
+            multi_modal_placeholder_index_maps=None,
+            seq_lens=None,
+            seq_lens_tensor=self.seq_lens_tensor[self.num_prefills:],
+            max_query_len=None,
+            max_prefill_seq_len=0,
+            max_decode_seq_len=self.max_decode_seq_len,
+            query_start_loc=None,
+            seq_start_loc=None,
+            context_lens_tensor=None,
+            block_tables=self.block_tables[self.num_prefills:],
+            use_cuda_graph=self.use_cuda_graph,
+        )
+        # Batch may be composed of prefill|decodes, adjust query start indices
+        # to refer to the start of decodes when the two are split apart.
+        # E.g. in tokens:[3 prefills|6 decodes], query_start_loc=[3,9] => [0,6].
+        if self._cached_decode_metadata.query_start_loc is not None:
+            qs = self._cached_decode_metadata.query_start_loc
+            self._cached_decode_metadata.query_start_loc = qs - qs[0]
+        return self._cached_decode_metadata
+
+    def advance_step(self,
+                     model_input: "ModelInputForGPUWithSamplingMetadata",
+                     sampled_token_ids: Optional[torch.Tensor],
+                     block_size: int,
+                     num_seqs: int,
+                     num_queries: int,
+                     turn_prefills_into_decodes: bool = False):
+        """
+        Update metadata in-place to advance one decode step.
+        """
+
+        assert not turn_prefills_into_decodes, \
+            ("Chunked prefill is not supported with rocm_flash_attn yet."
+             "turn_prefills_into_decodes is a Multi-Step + Chunked-Prefill "
+             "specific parameter.")
+
+        # When using cudagraph, the num_seqs is padded to the next captured
+        # batch sized, but num_queries tracks the actual number of requests in
+        # the batch. For --enforce-eager mode, num_seqs == num_queries
+        if num_seqs != num_queries:
+            assert num_seqs > num_queries
+            assert self.use_cuda_graph
+
+        assert self.num_prefills == 0
+        assert self.num_prefill_tokens == 0
+        assert self.num_decode_tokens == num_seqs
+        assert self.slot_mapping.shape == (num_seqs, )
+
+        assert self.seq_lens is not None
+        assert len(self.seq_lens) == num_seqs
+        assert self.seq_lens_tensor is not None
+        assert self.seq_lens_tensor.shape == (num_seqs, )
+        assert self.max_query_len == 1
+        assert self.max_prefill_seq_len == 0
+        assert self.max_decode_seq_len == max(self.seq_lens)
+
+        assert self.query_start_loc is not None
+        assert self.query_start_loc.shape == (num_queries + 1, )
+        assert self.seq_start_loc is not None
+        assert self.seq_start_loc.shape == (num_seqs + 1, )
+
+        assert self.context_lens_tensor is not None
+        assert self.context_lens_tensor.shape == (num_queries, )
+
+        assert self.block_tables is not None
+        assert self.block_tables.shape[0] == num_seqs
+
+        # Update query lengths. Note that we update only queries and not seqs,
+        # since tensors may be padded due to captured cuda graph batch size
+        for i in range(num_queries):
+            self.seq_lens[i] += 1
+        self.max_decode_seq_len = max(self.seq_lens)
+
+        ops.advance_step_flashattn(num_seqs=num_seqs,
+                                   num_queries=num_queries,
+                                   block_size=block_size,
+                                   input_tokens=model_input.input_tokens,
+                                   sampled_token_ids=sampled_token_ids,
+                                   input_positions=model_input.input_positions,
+                                   seq_lens=self.seq_lens_tensor,
+                                   slot_mapping=self.slot_mapping,
+                                   block_tables=self.block_tables)
+
+
+class ROCmFlashAttentionMetadataBuilder(
+        CommonMetadataBuilder[ROCmFlashAttentionMetadata]):
+
+    _metadata_cls = ROCmFlashAttentionMetadata
+
+
+def _make_alibi_bias(alibi_slopes: torch.Tensor,
+                     dtype: torch.dtype,
+                     seq_lens: Optional[List[int]],
+                     make_attn_mask: bool = True) -> List[torch.Tensor]:
+    attn_biases = []
+    if seq_lens:
+        for seq_len in seq_lens:
+            bias = torch.arange(seq_len, dtype=dtype)
+            # NOTE(zhuohan): HF uses
+            #     `bias = bias[None, :].repeat(seq_len, 1)`
+            # here. We find that both biases give the same results, but
+            # the bias below more accurately follows the original ALiBi
+            # paper.
+            bias = bias[None, :] - bias[:, None]
+
+            num_heads = alibi_slopes.shape[0]
+            bias = bias[None, :].repeat(
+                (num_heads, 1, 1)).to(alibi_slopes.device)
+            bias.mul_(alibi_slopes[:, None, None])
+            if make_attn_mask:
+                inf_mask = torch.empty(
+                    (1, seq_len, seq_len),
+                    dtype=bias.dtype).fill_(-torch.inf).triu_(diagonal=1).to(
+                        alibi_slopes.device)
+                attn_biases.append((bias + inf_mask).to(dtype))
+            else:
+                attn_biases.append(bias.to(dtype))
+
+    return attn_biases
+
+
+class ROCmFlashAttentionImpl(AttentionImpl):
+    """
+    If the input tensors contain prompt tokens, the layout is as follows:
+    |<--------------- num_prompt_tokens -------------->|
+    |<--prompt_0-->|<--prompt_1-->|...|<--prompt_N-1-->|
+
+    Otherwise, the layout is as follows:
+    |<------------------ num_generation_tokens (M) ----------------->|
+    |<--generation_0-->|..........|<--generation_M-1-->|<--padding-->|
+
+    Generation tokens can contain padding when cuda-graph is used.
+    Currently, prompt tokens don't contain any padding.
+
+    The prompts might have different lengths, while the generation tokens
+    always have length 1.
+
+    If chunked prefill is enabled, prefill tokens and decode tokens can be
+    batched together in a flattened 1D query.
+
+    |<----- num_prefill_tokens ---->|<------- num_decode_tokens ----------->|	
+    |<-prompt_0->|...|<-prompt_N-1->|<-generation_0->|...|<-generation_M-1->|
+
+    Currently, cuda graph is disabled for chunked prefill, meaning there's no
+    padding between prefill and decode tokens.
+    """
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: Optional[List[float]],
+        sliding_window: Optional[int],
+        kv_cache_dtype: str,
+        blocksparse_params: Optional[Dict[str, Any]] = None,
+        logits_soft_cap: Optional[float] = None,
+    ) -> None:
+        if blocksparse_params is not None:
+            raise ValueError(
+                "ROCmFlashAttention does not support blocksparse attention.")
+        if logits_soft_cap is not None:
+            raise ValueError(
+                "ROCmFlashAttention does not support attention logits soft "
+                "capping.")
+        self.num_heads = num_heads
+        self.head_size = head_size
+        self.scale = float(scale)
+        self.num_kv_heads = num_kv_heads
+        if alibi_slopes is not None:
+            alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
+        self.alibi_slopes = alibi_slopes
+        self.sliding_window = ((sliding_window, sliding_window)
+                               if sliding_window is not None else (-1, -1))
+        self.kv_cache_dtype = kv_cache_dtype
+
+        assert self.num_heads % self.num_kv_heads == 0
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+
+        supported_head_sizes = PagedAttention.get_supported_head_sizes()
+        if head_size not in supported_head_sizes:
+            raise ValueError(
+                f"Head size {head_size} is not supported by PagedAttention. "
+                f"Supported head sizes are: {supported_head_sizes}.")
+
+        self.use_naive_attn = False
+        # NOTE: Allow for switching between Triton and CK. Defaulting to triton.
+        self.use_triton_flash_attn = envs.VLLM_USE_TRITON_FLASH_ATTN
+        if self.use_triton_flash_attn:
+            from vllm.attention.ops.triton_flash_attention import (  # noqa: F401
+                triton_attention)
+            self.attn_func = triton_attention
+            logger.debug("Using Triton FA in ROCmBackend")
+            if self.sliding_window != (-1, -1):
+                logger.warning("ROCm Triton FA does not currently support "
+                               "sliding window attention. If using half "
+                               "precision, please try using the ROCm CK "
+                               "FA backend instead by setting the env var "
+                               "`VLLM_USE_TRITON_FLASH_ATTN=0`")
+        else:
+            # if not using triton, navi3x/navi21/navi10 do not use flash-attn
+            # either
+            if not current_platform.has_device_capability(90):
+                self.use_naive_attn = True
+            else:
+                try:
+                    from flash_attn import flash_attn_varlen_func  # noqa: F401
+                    self.attn_func = flash_attn_varlen_func
+                    logger.debug("Using CK FA in ROCmBackend")
+                except ModuleNotFoundError:
+                    self.use_naive_attn = True
+
+            if self.use_naive_attn:
+                self.attn_func = _sdpa_attention
+                logger.debug("Using naive attention in ROCmBackend")
+
+    def repeat_kv(self, x: torch.Tensor, n_rep: int) -> torch.Tensor:
+        """torch.repeat_interleave(x, dim=1, repeats=n_rep)"""
+        tokens, n_kv_heads, head_dim = x.shape
+        return (x[:, :,
+                  None, :].expand(tokens, n_kv_heads, n_rep,
+                                  head_dim).reshape(tokens, n_kv_heads * n_rep,
+                                                    head_dim))
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: ROCmFlashAttentionMetadata,
+        k_scale: float = 1.0,
+        v_scale: float = 1.0,
+        attn_type: AttentionType = AttentionType.DECODER,
+    ) -> torch.Tensor:
+        """Forward pass with FlashAttention and PagedAttention.
+
+        Args:
+            query: shape = [num_tokens, num_heads * head_size]
+            key: shape = [num_tokens, num_kv_heads * head_size]
+            value: shape = [num_tokens, num_kv_heads * head_size]
+            kv_cache = [2, num_blocks, block_size * num_kv_heads * head_size]
+                NOTE: kv_cache will be an empty tensor with shape [0]
+                for profiling run.
+            attn_metadata: Metadata for attention.
+        Returns:
+            shape = [num_tokens, num_heads * head_size]
+        """
+        # Reminder: Please update docs/source/serving/compatibility_matrix.rst
+        # If the feature combo become valid
+        if attn_type != AttentionType.DECODER:
+            raise NotImplementedError("Encoder self-attention and "
+                                      "encoder/decoder cross-attention "
+                                      "are not implemented for "
+                                      "ROCmFlashAttentionImpl")
+
+        num_tokens, hidden_size = query.shape
+        # Reshape the query, key, and value tensors.
+        query = query.view(-1, self.num_heads, self.head_size)
+        key = key.view(-1, self.num_kv_heads, self.head_size)
+        value = value.view(-1, self.num_kv_heads, self.head_size)
+
+        if kv_cache.numel() > 0:
+            key_cache, value_cache = PagedAttention.split_kv_cache(
+                kv_cache, self.num_kv_heads, self.head_size)
+
+            # Reshape the input keys and values and store them in the cache.
+            # If kv_cache is not provided, the new key and value tensors are
+            # not cached. This happens during the initial memory profiling run.
+            PagedAttention.write_to_paged_cache(
+                key,
+                value,
+                key_cache,
+                value_cache,
+                attn_metadata.slot_mapping,
+                self.kv_cache_dtype,
+                k_scale,
+                v_scale,
+            )
+
+        num_prefill_tokens = attn_metadata.num_prefill_tokens
+        num_decode_tokens = attn_metadata.num_decode_tokens
+        assert key.shape[0] == num_prefill_tokens + num_decode_tokens
+        assert value.shape[0] == num_prefill_tokens + num_decode_tokens
+
+        output = torch.empty_like(query)
+        # Query for decode. KV is not needed because it is already cached.
+        decode_query = query[num_prefill_tokens:]
+        # QKV for prefill.
+        query = query[:num_prefill_tokens]
+        key = key[:num_prefill_tokens]
+        value = value[:num_prefill_tokens]
+
+        assert query.shape[0] == num_prefill_tokens
+        assert decode_query.shape[0] == num_decode_tokens
+
+        if prefill_meta := attn_metadata.prefill_metadata:
+            # Prompt run.
+            assert prefill_meta.seq_lens is not None
+            if kv_cache.numel() == 0 or prefill_meta.block_tables.numel() == 0:
+                # triton attention
+                # When block_tables are not filled, it means q and k are the
+                # prompt, and they have the same length.
+                attn_masks = None
+                if self.use_triton_flash_attn:
+                    if self.alibi_slopes is not None:
+                        attn_masks = _make_alibi_bias(
+                            self.alibi_slopes,
+                            query.dtype,
+                            attn_metadata.seq_lens,
+                            make_attn_mask=False)  # type: ignore
+                    out, _ = self.attn_func(
+                        query,
+                        key,
+                        value,
+                        None,
+                        prefill_meta.seq_start_loc,
+                        prefill_meta.seq_start_loc,
+                        prefill_meta.max_prefill_seq_len,
+                        prefill_meta.max_prefill_seq_len,
+                        True,
+                        self.scale,
+                        attn_masks[0][None]
+                        if attn_masks is not None else None,
+                    )
+                elif self.use_naive_attn:
+                    if self.num_kv_heads != self.num_heads:
+                        # Interleave for MQA workaround.
+                        key = self.repeat_kv(key, self.num_queries_per_kv)
+                        value = self.repeat_kv(value, self.num_queries_per_kv)
+                    if self.alibi_slopes is not None:
+                        attn_masks = _make_alibi_bias(
+                            self.alibi_slopes,
+                            query.dtype,
+                            attn_metadata.seq_lens,
+                            make_attn_mask=True)  # type: ignore
+                    query = query.movedim(0, query.dim() - 2)
+                    key = key.movedim(0, key.dim() - 2)
+                    value = value.movedim(0, value.dim() - 2)
+                    # sdpa math backend attention
+                    out = self.attn_func(
+                        query,
+                        key,
+                        value,
+                        prefill_meta.seq_lens,
+                        num_tokens,
+                        self.num_heads,
+                        self.head_size,
+                        self.scale,
+                        attn_masks,
+                    )
+                else:
+                    out = self.attn_func(
+                        q=query,
+                        k=key,
+                        v=value,
+                        cu_seqlens_q=prefill_meta.seq_start_loc,
+                        cu_seqlens_k=prefill_meta.seq_start_loc,
+                        max_seqlen_q=prefill_meta.max_prefill_seq_len,
+                        max_seqlen_k=prefill_meta.max_prefill_seq_len,
+                        softmax_scale=self.scale,
+                        causal=True,
+                        window_size=self.sliding_window,
+                        alibi_slopes=self.alibi_slopes,
+                    )
+
+                # common code for prefill
+                assert output[:num_prefill_tokens].shape == out.shape
+                output[:num_prefill_tokens] = out
+            else:
+                # prefix-enabled attention
+                output[:num_prefill_tokens] = PagedAttention.forward_prefix(
+                    query,
+                    key,
+                    value,
+                    self.kv_cache_dtype,
+                    key_cache,
+                    value_cache,
+                    prefill_meta.block_tables,
+                    prefill_meta.query_start_loc,
+                    prefill_meta.seq_lens_tensor,
+                    prefill_meta.context_lens_tensor,
+                    prefill_meta.max_query_len,
+                    self.alibi_slopes,
+                    self.sliding_window[0],
+                    k_scale,
+                    v_scale,
+                )
+
+        if decode_meta := attn_metadata.decode_metadata:
+            # Decoding run.
+            # Whether to use rocm custom paged attention or not
+            num_seqs, num_heads, head_size = decode_query.shape
+            block_size = value_cache.shape[3]
+            gqa_ratio = num_heads // self.num_kv_heads
+            use_custom = _use_rocm_custom_paged_attention(
+                decode_query.dtype, head_size, block_size, gqa_ratio,
+                decode_meta.max_decode_seq_len)
+            if use_custom:
+                max_seq_len = decode_meta.max_decode_seq_len
+                max_num_partitions = (
+                    (max_seq_len + _PARTITION_SIZE_ROCM - 1) //
+                    _PARTITION_SIZE_ROCM)
+                assert _PARTITION_SIZE_ROCM % block_size == 0
+                tmp_output = torch.empty(
+                    size=(num_seqs, num_heads, max_num_partitions, head_size),
+                    dtype=output.dtype,
+                    device=output.device,
+                )
+                exp_sums = torch.empty(
+                    size=(num_seqs, num_heads, max_num_partitions),
+                    dtype=torch.float32,
+                    device=output.device,
+                )
+                max_logits = torch.empty_like(exp_sums)
+                ops.paged_attention_rocm(
+                    output[num_prefill_tokens:],
+                    exp_sums,
+                    max_logits,
+                    tmp_output,
+                    decode_query,
+                    key_cache,
+                    value_cache,
+                    self.num_kv_heads,
+                    self.scale,
+                    decode_meta.block_tables,
+                    decode_meta.seq_lens_tensor,
+                    block_size,
+                    max_seq_len,
+                    self.alibi_slopes,
+                    self.kv_cache_dtype,
+                    k_scale,
+                    v_scale,
+                )
+            else:
+                output[num_prefill_tokens:] = PagedAttention.forward_decode(
+                    decode_query,
+                    key_cache,
+                    value_cache,
+                    decode_meta.block_tables,
+                    decode_meta.seq_lens_tensor,
+                    decode_meta.max_decode_seq_len,
+                    self.kv_cache_dtype,
+                    self.num_kv_heads,
+                    self.scale,
+                    self.alibi_slopes,
+                    k_scale,
+                    v_scale,
+                )
+
+        # Reshape the output tensor.
+        return output.view(num_tokens, hidden_size)
+
+
+def _sdpa_attention(
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    seq_lens: List[int],
+    num_tokens: int,
+    num_heads: int,
+    head_size: int,
+    scale: float,
+    attn_masks: Optional[List[torch.Tensor]] = None,
+) -> torch.Tensor:
+    start = 0
+    output = torch.empty((num_tokens, num_heads, head_size),
+                         dtype=query.dtype,
+                         device=query.device)
+
+    for i, seq_len in enumerate(seq_lens):
+        end = start + seq_len
+        with torch.backends.cuda.sdp_kernel(enable_math=True,
+                                            enable_flash=False,
+                                            enable_mem_efficient=False):
+            sub_out = torch.nn.functional.scaled_dot_product_attention(
+                query[:, start:end, :],
+                key[:, start:end, :],
+                value[:, start:end, :],
+                dropout_p=0.0,
+                is_causal=attn_masks is None,
+                attn_mask=attn_masks[i] if attn_masks else None,
+                scale=scale).movedim(query.dim() - 2, 0)
+            output[start:end, :, :] = sub_out
+            start = end
+
+    return output
+
+
+def _use_rocm_custom_paged_attention(qtype: torch.dtype, head_size: int,
+                                     block_size: int, gqa_ratio: int,
+                                     max_seq_len: int) -> bool:
+    # rocm custom page attention not support on navi (gfx1*)
+    return (_ON_MI250_MI300 and not _ON_NAVI
+            and (qtype == torch.half or qtype == torch.bfloat16)
+            and (head_size == 64 or head_size == 128)
+            and (block_size == 16 or block_size == 32)
+            and (gqa_ratio >= 1 and gqa_ratio <= 16) and max_seq_len <= 32768)

vllm-v0.6.2/vllm/attention/backends/torch_sdpa.py

@@ -0,0 +1,553 @@
+""" Attention layer with torch scaled_dot_product_attention
+    and PagedAttention."""
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Type
+
+import torch
+from torch.nn.functional import scaled_dot_product_attention
+
+from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
+                                              AttentionMetadata, AttentionType)
+from vllm.attention.backends.utils import CommonAttentionState
+from vllm.attention.ops.paged_attn import PagedAttentionMetadata
+from vllm.platforms import current_platform
+
+if current_platform.is_cpu():
+    try:
+        from vllm.attention.ops.ipex_attn import PagedAttention
+    except ImportError:
+        from vllm.attention.ops.paged_attn import PagedAttention
+else:
+    from vllm.attention.ops.paged_attn import PagedAttention
+
+
+class TorchSDPABackend(AttentionBackend):
+
+    @staticmethod
+    def get_name() -> str:
+        return "TORCH_SDPA"
+
+    @staticmethod
+    def get_impl_cls() -> Type["TorchSDPABackendImpl"]:
+        return TorchSDPABackendImpl
+
+    @staticmethod
+    def get_metadata_cls() -> Type["AttentionMetadata"]:
+        return TorchSDPAMetadata
+
+    @staticmethod
+    def get_state_cls() -> Type["CommonAttentionState"]:
+        return CommonAttentionState
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        return PagedAttention.get_kv_cache_shape(num_blocks, block_size,
+                                                 num_kv_heads, head_size)
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: torch.Tensor,
+    ) -> None:
+        PagedAttention.swap_blocks(src_kv_cache, dst_kv_cache, src_to_dst)
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: torch.Tensor,
+    ) -> None:
+        PagedAttention.copy_blocks(kv_caches, src_to_dists)
+
+
+@dataclass
+class TorchSDPAMetadata(AttentionMetadata, PagedAttentionMetadata):
+    """Metadata for TorchSDPABackend.
+    """
+    # Currently, input sequences can only contain all prompts
+    # or all decoding. True if all sequences are prompts.
+    is_prompt: bool
+    slot_mapping: torch.Tensor
+    seq_lens: Optional[List[int]]
+
+    # Begin encoder attn & enc/dec cross-attn fields...
+    # Encoder sequence lengths representation
+    encoder_seq_lens: Optional[List[int]] = None
+    encoder_seq_lens_tensor: Optional[torch.Tensor] = None
+
+    # Maximum sequence length among encoder sequences
+    max_encoder_seq_len: Optional[int] = None
+
+    # Number of tokens input to encoder
+    num_encoder_tokens: Optional[int] = None
+
+    # Cross-attention memory-mapping data structures: slot mapping
+    # and block tables
+    cross_slot_mapping: Optional[torch.Tensor] = None
+    cross_block_tables: Optional[torch.Tensor] = None
+
+    def __post_init__(self):
+        # Set during the execution of the first attention op.
+        # It is a list because it is needed to set per prompt
+        # when alibi slopes is used. It is because of the limitation
+        # from xformer API.
+        # will not appear in the __repr__ and __init__
+        self.attn_bias: Optional[List[torch.Tensor]] = None
+        self.encoder_attn_bias: Optional[List[torch.Tensor]] = None
+        self.cross_attn_bias: Optional[List[torch.Tensor]] = None
+
+    @property
+    def is_all_encoder_attn_metadata_set(self):
+        '''
+        All attention metadata required for encoder attention is set.
+        '''
+        return ((self.encoder_seq_lens is not None)
+                and (self.encoder_seq_lens_tensor is not None)
+                and (self.max_encoder_seq_len is not None))
+
+    @property
+    def is_all_cross_attn_metadata_set(self):
+        '''
+        All attention metadata required for enc/dec cross-attention is set.
+
+        Superset of encoder attention required metadata.
+        '''
+        return (self.is_all_encoder_attn_metadata_set
+                and (self.cross_slot_mapping is not None)
+                and (self.cross_block_tables is not None))
+
+    @property
+    def prefill_metadata(self) -> Optional["TorchSDPAMetadata"]:
+        # Currently chunked prefill is not supported
+        if self.num_decode_tokens == 0:
+            assert self.num_prefills > 0
+            return self
+
+        return None
+
+    @property
+    def decode_metadata(self) -> Optional["TorchSDPAMetadata"]:
+        # Currently chunked prefill is not supported
+        if self.num_prefills > 0:
+            assert self.num_decode_tokens == 0
+            return None
+
+        return self
+
+    def get_seq_lens(
+        self,
+        attn_type: AttentionType,
+    ):
+        '''
+        Extract appropriate sequence lengths from attention metadata
+        according to attention type.
+
+        Arguments:
+
+        * attn_metadata: Attention metadata structure associated with attention
+        * attn_type: encoder attention, decoder self-attention,
+                    encoder/decoder cross-attention
+
+        Returns:
+        * Appropriate sequence lengths tensor for query
+        * Appropriate sequence lengths tensor for key & value
+        '''
+
+        if (attn_type == AttentionType.DECODER
+                or attn_type == AttentionType.ENCODER_ONLY):
+            seq_lens_q = self.seq_lens
+            seq_lens_kv = self.seq_lens
+        elif attn_type == AttentionType.ENCODER:
+            seq_lens_q = self.encoder_seq_lens
+            seq_lens_kv = self.encoder_seq_lens
+        elif attn_type == AttentionType.ENCODER_DECODER:
+            seq_lens_q = self.seq_lens
+            seq_lens_kv = self.encoder_seq_lens
+        else:
+            raise AttributeError(f"Invalid attention type {str(attn_type)}")
+        return seq_lens_q, seq_lens_kv
+
+    def get_attn_bias(
+        self,
+        attn_type: AttentionType,
+    ) -> Optional[List[torch.Tensor]]:
+        '''
+        Extract appropriate attention bias from attention metadata
+        according to attention type.
+
+        Arguments:
+
+        * attn_metadata: Attention metadata structure associated with attention
+        * attn_type: encoder attention, decoder self-attention,
+                    encoder/decoder cross-attention
+
+        Returns:
+        * Appropriate attention bias value given the attention type
+        '''
+
+        if (attn_type == AttentionType.DECODER
+                or attn_type == AttentionType.ENCODER_ONLY):
+            return self.attn_bias
+        elif attn_type == AttentionType.ENCODER:
+            return self.encoder_attn_bias
+        elif attn_type == AttentionType.ENCODER_DECODER:
+            return self.cross_attn_bias
+        else:
+            raise AttributeError(f"Invalid attention type {str(attn_type)}")
+
+    def set_attn_bias(
+        self,
+        attn_bias: List[torch.Tensor],
+        attn_type: AttentionType,
+    ) -> None:
+        '''
+        Update appropriate attention bias field of attention metadata,
+        according to attention type.
+
+        Arguments:
+
+        * attn_metadata: Attention metadata structure associated with attention
+        * attn_bias: The desired attention bias value
+        * attn_type: encoder attention, decoder self-attention,
+                    encoder/decoder cross-attention
+        '''
+
+        if (attn_type == AttentionType.DECODER
+                or attn_type == AttentionType.ENCODER_ONLY):
+            self.attn_bias = attn_bias
+        elif attn_type == AttentionType.ENCODER:
+            self.encoder_attn_bias = attn_bias
+        elif attn_type == AttentionType.ENCODER_DECODER:
+            self.cross_attn_bias = attn_bias
+        else:
+            raise AttributeError(f"Invalid attention type {str(attn_type)}")
+
+    def get_seq_len_block_table_args(
+        self,
+        attn_type: AttentionType,
+    ) -> tuple:
+        '''
+        The particular choice of sequence-length- and block-table-related
+        attributes which should be extracted from attn_metadata is dependent
+        on the type of attention operation.
+
+        Decoder attn -> select entirely decoder self-attention-related fields
+        Encoder/decoder cross-attn -> select encoder sequence lengths &
+                                    cross-attn block-tables fields
+        Encoder attn -> select encoder sequence lengths fields & no block tables
+
+        Arguments:
+
+        * attn_metadata: Attention metadata structure associated with attention
+        * is_prompt: True if prefill, False otherwise
+        * attn_type: encoder attention, decoder self-attention,
+                    encoder/decoder cross-attention
+
+        Returns:
+
+        * Appropriate sequence-lengths tensor
+        * Appropriate max sequence-length scalar
+        * Appropriate block tables (or None)
+        '''
+
+        if (attn_type == AttentionType.DECODER
+                or attn_type == AttentionType.ENCODER_ONLY):
+            # Decoder self-attention
+            # Choose max_seq_len based on whether we are in prompt_run
+            return (self.seq_lens_tensor, self.max_decode_seq_len,
+                    self.block_tables)
+        elif attn_type == AttentionType.ENCODER_DECODER:
+            # Enc/dec cross-attention KVs match encoder sequence length;
+            # cross-attention utilizes special "cross" block tables
+            return (self.encoder_seq_lens_tensor, self.max_encoder_seq_len,
+                    self.cross_block_tables)
+        elif attn_type == AttentionType.ENCODER:
+            # No block tables associated with encoder attention
+            return (self.encoder_seq_lens_tensor, self.max_encoder_seq_len,
+                    None)
+        else:
+            raise AttributeError(f"Invalid attention type {str(attn_type)}")
+
+
+class TorchSDPABackendImpl(AttentionImpl[TorchSDPAMetadata]):
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: Optional[List[float]],
+        sliding_window: Optional[int],
+        kv_cache_dtype: str,
+        blocksparse_params: Optional[Dict[str, Any]] = None,
+        logits_soft_cap: Optional[float] = None,
+    ) -> None:
+        if blocksparse_params is not None:
+            raise ValueError(
+                "Torch SPDA does not support block-sparse attention.")
+        if logits_soft_cap is not None:
+            raise ValueError("Torch SPDA does not support logits soft cap.")
+        self.num_heads = num_heads
+        self.head_size = head_size
+        self.scale = float(scale)
+        self.num_kv_heads = num_kv_heads
+        if alibi_slopes is not None:
+            alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
+        self.alibi_slopes = alibi_slopes
+        self.sliding_window = sliding_window
+        self.kv_cache_dtype = kv_cache_dtype
+
+        assert self.num_heads % self.num_kv_heads == 0
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+        self.need_mask = (self.alibi_slopes is not None
+                          or self.sliding_window is not None)
+
+        supported_head_sizes = PagedAttention.get_supported_head_sizes()
+        if head_size not in supported_head_sizes:
+            raise ValueError(
+                f"Head size {head_size} is not supported by PagedAttention. "
+                f"Supported head sizes are: {supported_head_sizes}.")
+        if kv_cache_dtype != "auto":
+            raise NotImplementedError(
+                "Torch SDPA backend does not support FP8 KV cache. "
+                "Please use xFormers backend instead.")
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: TorchSDPAMetadata,  # type: ignore
+        k_scale: float = 1.0,
+        v_scale: float = 1.0,
+        attn_type: AttentionType = AttentionType.DECODER,
+    ) -> torch.Tensor:
+        """Forward pass with torch SDPA and PagedAttention.
+
+        Args:
+            query: shape = [num_tokens, num_heads * head_size]
+            key: shape = [num_tokens, num_kv_heads * head_size]
+            value: shape = [num_tokens, num_kv_heads * head_size]
+            kv_cache = [2, num_blocks, block_size * num_kv_heads * head_size]
+                NOTE: kv_cache will be an empty tensor with shape [0]
+                for profiling run.
+            attn_metadata: Metadata for attention.
+        Returns:
+            shape = [num_tokens, num_heads * head_size]
+        """
+        assert k_scale == 1.0 and v_scale == 1.0
+        if (attn_type == AttentionType.ENCODER
+                and (not attn_metadata.is_all_encoder_attn_metadata_set)):
+            raise AttributeError("Encoder attention requires setting "
+                                 "encoder metadata attributes.")
+        elif (attn_type == AttentionType.ENCODER_DECODER
+              and (not attn_metadata.is_all_cross_attn_metadata_set)):
+            raise AttributeError("Encoder/decoder cross-attention "
+                                 "requires setting cross-attention "
+                                 "metadata attributes.")
+
+        # Reshape the query, key, and value tensors.
+        query = query.view(-1, self.num_heads, self.head_size)
+        if key is not None:
+            assert value is not None
+            key = key.view(-1, self.num_kv_heads, self.head_size)
+            value = value.view(-1, self.num_kv_heads, self.head_size)
+        else:
+            assert value is None
+
+        if (attn_type != AttentionType.ENCODER and kv_cache.numel() > 0):
+            # KV-cache during decoder-self- or
+            # encoder-decoder-cross-attention, but not
+            # during encoder attention.
+            #
+            # Even if there are no new key/value pairs to cache,
+            # we still need to break out key_cache and value_cache
+            # i.e. for later use by paged attention
+            key_cache, value_cache = PagedAttention.split_kv_cache(
+                kv_cache, self.num_kv_heads, self.head_size)
+
+            if (key is not None) and (value is not None):
+                if attn_type == AttentionType.ENCODER_DECODER:
+                    # Update cross-attention KV cache (prefill-only)
+                    # During cross-attention decode, key & value will be None,
+                    # preventing this IF-statement branch from running
+                    updated_slot_mapping = attn_metadata.cross_slot_mapping
+                else:
+                    # Update self-attention KV cache (prefill/decode)
+                    updated_slot_mapping = attn_metadata.slot_mapping
+
+                PagedAttention.write_to_paged_cache(key, value, key_cache,
+                                                    value_cache,
+                                                    updated_slot_mapping,
+                                                    self.kv_cache_dtype,
+                                                    k_scale, v_scale)
+
+        if attn_type != AttentionType.ENCODER:
+            # Decoder self-attention supports chunked prefill.
+            # Encoder/decoder cross-attention requires no chunked
+            # prefill (100% prefill or 100% decode tokens, no mix)
+            num_prefill_tokens = attn_metadata.num_prefill_tokens
+            num_decode_tokens = attn_metadata.num_decode_tokens
+        else:
+            # Encoder attention - chunked prefill is not applicable;
+            # derive token-count from query shape & and treat them
+            # as 100% prefill tokens
+            assert attn_metadata.num_encoder_tokens is not None
+            num_prefill_tokens = attn_metadata.num_encoder_tokens
+            num_decode_tokens = 0
+
+        if attn_type == AttentionType.DECODER:
+            # Only enforce this shape-constraint for decoder
+            # self-attention
+            assert key.shape[0] == num_prefill_tokens + num_decode_tokens
+            assert value.shape[0] == num_prefill_tokens + num_decode_tokens
+
+        if prefill_meta := attn_metadata.prefill_metadata:
+            assert attn_metadata.seq_lens is not None
+            if (kv_cache.numel() == 0
+                    or prefill_meta.block_tables.numel() == 0):
+                output = self._run_sdpa_forward(query,
+                                                key,
+                                                value,
+                                                prefill_meta,
+                                                attn_type=attn_type)
+            else:
+                # prefix-enabled attention
+                raise RuntimeError(
+                    "Torch SDPA backend doesn't support prefix decoding.")
+
+        if decode_meta := attn_metadata.decode_metadata:
+            assert attn_type != AttentionType.ENCODER_ONLY, (
+                "Encoder-only models should not have decode metadata.")
+            # Decoding run.
+            (
+                seq_lens_arg,
+                max_seq_len_arg,
+                block_tables_arg,
+            ) = decode_meta.get_seq_len_block_table_args(attn_type)
+
+            output = PagedAttention.forward_decode(
+                query,
+                key_cache,
+                value_cache,
+                block_tables_arg,
+                seq_lens_arg,
+                max_seq_len_arg,
+                self.kv_cache_dtype,
+                self.num_kv_heads,
+                self.scale,
+                self.alibi_slopes,
+                k_scale,
+                v_scale,
+            )
+
+        # Reshape the output tensor.
+        return output.view(-1, self.num_heads * self.head_size)
+
+    def _run_sdpa_forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attn_metadata: TorchSDPAMetadata,
+        attn_type: AttentionType = AttentionType.DECODER,
+    ):
+        if self.num_kv_heads != self.num_heads:
+            key = key.repeat_interleave(self.num_queries_per_kv, dim=1)
+            value = value.repeat_interleave(self.num_queries_per_kv, dim=1)
+
+        attn_masks = attn_metadata.get_attn_bias(attn_type)
+        if attn_masks is None:
+            if self.alibi_slopes is not None:
+                attn_masks = _make_alibi_bias(
+                    self.alibi_slopes, query.dtype,
+                    attn_metadata.seq_lens)  # type: ignore
+            elif self.sliding_window is not None:
+                assert attn_metadata.seq_lens is not None
+                attn_masks = _make_sliding_window_bias(
+                    attn_metadata.seq_lens, self.sliding_window,
+                    query.dtype)  # type: ignore
+            else:
+                seq_lens, _ = attn_metadata.get_seq_lens(attn_type)
+                attn_masks = [None] * len(seq_lens)
+            attn_metadata.set_attn_bias(attn_masks, attn_type)
+
+        output = torch.empty_like(query)
+        query = query.movedim(0, query.dim() - 2)
+        key = key.movedim(0, key.dim() - 2)
+        value = value.movedim(0, value.dim() - 2)
+
+        causal_attn = (attn_type == AttentionType.DECODER)
+
+        seq_lens_q, seq_lens_kv = attn_metadata.get_seq_lens(attn_type)
+        start_q, start_kv = 0, 0
+        for seq_len_q, seq_len_kv, mask in zip(seq_lens_q, seq_lens_kv,
+                                               attn_masks):
+            end_q = start_q + seq_len_q
+            end_kv = start_kv + seq_len_kv
+            sub_out = scaled_dot_product_attention(
+                query[None, :, start_q:end_q, :],
+                key[None, :, start_kv:end_kv, :],
+                value[None, :, start_kv:end_kv, :],
+                attn_mask=mask,
+                dropout_p=0.0,
+                is_causal=causal_attn and not self.need_mask,
+                scale=self.scale).squeeze(0).movedim(query.dim() - 2, 0)
+            output[start_q:end_q, :, :] = sub_out
+            start_q, start_kv = end_q, end_kv
+        return output
+
+
+def _make_alibi_bias(
+    alibi_slopes: torch.Tensor,
+    dtype: torch.dtype,
+    seq_lens: List[int],
+) -> List[torch.Tensor]:
+    attn_biases: List[torch.Tensor] = []
+    for seq_len in seq_lens:
+        bias = torch.arange(seq_len, dtype=dtype)
+        # NOTE(zhuohan): HF uses
+        #     `bias = bias[None, :].repeat(seq_len, 1)`
+        # here. We find that both biases give the same results, but
+        # the bias below more accurately follows the original ALiBi
+        # paper.
+        bias = bias[None, :] - bias[:, None]
+
+        num_heads = alibi_slopes.shape[0]
+        bias = bias[None, :].repeat((num_heads, 1, 1))
+        bias.mul_(alibi_slopes[:, None, None]).unsqueeze_(0)
+        inf_mask = torch.empty(
+            (1, seq_len, seq_len),
+            dtype=bias.dtype).fill_(-torch.inf).triu_(diagonal=1)
+        attn_biases.append((bias + inf_mask).to(dtype))
+
+    return attn_biases
+
+
+def _make_sliding_window_bias(
+    seq_lens: List[int],
+    window_size: Optional[int],
+    dtype: torch.dtype,
+) -> List[torch.Tensor]:
+    attn_biases: List[torch.Tensor] = []
+    for seq_len in seq_lens:
+        tensor = torch.full(
+            (1, seq_len, seq_len),
+            dtype=dtype,
+            fill_value=1,
+        )
+        shift = 0
+        mask = torch.tril(tensor, diagonal=shift).to(dtype)  # type: ignore
+        if window_size is not None:
+            mask = torch.triu(mask, diagonal=shift - window_size + 1)
+        mask = torch.log(mask)
+        attn_biases.append(mask.to(dtype))
+
+    return attn_biases

vllm-v0.6.2/vllm/attention/backends/utils.py

@@ -0,0 +1,574 @@
+"""Attention backend utils"""
+from collections import defaultdict
+from contextlib import contextmanager
+from itertools import accumulate
+from typing import TYPE_CHECKING, Any, Dict, List, Tuple, Type, TypeVar, Union
+
+import numpy as np
+import torch
+
+from vllm.attention import (AttentionMetadata, AttentionMetadataBuilder,
+                            AttentionState)
+from vllm.attention.backends.abstract import AttentionType
+from vllm.multimodal import MultiModalPlaceholderMap
+from vllm.utils import async_tensor_h2d, make_tensor_with_pad
+
+if TYPE_CHECKING:
+    from vllm.worker.model_runner_base import ModelRunnerBase
+
+# Error string(s) for encoder/decoder
+# unsupported attention scenarios
+STR_NOT_IMPL_ENC_DEC_ROCM_HIP = ("ROCm/HIP is not currently supported "
+                                 "with encoder/decoder models.")
+
+PAD_SLOT_ID = -1
+
+# Switch to numpy implementation of compute_slot_mapping
+# if we have at least this many elements. Could be tuned further.
+_COMPUTE_SLOT_MAPPING_NUMPY_NUMEL = 256
+
+if TYPE_CHECKING:
+    from vllm.worker.model_runner import ModelInputForGPUBuilder
+
+
+def is_block_tables_empty(block_tables: Union[None, Dict]):
+    """
+    Check if block_tables is None or a dictionary with all None values.
+    """
+    if block_tables is None:
+        return True
+    return (isinstance(block_tables, dict)
+            and all(value is None for value in block_tables.values()))
+
+
+def compute_slot_mapping_start_idx(is_prompt: bool, query_len: int,
+                                   context_len: int, sliding_window: int):
+    """
+    Compute the start index of slot mapping.
+    """
+    start_idx = 0
+    if is_prompt and sliding_window is not None:
+        start_idx = max(0, query_len - sliding_window)
+    return start_idx
+
+
+def _compute_slot_mapping_python(slot_mapping: List[int],
+                                 block_table: List[int], range_start: int,
+                                 range_end: int, block_size: int):
+    for i in range(range_start, range_end):
+        block_number = block_table[i // block_size]
+        block_offset = i % block_size
+        slot = block_number * block_size + block_offset
+        slot_mapping.append(slot)
+
+
+def _compute_slot_mapping_numpy(slot_mapping: List[int],
+                                block_table: List[int], range_start: int,
+                                range_end: int, block_size: int):
+    block_table_array = np.array(block_table)
+    idx = np.arange(range_start, range_end)
+    block_offset = idx % block_size
+    idx //= block_size
+    seq_slot_mapping_array = block_table_array[idx]
+    seq_slot_mapping_array *= block_size
+    seq_slot_mapping_array += block_offset
+    slot_mapping.extend(seq_slot_mapping_array)
+
+
+def compute_slot_mapping(is_profile_run: bool, slot_mapping: List[int],
+                         seq_id: int, seq_len: int, context_len: int,
+                         start_idx: int, block_size: int,
+                         block_tables: Dict[int, List[int]]):
+    """
+    Compute slot mapping.
+    """
+    if is_profile_run:
+        # During memory profiling, the block tables are not
+        # initialized yet. In this case, we just use a dummy
+        # slot mapping.
+        # In embeddings, the block tables are {seq_id: None}.
+        slot_mapping.extend([PAD_SLOT_ID] * seq_len)
+        return
+
+    # Mask the [0, start_idx) tokens of the prompt with
+    # PAD_SLOT_ID, where start_idx is max(0, seq_len -
+    # sliding_window). For example, if the prompt len is 10,
+    # sliding window is 8, and block size is 4, the first two
+    # tokens are masked and the slot mapping will be
+    # [-1, -1, 2, 3, 4, 5, 6, 7, 0, 1].
+    padding_mask_len = max(0, start_idx - context_len)
+    slot_mapping.extend([PAD_SLOT_ID] * padding_mask_len)
+
+    range_start = max(start_idx, context_len)
+    range_end = seq_len
+    numel = range_end - range_start
+    block_table = block_tables[seq_id]
+
+    # numpy implementation will be faster than python if we have
+    # many elements, otherwise it will be slower.
+    if numel < _COMPUTE_SLOT_MAPPING_NUMPY_NUMEL:
+        _compute_slot_mapping_python(slot_mapping, block_table, range_start,
+                                     range_end, block_size)
+    else:
+        _compute_slot_mapping_numpy(slot_mapping, block_table, range_start,
+                                    range_end, block_size)
+
+
+TAttentionMetadata = TypeVar("TAttentionMetadata", bound='AttentionMetadata')
+
+
+class CommonMetadataBuilder(AttentionMetadataBuilder[TAttentionMetadata]):
+
+    _metadata_cls: Type[TAttentionMetadata]
+
+    def __init__(self, input_builder: "ModelInputForGPUBuilder"):
+        self.slot_mapping: List[int] = []
+        self.prefill_seq_lens: List[int] = []
+        self.context_lens: List[int] = []
+        self.block_tables: List[List[int]] = []
+        self.curr_seq_lens: List[int] = []
+        self.multimodal_placeholder_maps: Dict[
+            str,
+            MultiModalPlaceholderMap] = defaultdict(MultiModalPlaceholderMap)
+        self.num_prefills = 0
+        self.num_prefill_tokens = 0
+        self.num_decode_tokens = 0
+
+        self.input_builder = input_builder
+        self.runner = input_builder.runner
+
+        self.sliding_window = input_builder.sliding_window
+        self.block_size = input_builder.block_size
+
+    def _add_seq_group(
+            self, inter_data: "ModelInputForGPUBuilder.InterDataForSeqGroup",
+            chunked_prefill_enabled: bool):
+        is_prompt = inter_data.is_prompt
+        block_tables = inter_data.block_tables
+
+        for (seq_id, token_len, seq_len, curr_seq_len, query_len, context_len,
+             curr_sliding_window_block) in zip(
+                 inter_data.seq_ids, [len(t) for t in inter_data.input_tokens],
+                 inter_data.orig_seq_lens, inter_data.seq_lens,
+                 inter_data.query_lens, inter_data.context_lens,
+                 inter_data.curr_sliding_window_blocks):
+            self.context_lens.append(context_len)
+            if is_prompt:
+                mm_maps = inter_data.multi_modal_placeholder_maps
+                if mm_maps:
+                    for modality, placeholders in mm_maps.items():
+                        self.multimodal_placeholder_maps[modality].extend(
+                            placeholders)
+
+                self.num_prefills += 1
+                self.num_prefill_tokens += token_len
+                self.prefill_seq_lens.append(seq_len)
+            else:
+                assert query_len == 1, (
+                    "seq_len: {}, context_len: {}, query_len: {}".format(
+                        seq_len, context_len, query_len))
+                self.num_decode_tokens += query_len
+                self.curr_seq_lens.append(curr_seq_len)
+
+            # Compute block table.
+            # TODO(sang): Combine chunked prefill and prefix caching by
+            # only allowing multiple of block_size chunk size.
+            # NOTE: This only works for oooooooxxx style attention.
+            block_table = []
+            if inter_data.prefix_cache_hit:
+                block_table = block_tables[seq_id]
+            elif ((chunked_prefill_enabled or not is_prompt)
+                  and block_tables is not None):
+                if curr_sliding_window_block == 0:
+                    block_table = block_tables[seq_id]
+                else:
+                    block_table = block_tables[seq_id][
+                        -curr_sliding_window_block:]
+            self.block_tables.append(block_table)
+
+            # Compute slot mapping.
+            is_profile_run = is_block_tables_empty(block_tables)
+            start_idx = compute_slot_mapping_start_idx(is_prompt, query_len,
+                                                       context_len,
+                                                       self.sliding_window)
+            compute_slot_mapping(is_profile_run, self.slot_mapping, seq_id,
+                                 seq_len, context_len, start_idx,
+                                 self.block_size, inter_data.block_tables)
+
+    def build(self, seq_lens: List[int], query_lens: List[int],
+              cuda_graph_pad_size: int, batch_size: int):
+        """Build attention metadata with on-device tensors.
+
+        Args:
+            seq_lens: The maybe padded sequence lengths of the input sequences.
+            query_lens: The query lengths of the input sequences.
+            cuda_graph_pad_size: The padding size for cuda graph.
+                                 -1 if cuda graph is not used.
+            batch_size: The maybe padded batch size.
+        """
+        for inter_data in self.input_builder.inter_data_list:
+            self._add_seq_group(inter_data,
+                                self.input_builder.chunked_prefill_enabled)
+
+        device = self.runner.device
+        use_captured_graph = cuda_graph_pad_size != -1
+
+        max_query_len = max(query_lens)
+        max_prefill_seq_len = max(self.prefill_seq_lens, default=0)
+        max_decode_seq_len = max(self.curr_seq_lens, default=0)
+        num_decode_tokens = self.num_decode_tokens
+        query_start_loc = list(accumulate(query_lens, initial=0))
+        seq_start_loc = list(accumulate(seq_lens, initial=0))
+
+        if use_captured_graph:
+            self.slot_mapping.extend([PAD_SLOT_ID] * cuda_graph_pad_size)
+            self.block_tables.extend([] * cuda_graph_pad_size)
+            num_decode_tokens = batch_size
+
+            # The shape of graph_block_tables is
+            # [max batch size, max context len // block size].
+            input_block_tables = self.runner.graph_block_tables[:batch_size]
+            for i, block_table in enumerate(self.block_tables):
+                if block_table:
+                    input_block_tables[i, :len(block_table)] = block_table
+            block_tables = torch.from_numpy(input_block_tables).to(
+                device, non_blocking=True)
+        else:
+            block_tables = make_tensor_with_pad(
+                self.block_tables,
+                pad=0,
+                dtype=torch.int,
+                device=device,
+            )
+        assert max_query_len > 0, "query_lens: {}".format(query_lens)
+
+        assert device is not None
+        context_lens_tensor = async_tensor_h2d(self.context_lens, torch.int,
+                                               device, self.runner.pin_memory)
+        seq_lens_tensor = async_tensor_h2d(seq_lens, torch.int, device,
+                                           self.runner.pin_memory)
+        slot_mapping_tensor = async_tensor_h2d(self.slot_mapping, torch.long,
+                                               device, self.runner.pin_memory)
+        query_start_loc_tensor = async_tensor_h2d(query_start_loc, torch.int32,
+                                                  device,
+                                                  self.runner.pin_memory)
+        seq_start_loc_tensor = async_tensor_h2d(seq_start_loc, torch.int32,
+                                                device, self.runner.pin_memory)
+        placeholder_index_maps = {
+            modality: placeholder_map.index_map()
+            for modality, placeholder_map in
+            self.multimodal_placeholder_maps.items()
+        }
+
+        return self._metadata_cls(  # type: ignore
+            num_prefills=self.num_prefills,
+            slot_mapping=slot_mapping_tensor,
+            multi_modal_placeholder_index_maps=placeholder_index_maps,
+            num_prefill_tokens=self.num_prefill_tokens,
+            num_decode_tokens=num_decode_tokens,
+            seq_lens=seq_lens,
+            seq_lens_tensor=seq_lens_tensor,
+            max_query_len=max_query_len,
+            max_prefill_seq_len=max_prefill_seq_len,
+            max_decode_seq_len=max_decode_seq_len,
+            query_start_loc=query_start_loc_tensor,
+            seq_start_loc=seq_start_loc_tensor,
+            context_lens_tensor=context_lens_tensor,
+            block_tables=block_tables,
+            use_cuda_graph=use_captured_graph,
+        )
+
+
+class CommonAttentionState(AttentionState):
+
+    def __init__(self, runner: "ModelRunnerBase"):
+        self.runner = runner
+        self._is_graph_capturing = False
+
+    @contextmanager
+    def graph_capture(self, max_batch_size: int):
+        self._is_graph_capturing = True
+        self._graph_slot_mapping = torch.full((max_batch_size, ),
+                                              PAD_SLOT_ID,
+                                              dtype=torch.long,
+                                              device=self.runner.device)
+        self._graph_seq_lens = torch.ones(max_batch_size,
+                                          dtype=torch.int32,
+                                          device=self.runner.device)
+        self._graph_block_tables = torch.from_numpy(
+            self.runner.graph_block_tables).to(device=self.runner.device)
+        yield
+        self._is_graph_capturing = False
+        del self._graph_slot_mapping
+        del self._graph_seq_lens
+        del self._graph_block_tables
+
+    def graph_clone(self, batch_size: int) -> "CommonAttentionState":
+        assert self._is_graph_capturing
+        return self.__class__(self.runner)
+
+    def graph_capture_get_metadata_for_batch(
+            self, batch_size: int, is_encoder_decoder_model: bool = False):
+        assert self._is_graph_capturing
+        attn_metadata = self.runner.attn_backend.make_metadata(
+            num_prefills=0,
+            num_prefill_tokens=0,
+            num_decode_tokens=batch_size,
+            slot_mapping=self._graph_slot_mapping[:batch_size],
+            multi_modal_placeholder_index_maps=None,
+            seq_lens=None,
+            seq_lens_tensor=self._graph_seq_lens[:batch_size],
+            max_query_len=1,
+            max_decode_query_len=1,
+            max_prefill_seq_len=0,
+            max_decode_seq_len=self.runner.max_seq_len_to_capture,
+            query_start_loc=None,
+            seq_start_loc=None,
+            context_lens_tensor=None,
+            block_tables=self._graph_block_tables[:batch_size],
+            use_cuda_graph=True,
+        )
+        if is_encoder_decoder_model:
+            # The encoder decoder model works only with XFormers and
+            # Flash Attention backend. Assert the same.
+            assert self.runner.attn_backend.get_name() in\
+                ["XFORMERS", "FLASH_ATTN"], \
+                f"Expected attn_backend name to be either 'XFORMERS' or " \
+                f"'FLASH_ATTN', but "\
+                f"got '{self.runner.attn_backend.get_name()}'"
+            self._update_captured_metadata_for_enc_dec_model(
+                batch_size=batch_size, attn_metadata=attn_metadata)
+
+        return attn_metadata
+
+    def get_graph_input_buffers(
+            self,
+            attn_metadata,
+            is_encoder_decoder_model: bool = False) -> Dict[str, Any]:
+        input_buffers = {
+            "slot_mapping": attn_metadata.slot_mapping,
+            "seq_lens_tensor": attn_metadata.decode_metadata.seq_lens_tensor,
+            "block_tables": attn_metadata.decode_metadata.block_tables,
+        }
+        if is_encoder_decoder_model:
+            # The encoder decoder model works only with XFormers and
+            # Flash Attention backend. Assert the same.
+            assert self.runner.attn_backend.get_name() in\
+                ["XFORMERS", "FLASH_ATTN"], \
+                f"Expected attn_backend name to be either 'XFORMERS' or "\
+                f"'FLASH_ATTN', but "\
+                f"got '{self.runner.attn_backend.get_name()}'"
+            self._add_additonal_input_buffers_for_enc_dec_model(
+                attn_metadata=attn_metadata, input_buffers=input_buffers)
+        return input_buffers
+
+    def prepare_graph_input_buffers(
+            self,
+            input_buffers,
+            attn_metadata,
+            is_encoder_decoder_model: bool = False) -> None:
+        input_buffers["seq_lens_tensor"].copy_(
+            attn_metadata.decode_metadata.seq_lens_tensor, non_blocking=True)
+        input_buffers["block_tables"].copy_(
+            attn_metadata.decode_metadata.block_tables, non_blocking=True)
+        if is_encoder_decoder_model:
+            # The encoder decoder model works only with XFormers and
+            # Flash Attention backend. Assert the same.
+            assert self.runner.attn_backend.get_name() in\
+                ["XFORMERS", "FLASH_ATTN"], \
+                f"Expected attn_backend name to be either 'XFORMERS' or "\
+                f"'FLASH_ATTN', but "\
+                f"got '{self.runner.attn_backend.get_name()}'"
+            self._prepare_input_buffers_for_enc_dec_model(
+                attn_metadata, input_buffers)
+
+    def begin_forward(self, model_input) -> None:
+        return
+
+    def _update_captured_metadata_for_enc_dec_model(self, batch_size: int,
+                                                    attn_metadata):
+        """
+        Updates the attention metadata parameters for CUDA graph capture in an
+        encoder-decoder model.
+
+        This method modifies attention-related tensors and metadata required
+        for CUDA graph capture in encoder-decoder models. Specifically, it
+        updates the cross-attention and encoder sequence tensors in the 
+        AttentionMetadata object.
+        """
+        # During decode phase the cross_slot_mapping will be empty. Hence set
+        # an empty tensor for CUDA Graph capture.
+        attn_metadata.cross_slot_mapping = torch.tensor(
+            [], dtype=torch.int).cuda()
+        attn_metadata.cross_block_tables = torch.full(
+            (batch_size, self.runner.get_max_block_per_batch()),
+            1,
+            dtype=torch.int).cuda()
+        attn_metadata.encoder_seq_lens = torch.full((batch_size, ),
+                                                    1,
+                                                    dtype=torch.int).cuda()
+        attn_metadata.encoder_seq_lens_tensor = torch.full(
+            (batch_size, ), 1, dtype=torch.int).cuda()
+        attn_metadata.max_encoder_seq_len = self.runner.max_seq_len_to_capture
+        attn_metadata.num_encoder_tokens = 0
+
+    def _add_additonal_input_buffers_for_enc_dec_model(
+            self, attn_metadata, input_buffers: Dict[str, Any]):
+        """
+        Saves additional input buffers specific to the encoder-decoder model
+        from the attention metadata.
+
+        This method extracts and stores encoder-decoder related input buffers
+        from the `attn_metadata` into the `input_buffers` dictionary. The
+        buffers include encoder sequence lengths, cross-slot mappings, and
+        cross-block tables, which are essential for the encoder-decoder model
+        during CUDA graph replay.
+        """
+        input_buffers["encoder_seq_lens_tensor"] = (
+            attn_metadata.decode_metadata.encoder_seq_lens_tensor)
+        input_buffers["cross_slot_mapping"] = (
+            attn_metadata.decode_metadata.cross_slot_mapping)
+        input_buffers["cross_block_tables"] = (
+            attn_metadata.decode_metadata.cross_block_tables)
+
+    def _prepare_input_buffers_for_enc_dec_model(self, attn_metadata,
+                                                 input_buffers: Dict[str,
+                                                                     Any]):
+        """
+        Populates input buffers with data from the encoder-decoder model's
+        attention metadata.
+
+        This method fills the input buffers with encoder-decoder specific
+        tensors. It copies data from the `attn_metadata` and keyword arguments
+        (`kwargs`) into corresponding buffers in the `input_buffers` dictionary.
+        The copied data includes attention-related metadata as well as input 
+        IDs and positional information for the encoder.
+        """
+        input_buffers["encoder_seq_lens_tensor"].copy_(
+            attn_metadata.decode_metadata.encoder_seq_lens_tensor,
+            non_blocking=True)
+        input_buffers["cross_slot_mapping"].copy_(
+            attn_metadata.decode_metadata.cross_slot_mapping,
+            non_blocking=True)
+        input_buffers["cross_block_tables"].copy_(
+            attn_metadata.decode_metadata.cross_block_tables,
+            non_blocking=True)
+
+
+def is_all_encoder_attn_metadata_set(attn_metadata):
+    '''
+    All attention metadata required for encoder attention is set.
+    '''
+    return ((attn_metadata.encoder_seq_lens is not None)
+            and (attn_metadata.encoder_seq_lens_tensor is not None)
+            and (attn_metadata.max_encoder_seq_len is not None))
+
+
+def is_all_cross_attn_metadata_set(attn_metadata):
+    '''
+    All attention metadata required for enc/dec cross-attention is set.
+
+    Superset of encoder attention required metadata.
+    '''
+    return (attn_metadata.is_all_encoder_attn_metadata_set
+            and (attn_metadata.cross_slot_mapping is not None)
+            and (attn_metadata.cross_block_tables is not None))
+
+
+def get_seq_len_block_table_args(
+    attn_metadata,
+    is_prompt: bool,
+    attn_type: AttentionType,
+) -> tuple:
+    '''
+    The particular choice of sequence-length- and block-table-related
+    attributes which should be extracted from attn_metadata is dependent
+    on the type of attention operation.
+
+    Decoder attn -> select entirely decoder self-attention-related fields
+    Encoder/decoder cross-attn -> select encoder sequence lengths & 
+                                  cross-attn block-tables fields
+    Encoder attn -> select encoder sequence lengths fields & no block tables
+    
+    Arguments:
+
+    * attn_metadata: Attention metadata structure associated with attention op
+    * is_prompt: True if prefill, False otherwise
+    * attn_type: encoder attention, decoder self-attention,
+                 encoder/decoder cross-attention
+
+    Returns:
+
+    * Appropriate sequence-lengths tensor
+    * Appropriate max sequence-length scalar
+    * Appropriate block tables (or None)
+    '''
+
+    if attn_type == AttentionType.DECODER:
+        # Decoder self-attention
+        # Choose max_seq_len based on whether we are in prompt_run
+        if is_prompt:
+            max_seq_len = attn_metadata.max_prefill_seq_len
+        else:
+            max_seq_len = attn_metadata.max_decode_seq_len
+        return (attn_metadata.seq_lens_tensor, max_seq_len,
+                attn_metadata.block_tables)
+    elif attn_type == AttentionType.ENCODER_DECODER:
+        # Enc/dec cross-attention KVs match encoder sequence length;
+        # cross-attention utilizes special "cross" block tables
+        return (attn_metadata.encoder_seq_lens_tensor,
+                attn_metadata.max_encoder_seq_len,
+                attn_metadata.cross_block_tables)
+    elif attn_type == AttentionType.ENCODER:
+        # No block tables associated with encoder attention
+        return (attn_metadata.encoder_seq_lens_tensor,
+                attn_metadata.max_encoder_seq_len, None)
+    else:
+        raise AttributeError(f"Invalid attention type {str(attn_type)}")
+
+
+def get_num_prefill_decode_query_kv_tokens(
+    attn_metadata,
+    attn_type: AttentionType,
+) -> Tuple[int, int, int]:
+    """
+    Calculate the number of prefill and decode tokens for query, key/value
+    based on the attention metadata and the specified attention type.
+
+    Args:
+        attn_metadata (FlashAttentionMetadata): Attention Metadata object.
+        attn_type (AttentionType): The type of attention being used.
+    Returns:
+        Tuple[int, int, int]: A tuple containing three integers:
+            - The number of prefill query tokens.
+            - The number of prefill key/value tokens.
+            - The number of decode query tokens.
+
+    Raises:
+        AssertionError: If the number of encoder tokens in `attn_metadata` 
+        is `None` when required for the calculations.
+    """
+    num_prefill_query_tokens = 0
+    num_decode_query_tokens = 0
+    num_prefill_kv_tokens = 0
+    if attn_type == AttentionType.ENCODER:
+        # Encoder attention is only invoked during prefill phase.
+        # The same input servers a both query and key.
+        assert attn_metadata.num_encoder_tokens is not None
+        num_prefill_query_tokens = attn_metadata.num_encoder_tokens
+        num_prefill_kv_tokens = attn_metadata.num_encoder_tokens
+        num_decode_query_tokens = 0
+    elif attn_type == AttentionType.ENCODER_DECODER:
+        assert attn_metadata.num_encoder_tokens is not None
+        num_prefill_query_tokens = attn_metadata.num_prefill_tokens
+        # The key is the encoder/cross-attention.
+        num_prefill_kv_tokens = attn_metadata.num_encoder_tokens
+        num_decode_query_tokens = attn_metadata.num_decode_tokens
+    else:  # attn_type == AttentionType.DECODER or
+        # attn_type == AttentionType.ENCODER_ONLY
+        num_prefill_query_tokens = attn_metadata.num_prefill_tokens
+        num_prefill_kv_tokens = attn_metadata.num_prefill_tokens
+        num_decode_query_tokens = attn_metadata.num_decode_tokens
+
+    return (num_prefill_query_tokens, num_prefill_kv_tokens,
+            num_decode_query_tokens)

vllm-v0.6.2/vllm/attention/backends/xformers.py

@@ -0,0 +1,787 @@
+"""Attention layer with xFormers and PagedAttention."""
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Type
+
+import torch
+from xformers import ops as xops
+from xformers.ops.fmha.attn_bias import (AttentionBias,
+                                         BlockDiagonalCausalMask,
+                                         BlockDiagonalMask,
+                                         LowerTriangularMaskWithTensorBias)
+
+from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
+                                              AttentionMetadata, AttentionType)
+from vllm.attention.backends.utils import (
+    CommonAttentionState, CommonMetadataBuilder,
+    get_num_prefill_decode_query_kv_tokens, get_seq_len_block_table_args,
+    is_all_cross_attn_metadata_set, is_all_encoder_attn_metadata_set)
+from vllm.attention.ops.paged_attn import (PagedAttention,
+                                           PagedAttentionMetadata)
+from vllm.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+class XFormersBackend(AttentionBackend):
+
+    @staticmethod
+    def get_name() -> str:
+        return "XFORMERS"
+
+    @staticmethod
+    def get_impl_cls() -> Type["XFormersImpl"]:
+        return XFormersImpl
+
+    @staticmethod
+    def get_metadata_cls() -> Type["AttentionMetadata"]:
+        return XFormersMetadata
+
+    @staticmethod
+    def get_builder_cls() -> Type["XFormersMetadataBuilder"]:
+        return XFormersMetadataBuilder
+
+    @staticmethod
+    def get_state_cls() -> Type["CommonAttentionState"]:
+        return CommonAttentionState
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        return PagedAttention.get_kv_cache_shape(num_blocks, block_size,
+                                                 num_kv_heads, head_size)
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: Dict[int, int],
+    ) -> None:
+        PagedAttention.swap_blocks(src_kv_cache, dst_kv_cache, src_to_dst)
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: torch.Tensor,
+    ) -> None:
+        PagedAttention.copy_blocks(kv_caches, src_to_dists)
+
+
+@dataclass
+class XFormersMetadata(AttentionMetadata, PagedAttentionMetadata):
+    """Metadata for XFormersbackend.
+
+    NOTE: Any python object stored here is not updated when it is
+    cuda-graph replayed. If you have values that need to be changed
+    dynamically, it should be stored in tensor. The tensor has to be
+    updated from `CUDAGraphRunner.forward` API.
+    """
+
+    # |---------- N-1 iteration --------|
+    # |---------------- N iteration ---------------------|
+    # |- tokenA -|......................|-- newTokens ---|
+    # |---------- context_len ----------|
+    # |-------------------- seq_len ----------------------|
+    #                                   |-- query_len ---|
+
+    # seq_lens stored as a tensor.
+    seq_lens_tensor: Optional[torch.Tensor]
+
+    # FIXME: It is for flash attn.
+    # Maximum sequence length among prefill batch. 0 if there are decoding
+    # requests only.
+    max_prefill_seq_len: int
+    # Maximum sequence length among decode batch. 0 if there are prefill
+    # requests only.
+    max_decode_seq_len: int
+
+    # Whether or not if cuda graph is enabled.
+    # Cuda-graph is currently enabled for decoding only.
+    # TODO(woosuk): Move `use_cuda_graph` out since it's unrelated to attention.
+    use_cuda_graph: bool
+
+    # (batch_size,). The sequence length per sequence. Sequence length means
+    # the computed tokens + new tokens None if it is a decoding.
+    seq_lens: Optional[List[int]] = None
+
+    # FIXME: It is for flash attn.
+    # (batch_size + 1,). The cumulative sequence lengths of the sequences in
+    # the batch, used to index into sequence. E.g., if the sequence length is
+    # [4, 6], it is [0, 4, 10].
+    seq_start_loc: Optional[torch.Tensor] = None
+
+    # (batch_size,) A tensor of context lengths (tokens that are computed
+    # so far).
+    context_lens_tensor: Optional[torch.Tensor] = None
+
+    # Maximum query length in the batch. None for decoding.
+    max_query_len: Optional[int] = None
+
+    # Max number of query tokens among request in the batch.
+    max_decode_query_len: Optional[int] = None
+
+    # (batch_size + 1,). The cumulative subquery lengths of the sequences in
+    # the batch, used to index into subquery. E.g., if the subquery length
+    # is [4, 6], it is [0, 4, 10].
+    query_start_loc: Optional[torch.Tensor] = None
+
+    # Self-attention prefill/decode metadata cache
+    _cached_prefill_metadata: Optional["XFormersMetadata"] = None
+    _cached_decode_metadata: Optional["XFormersMetadata"] = None
+
+    # Begin encoder attn & enc/dec cross-attn fields...
+
+    # Encoder sequence lengths representation
+    encoder_seq_lens: Optional[List[int]] = None
+    encoder_seq_lens_tensor: Optional[torch.Tensor] = None
+    # FIXME: It is for flash attn.
+    # (batch_size + 1,). The cumulative sequence lengths of the sequences in
+    # the batch, used to index into sequence. E.g., if the sequence length is
+    # [4, 6], it is [0, 4, 10].
+    encoder_seq_start_loc: Optional[torch.Tensor] = None
+
+    # Maximum sequence length among encoder sequences
+    max_encoder_seq_len: Optional[int] = None
+
+    # Number of tokens input to encoder
+    num_encoder_tokens: Optional[int] = None
+
+    # Cross-attention memory-mapping data structures: slot mapping
+    # and block tables
+    cross_slot_mapping: Optional[torch.Tensor] = None
+    cross_block_tables: Optional[torch.Tensor] = None
+
+    def __post_init__(self):
+        # Set during the execution of the first attention op.
+        # It is a list because it is needed to set per prompt
+        # when alibi slopes is used. It is because of the limitation
+        # from xformer API.
+        # will not appear in the __repr__ and __init__
+        self.attn_bias: Optional[List[AttentionBias]] = None
+        self.encoder_attn_bias: Optional[List[AttentionBias]] = None
+        self.cross_attn_bias: Optional[List[AttentionBias]] = None
+
+    @property
+    def is_all_encoder_attn_metadata_set(self):
+        '''
+        All attention metadata required for encoder attention is set.
+        '''
+        return is_all_encoder_attn_metadata_set(self)
+
+    @property
+    def is_all_cross_attn_metadata_set(self):
+        '''
+        All attention metadata required for enc/dec cross-attention is set.
+
+        Superset of encoder attention required metadata.
+        '''
+        return is_all_cross_attn_metadata_set(self)
+
+    @property
+    def prefill_metadata(self) -> Optional["XFormersMetadata"]:
+        if self.num_prefills == 0:
+            return None
+
+        if self._cached_prefill_metadata is not None:
+            # Recover cached prefill-phase attention
+            # metadata structure
+            return self._cached_prefill_metadata
+
+        assert ((self.seq_lens is not None)
+                or (self.encoder_seq_lens is not None))
+        assert ((self.seq_lens_tensor is not None)
+                or (self.encoder_seq_lens_tensor is not None))
+
+        # Compute some attn_metadata fields which default to None
+        query_start_loc = (None if self.query_start_loc is None else
+                           self.query_start_loc[:self.num_prefills + 1])
+        slot_mapping = (None if self.slot_mapping is None else
+                        self.slot_mapping[:self.num_prefill_tokens])
+        seq_lens = (None if self.seq_lens is None else
+                    self.seq_lens[:self.num_prefills])
+        seq_lens_tensor = (None if self.seq_lens_tensor is None else
+                           self.seq_lens_tensor[:self.num_prefills])
+        context_lens_tensor = (None if self.context_lens_tensor is None else
+                               self.context_lens_tensor[:self.num_prefills])
+        block_tables = (None if self.block_tables is None else
+                        self.block_tables[:self.num_prefills])
+
+        # Construct & cache prefill-phase attention metadata structure
+        self._cached_prefill_metadata = XFormersMetadata(
+            num_prefills=self.num_prefills,
+            num_prefill_tokens=self.num_prefill_tokens,
+            num_decode_tokens=0,
+            slot_mapping=slot_mapping,
+            multi_modal_placeholder_index_maps=self.
+            multi_modal_placeholder_index_maps,
+            seq_lens=seq_lens,
+            seq_lens_tensor=seq_lens_tensor,
+            max_query_len=self.max_query_len,
+            max_prefill_seq_len=self.max_prefill_seq_len,
+            max_decode_seq_len=0,
+            query_start_loc=query_start_loc,
+            context_lens_tensor=context_lens_tensor,
+            block_tables=block_tables,
+            use_cuda_graph=False,
+            # Begin encoder & cross attn fields below...
+            encoder_seq_lens=self.encoder_seq_lens,
+            encoder_seq_lens_tensor=self.encoder_seq_lens_tensor,
+            max_encoder_seq_len=self.max_encoder_seq_len,
+            cross_slot_mapping=self.cross_slot_mapping,
+            cross_block_tables=self.cross_block_tables)
+        return self._cached_prefill_metadata
+
+    @property
+    def decode_metadata(self) -> Optional["XFormersMetadata"]:
+        if self.num_decode_tokens == 0:
+            return None
+
+        if self._cached_decode_metadata is not None:
+            # Recover cached decode-phase attention
+            # metadata structure
+            return self._cached_decode_metadata
+        assert ((self.seq_lens_tensor is not None)
+                or (self.encoder_seq_lens_tensor is not None))
+
+        # Compute some attn_metadata fields which default to None
+        slot_mapping = (None if self.slot_mapping is None else
+                        self.slot_mapping[self.num_prefill_tokens:])
+        seq_lens_tensor = (None if self.seq_lens_tensor is None else
+                           self.seq_lens_tensor[self.num_prefills:])
+        block_tables = (None if self.block_tables is None else
+                        self.block_tables[self.num_prefills:])
+
+        # Construct & cache decode-phase attention metadata structure
+        self._cached_decode_metadata = XFormersMetadata(
+            num_prefills=0,
+            num_prefill_tokens=0,
+            num_decode_tokens=self.num_decode_tokens,
+            slot_mapping=slot_mapping,
+            multi_modal_placeholder_index_maps=None,
+            seq_lens_tensor=seq_lens_tensor,
+            max_prefill_seq_len=0,
+            max_decode_seq_len=self.max_decode_seq_len,
+            block_tables=block_tables,
+            use_cuda_graph=self.use_cuda_graph,
+            # Begin encoder & cross attn fields below...
+            encoder_seq_lens=self.encoder_seq_lens,
+            encoder_seq_lens_tensor=self.encoder_seq_lens_tensor,
+            max_encoder_seq_len=self.max_encoder_seq_len,
+            cross_slot_mapping=self.cross_slot_mapping,
+            cross_block_tables=self.cross_block_tables)
+
+        # Batch may be composed of prefill|decodes, adjust query start indices
+        # to refer to the start of decodes when the two are split apart.
+        # E.g. in tokens:[3 prefills|6 decodes], query_start_loc=[3,9] => [0,6].
+        if self._cached_decode_metadata.query_start_loc is not None:
+            qs = self._cached_decode_metadata.query_start_loc
+            self._cached_decode_metadata.query_start_loc = qs - qs[0]
+        return self._cached_decode_metadata
+
+
+def _get_attn_bias(
+    attn_metadata: XFormersMetadata,
+    attn_type: AttentionType,
+) -> Optional[AttentionBias]:
+    '''
+    Extract appropriate attention bias from attention metadata
+    according to attention type.
+
+    Arguments:
+
+    * attn_metadata: Attention metadata structure associated with attention
+    * attn_type: encoder attention, decoder self-attention,
+                 encoder/decoder cross-attention
+
+    Returns:
+    * Appropriate attention bias value given the attention type
+    '''
+
+    if (attn_type == AttentionType.DECODER
+            or attn_type == AttentionType.ENCODER_ONLY):
+        return attn_metadata.attn_bias
+    elif attn_type == AttentionType.ENCODER:
+        return attn_metadata.encoder_attn_bias
+    elif attn_type == AttentionType.ENCODER_DECODER:
+        return attn_metadata.cross_attn_bias
+    else:
+        raise AttributeError(f"Invalid attention type {str(attn_type)}")
+
+
+def _set_attn_bias(
+    attn_metadata: XFormersMetadata,
+    attn_bias: List[Optional[AttentionBias]],
+    attn_type: AttentionType,
+) -> None:
+    '''
+    Update appropriate attention bias field of attention metadata,
+    according to attention type.
+
+    Arguments:
+
+    * attn_metadata: Attention metadata structure associated with attention
+    * attn_bias: The desired attention bias value
+    * attn_type: encoder attention, decoder self-attention,
+                 encoder/decoder cross-attention
+    '''
+
+    if (attn_type == AttentionType.DECODER
+            or attn_type == AttentionType.ENCODER_ONLY):
+        attn_metadata.attn_bias = attn_bias
+    elif attn_type == AttentionType.ENCODER:
+        attn_metadata.encoder_attn_bias = attn_bias
+    elif attn_type == AttentionType.ENCODER_DECODER:
+        attn_metadata.cross_attn_bias = attn_bias
+    else:
+        raise AttributeError(f"Invalid attention type {str(attn_type)}")
+
+
+class XFormersMetadataBuilder(CommonMetadataBuilder[XFormersMetadata]):
+
+    _metadata_cls = XFormersMetadata
+
+
+class XFormersImpl(AttentionImpl[XFormersMetadata]):
+    """
+    If the input tensors contain prompt tokens, the layout is as follows:
+    |<--------------- num_prefill_tokens ----------------->|	
+    |<--prefill_0-->|<--prefill_1-->|...|<--prefill_N-1--->|
+
+    Otherwise, the layout is as follows:	
+    |<----------------- num_decode_tokens ------------------>|	
+    |<--decode_0-->|..........|<--decode_M-1-->|<--padding-->|
+
+    Generation tokens can contain padding when cuda-graph is used.
+    Currently, prompt tokens don't contain any padding.
+
+    The prompts might have different lengths, while the generation tokens
+    always have length 1.
+
+    If chunked prefill is enabled, prefill tokens and decode tokens can be
+    batched together in a flattened 1D query.
+
+    |<----- num_prefill_tokens ---->|<------- num_decode_tokens --------->|
+    |<-prefill_0->|...|<-prefill_N-1->|<--decode_0-->|...|<--decode_M-1-->|
+
+    Currently, cuda graph is disabled for chunked prefill, meaning there's no
+    padding between prefill and decode tokens.
+    """
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: int,
+        alibi_slopes: Optional[List[float]],
+        sliding_window: Optional[int],
+        kv_cache_dtype: str,
+        blocksparse_params: Optional[Dict[str, Any]] = None,
+        logits_soft_cap: Optional[float] = None,
+    ) -> None:
+        if blocksparse_params is not None:
+            raise ValueError(
+                "XFormers does not support block-sparse attention.")
+        if logits_soft_cap is not None:
+            raise ValueError(
+                "XFormers does not support attention logits soft capping.")
+        self.num_heads = num_heads
+        self.head_size = head_size
+        self.scale = float(scale)
+        self.num_kv_heads = num_kv_heads
+        if alibi_slopes is not None:
+            alibi_slopes = torch.tensor(alibi_slopes, dtype=torch.float32)
+        self.alibi_slopes = alibi_slopes
+        self.sliding_window = sliding_window
+        self.kv_cache_dtype = kv_cache_dtype
+
+        assert self.num_heads % self.num_kv_heads == 0
+        self.num_queries_per_kv = self.num_heads // self.num_kv_heads
+
+        suppored_head_sizes = PagedAttention.get_supported_head_sizes()
+        if head_size not in suppored_head_sizes:
+            raise ValueError(
+                f"Head size {head_size} is not supported by PagedAttention. "
+                f"Supported head sizes are: {suppored_head_sizes}.")
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: Optional[torch.Tensor],
+        value: Optional[torch.Tensor],
+        kv_cache: torch.Tensor,
+        attn_metadata: "XFormersMetadata",
+        k_scale: float = 1.0,
+        v_scale: float = 1.0,
+        attn_type: AttentionType = AttentionType.DECODER,
+    ) -> torch.Tensor:
+        """Forward pass with xFormers and PagedAttention.
+
+        For decoder-only models: query, key and value must be non-None.
+
+        For encoder/decoder models:
+        * XFormersImpl.forward() may be invoked for both self- and cross-
+          attention layers.
+        * For self-attention: query, key and value must be non-None.
+        * For cross-attention:
+            * Query must be non-None
+            * During prefill, key and value must be non-None; key and value
+              get cached for use during decode.
+            * During decode, key and value may be None, since:
+              (1) key and value tensors were cached during prefill, and
+              (2) cross-attention key and value tensors do not grow during
+                  decode
+        
+        A note on how the attn_type (attention type enum) argument impacts
+        attention forward() behavior:
+    
+            * DECODER: normal decoder-only behavior;
+                use decoder self-attention block table
+            * ENCODER: no KV caching; pass encoder sequence
+                attributes (encoder_seq_lens/encoder_seq_lens_tensor/
+                max_encoder_seq_len) to kernel, in lieu of decoder
+                sequence attributes (seq_lens/seq_lens_tensor/max_seq_len).
+                Used for encoder branch of encoder-decoder models.
+            * ENCODER_ONLY: no kv_caching, uses the normal attention 
+                attributes (seq_lens/seq_lens_tensor/max_seq_len).
+            * ENCODER_DECODER: cross-attention behavior;
+                use cross-attention block table for caching KVs derived
+                from encoder hidden states; since KV sequence lengths
+                will match encoder sequence lengths, pass encoder sequence
+                attributes to kernel (encoder_seq_lens/encoder_seq_lens_tensor/
+                max_encoder_seq_len)
+    
+        Args:
+            query: shape = [num_tokens, num_heads * head_size]
+            key: shape = [num_tokens, num_kv_heads * head_size]
+            value: shape = [num_tokens, num_kv_heads * head_size]
+            kv_cache = [2, num_blocks, block_size * num_kv_heads * head_size]
+                NOTE: kv_cache will be an empty tensor with shape [0]
+                for profiling run.
+            attn_metadata: Metadata for attention.
+            attn_type: Select attention type, between encoder attention,
+                       decoder self-attention, or encoder/decoder cross-
+                       attention. Defaults to decoder self-attention,
+                       which is the vLLM default generally
+        Returns:
+            shape = [num_tokens, num_heads * head_size]
+        """
+
+        # Check that appropriate attention metadata attributes are
+        # selected for the desired attention type
+        if (attn_type == AttentionType.ENCODER
+                and (not attn_metadata.is_all_encoder_attn_metadata_set)):
+            raise AttributeError("Encoder attention requires setting "
+                                 "encoder metadata attributes.")
+
+        elif (attn_type == AttentionType.ENCODER_DECODER
+              and (not attn_metadata.is_all_cross_attn_metadata_set)):
+            raise AttributeError("Encoder/decoder cross-attention "
+                                 "requires setting cross-attention "
+                                 "metadata attributes.")
+
+        query = query.view(-1, self.num_heads, self.head_size)
+        if key is not None:
+            assert value is not None
+            key = key.view(-1, self.num_kv_heads, self.head_size)
+            value = value.view(-1, self.num_kv_heads, self.head_size)
+        else:
+            assert value is None
+
+        # Self-attention vs. cross-attention will impact
+        # which KV cache memory-mapping & which
+        # seqlen datastructures we utilize
+
+        if (attn_type != AttentionType.ENCODER and kv_cache.numel() > 0):
+            # KV-cache during decoder-self- or
+            # encoder-decoder-cross-attention, but not
+            # during encoder attention.
+            #
+            # Even if there are no new key/value pairs to cache,
+            # we still need to break out key_cache and value_cache
+            # i.e. for later use by paged attention
+            key_cache, value_cache = PagedAttention.split_kv_cache(
+                kv_cache, self.num_kv_heads, self.head_size)
+
+            if (key is not None) and (value is not None):
+
+                if attn_type == AttentionType.ENCODER_DECODER:
+                    # Update cross-attention KV cache (prefill-only)
+                    # During cross-attention decode, key & value will be None,
+                    # preventing this IF-statement branch from running
+                    updated_slot_mapping = attn_metadata.cross_slot_mapping
+                else:
+                    # Update self-attention KV cache (prefill/decode)
+                    updated_slot_mapping = attn_metadata.slot_mapping
+
+                # Reshape the input keys and values and store them in the cache.
+                # If kv_cache is not provided, the new key and value tensors are
+                # not cached. This happens during the initial memory
+                # profiling run.
+                PagedAttention.write_to_paged_cache(key, value, key_cache,
+                                                    value_cache,
+                                                    updated_slot_mapping,
+                                                    self.kv_cache_dtype,
+                                                    k_scale, v_scale)
+        (num_prefill_query_tokens, num_prefill_kv_tokens,
+        num_decode_query_tokens) = \
+            get_num_prefill_decode_query_kv_tokens(attn_metadata, attn_type)
+
+        output = torch.empty_like(query)
+        # Query for decode. KV is not needed because it is already cached.
+        decode_query = query[num_prefill_query_tokens:]
+        # QKV for prefill.
+        query = query[:num_prefill_query_tokens]
+        if key is not None and value is not None:
+            key = key[:num_prefill_kv_tokens]
+            value = value[:num_prefill_kv_tokens]
+
+        assert query.shape[0] == num_prefill_query_tokens
+        assert decode_query.shape[0] == num_decode_query_tokens
+
+        if prefill_meta := attn_metadata.prefill_metadata:
+            # Prompt run.
+            if kv_cache.numel() == 0 or prefill_meta.block_tables.numel() == 0:
+                # normal attention.
+                # block tables are empty if the prompt does not have a cached
+                # prefix.
+                out = self._run_memory_efficient_xformers_forward(
+                    query, key, value, prefill_meta, attn_type=attn_type)
+                assert out.shape == output[:num_prefill_query_tokens].shape
+                output[:num_prefill_query_tokens] = out
+            else:
+                assert attn_type != AttentionType.ENCODER_ONLY, (
+                    "Encoder-only models should not have prefix attention.")
+
+                assert prefill_meta.query_start_loc is not None
+                assert prefill_meta.max_query_len is not None
+
+                # prefix-enabled attention
+                # TODO(Hai) this triton kernel has regression issue (broke) to
+                # deal with different data types between KV and FP8 KV cache,
+                # to be addressed separately.
+                out = PagedAttention.forward_prefix(
+                    query,
+                    key,
+                    value,
+                    self.kv_cache_dtype,
+                    key_cache,
+                    value_cache,
+                    prefill_meta.block_tables,
+                    prefill_meta.query_start_loc,
+                    prefill_meta.seq_lens_tensor,
+                    prefill_meta.context_lens_tensor,
+                    prefill_meta.max_query_len,
+                    self.alibi_slopes,
+                    self.sliding_window,
+                    k_scale,
+                    v_scale,
+                )
+                assert output[:num_prefill_query_tokens].shape == out.shape
+                output[:num_prefill_query_tokens] = out
+
+        if decode_meta := attn_metadata.decode_metadata:
+            assert attn_type != AttentionType.ENCODER_ONLY, (
+                "Encoder-only models should not have decode metadata.")
+
+            (
+                seq_lens_arg,
+                max_seq_len_arg,
+                block_tables_arg,
+            ) = get_seq_len_block_table_args(decode_meta, False, attn_type)
+
+            output[num_prefill_query_tokens:] = PagedAttention.forward_decode(
+                decode_query,
+                key_cache,
+                value_cache,
+                block_tables_arg,
+                seq_lens_arg,
+                max_seq_len_arg,
+                self.kv_cache_dtype,
+                self.num_kv_heads,
+                self.scale,
+                self.alibi_slopes,
+                k_scale,
+                v_scale,
+            )
+
+        # Reshape the output tensor.
+        return output.view(-1, self.num_heads * self.head_size)
+
+    def _run_memory_efficient_xformers_forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        attn_metadata: XFormersMetadata,
+        attn_type: AttentionType = AttentionType.DECODER,
+    ) -> torch.Tensor:
+        """Attention for 1D query of multiple prompts. Multiple prompt
+        tokens are flattened in to `query` input.
+
+        See https://facebookresearch.github.io/xformers/components/ops.html
+        for API spec.
+
+        Args:
+            output: shape = [num_prefill_tokens, num_heads, head_size]
+            query: shape = [num_prefill_tokens, num_heads, head_size]
+            key: shape = [num_prefill_tokens, num_kv_heads, head_size]
+            value: shape = [num_prefill_tokens, num_kv_heads, head_size]
+            attn_metadata: Metadata for attention.
+            attn_type: Select attention type, between encoder attention,
+                       decoder self-attention, or encoder/decoder cross-
+                       attention. Defaults to decoder self-attention,
+                       which is the vLLM default generally
+        """
+
+        original_query = query
+        if self.num_kv_heads != self.num_heads:
+            # GQA/MQA requires the shape [B, M, G, H, K].
+            # Note that the output also has the same shape (which is different
+            # from a spec from the doc).
+            query = query.view(query.shape[0], self.num_kv_heads,
+                               self.num_queries_per_kv, query.shape[-1])
+            key = key[:, :,
+                      None, :].expand(key.shape[0], self.num_kv_heads,
+                                      self.num_queries_per_kv, key.shape[-1])
+            value = value[:, :,
+                          None, :].expand(value.shape[0], self.num_kv_heads,
+                                          self.num_queries_per_kv,
+                                          value.shape[-1])
+
+        # Set attention bias if not provided. This typically happens at
+        # the very attention layer of every iteration.
+        # FIXME(woosuk): This is a hack.
+        attn_bias = _get_attn_bias(attn_metadata, attn_type)
+        if attn_bias is None:
+            if self.alibi_slopes is None:
+
+                # Cross attention block of decoder branch of encoder-decoder
+                # model uses seq_lens for dec / encoder_seq_lens for enc
+                if (attn_type == AttentionType.ENCODER_DECODER):
+                    assert attn_metadata.seq_lens is not None
+                    assert attn_metadata.encoder_seq_lens is not None
+
+                    # Cross-attention mask is non-causal
+                    attn_bias = BlockDiagonalMask.from_seqlens(
+                        attn_metadata.seq_lens, attn_metadata.encoder_seq_lens)
+
+                # Encoder branch of encoder-decoder model uses
+                # attn_metadata.encoder_seq_lens
+                elif attn_type == AttentionType.ENCODER:
+
+                    assert attn_metadata.encoder_seq_lens is not None
+
+                    # Encoder self-attention mask is non-causal
+                    attn_bias = BlockDiagonalMask.from_seqlens(
+                        attn_metadata.encoder_seq_lens)
+
+                # Self-attention block of encoder-only model just
+                # uses the seq_lens directly.
+                elif attn_type == AttentionType.ENCODER_ONLY:
+                    assert attn_metadata.seq_lens is not None
+
+                    # Encoder self-attention mask is non-causal
+                    attn_bias = BlockDiagonalMask.from_seqlens(
+                        attn_metadata.seq_lens)
+
+                # Self-attention block of decoder branch just
+                # uses the seq_lens directly
+                elif attn_type == AttentionType.DECODER:
+                    assert attn_metadata.seq_lens is not None
+
+                    # Decoder self-attention mask is causal
+                    attn_bias = BlockDiagonalCausalMask.from_seqlens(
+                        attn_metadata.seq_lens)
+                else:
+                    raise ValueError("Unknown AttentionType: %s", attn_type)
+
+                if self.sliding_window is not None:
+                    attn_bias = attn_bias.make_local_attention(
+                        self.sliding_window)
+                attn_bias = [attn_bias]
+            else:
+                assert attn_type == AttentionType.DECODER
+                assert attn_metadata.seq_lens is not None
+                attn_bias = _make_alibi_bias(self.alibi_slopes,
+                                             self.num_kv_heads, query.dtype,
+                                             attn_metadata.seq_lens)
+
+            _set_attn_bias(attn_metadata, attn_bias, attn_type)
+
+        # No alibi slopes.
+        # TODO(woosuk): Too many view operations. Let's try to reduce
+        # them in the future for code readability.
+        if self.alibi_slopes is None:
+            # Add the batch dimension.
+            query = query.unsqueeze(0)
+            key = key.unsqueeze(0)
+            value = value.unsqueeze(0)
+            out = xops.memory_efficient_attention_forward(
+                query,
+                key,
+                value,
+                attn_bias=attn_bias[0],
+                p=0.0,
+                scale=self.scale)
+            return out.view_as(original_query)
+
+        # Attention with alibi slopes.
+        # FIXME(woosuk): Because xformers does not support dynamic sequence
+        # lengths with custom attention bias, we process each prompt one by
+        # one. This is inefficient, especially when we have many short prompts.
+        assert attn_metadata.seq_lens is not None
+        output = torch.empty_like(original_query)
+        start = 0
+        for i, seq_len in enumerate(attn_metadata.seq_lens):
+            end = start + seq_len
+            out = xops.memory_efficient_attention_forward(
+                query[None, start:end],
+                key[None, start:end],
+                value[None, start:end],
+                attn_bias=attn_bias[i],
+                p=0.0,
+                scale=self.scale)
+            # TODO(woosuk): Unnecessary copy. Optimize.
+            output[start:end].copy_(out.view_as(original_query[start:end]))
+            start += seq_len
+        return output
+
+
+def _make_alibi_bias(
+    alibi_slopes: torch.Tensor,
+    num_kv_heads: int,
+    dtype: torch.dtype,
+    seq_lens: List[int],
+) -> List[AttentionBias]:
+    attn_biases: List[AttentionBias] = []
+    for seq_len in seq_lens:
+        bias = torch.arange(seq_len, dtype=dtype)
+        # NOTE(zhuohan): HF uses
+        #     `bias = bias[None, :].repeat(seq_len, 1)`
+        # here. We find that both biases give the same results, but
+        # the bias below more accurately follows the original ALiBi
+        # paper.
+        # Calculate a matrix where each element represents ith element- jth
+        # element.
+        bias = bias[None, :] - bias[:, None]
+
+        padded_len = (seq_len + 7) // 8 * 8
+        num_heads = alibi_slopes.shape[0]
+        bias = torch.empty(
+            1,  # batch size
+            num_heads,
+            seq_len,
+            padded_len,
+            device=alibi_slopes.device,
+            dtype=dtype,
+        )[:, :, :, :seq_len].copy_(bias)
+        bias.mul_(alibi_slopes[:, None, None])
+        if num_heads != num_kv_heads:
+            bias = bias.unflatten(1, (num_kv_heads, num_heads // num_kv_heads))
+        attn_biases.append(LowerTriangularMaskWithTensorBias(bias))
+
+    return attn_biases

vllm-v0.6.2/vllm/attention/layer.py

@@ -0,0 +1,114 @@
+"""Attention layer."""
+from typing import Any, Dict, List, Optional
+
+import torch
+import torch.nn as nn
+
+from vllm.attention import AttentionMetadata, AttentionType
+from vllm.attention.selector import get_attn_backend
+from vllm.config import CacheConfig
+from vllm.model_executor.layers.quantization.base_config import (
+    QuantizationConfig)
+from vllm.model_executor.layers.quantization.kv_cache import BaseKVCacheMethod
+
+
+class Attention(nn.Module):
+    """Attention layer.
+
+    This class takes query, key, and value tensors as input. The input tensors
+    can either contain prompt tokens or generation tokens.
+    The class does the following:
+
+    1. Store the input key and value tensors in the KV cache.
+    2. Perform (multi-head/multi-query/grouped-query) attention.
+    3. Return the output tensor.
+    """
+
+    def __init__(
+        self,
+        num_heads: int,
+        head_size: int,
+        scale: float,
+        num_kv_heads: Optional[int] = None,
+        alibi_slopes: Optional[List[float]] = None,
+        cache_config: Optional[CacheConfig] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+        blocksparse_params: Optional[Dict[str, Any]] = None,
+        logits_soft_cap: Optional[float] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        if cache_config is not None:
+            kv_cache_dtype = cache_config.cache_dtype
+            block_size = cache_config.block_size
+            sliding_window = cache_config.sliding_window
+            is_attention_free = cache_config.is_attention_free
+        else:
+            kv_cache_dtype = "auto"
+            block_size = 16
+            sliding_window = None
+            is_attention_free = False
+        if num_kv_heads is None:
+            num_kv_heads = num_heads
+
+        # The default k/v_scale is set to 1.0. This is ignored
+        # when kv-cache is not fp8, and should be used with
+        # kv-cache in fp8_e5m2. For kv-cache in fp8_e4m3, we
+        # expect the pre-quantized k/v_scale to be loaded along
+        # with the model weights.
+        self.kv_cache_dtype = kv_cache_dtype
+        self._k_scale = 1.0
+        self._v_scale = 1.0
+        quant_method = quant_config.get_quant_method(
+            self, prefix=prefix) if quant_config else None
+        if quant_method is not None:
+            assert isinstance(quant_method, BaseKVCacheMethod)
+            # TODO (mgoin): kv cache dtype should be specified in the FP8
+            # checkpoint config and become the "auto" behavior
+            if self.kv_cache_dtype == "fp8_e5m2":
+                raise ValueError("fp8_e5m2 kv-cache is not supported with "
+                                 "fp8 checkpoints.")
+            # If quantization is enabled, we make "k_scale" and "v_scale"
+            # parameters so that it can be loaded from the model checkpoint.
+            # The k/v_scale will then be converted back to native float32
+            # values after weight loading.
+            self.quant_method = quant_method
+            self.quant_method.create_weights(self)
+
+        # During model initialization, the default dtype is set as the model
+        # weight and activation dtype.
+        dtype = torch.get_default_dtype()
+        attn_backend = get_attn_backend(head_size, dtype, kv_cache_dtype,
+                                        block_size, is_attention_free,
+                                        blocksparse_params is not None)
+        impl_cls = attn_backend.get_impl_cls()
+        self.impl = impl_cls(num_heads, head_size, scale, num_kv_heads,
+                             alibi_slopes, sliding_window, kv_cache_dtype,
+                             blocksparse_params, logits_soft_cap)
+
+    def forward(
+        self,
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: AttentionMetadata,
+        attn_type: AttentionType = AttentionType.DECODER,
+    ) -> torch.Tensor:
+
+        return self.impl.forward(query,
+                                 key,
+                                 value,
+                                 kv_cache,
+                                 attn_metadata,
+                                 self._k_scale,
+                                 self._v_scale,
+                                 attn_type=attn_type)
+
+    def extra_repr(self) -> str:
+        s = f"head_size={self.impl.head_size}"  # type: ignore
+        s += f", num_heads={self.impl.num_heads}"  # type: ignore
+        s += f", num_kv_heads={self.impl.num_kv_heads}"  # type: ignore
+        s += f", scale={self.impl.scale}"  # type: ignore
+        s += f", backend={self.impl.__class__.__name__}"
+        return s

vllm-v0.6.2/vllm/attention/ops/blocksparse_attention/blocksparse_attention_kernel.py

@@ -0,0 +1,423 @@
+import torch
+import triton
+import triton.language as tl
+
+
+def blocksparse_flash_attn_varlen_fwd(
+        q,
+        k,
+        v,  # (#tokens, n_heads, head_size)
+        cu_seqlens_k,
+        cu_seqlens_q,
+        sm_scale,
+        sparse_layout,
+        *,
+        block_size=64,
+        q_block_size=None,
+        max_seqlen=None):
+    # split q to blocks
+
+    assert isinstance(sparse_layout, (list, tuple))
+
+    _, n_heads, head_size = q.shape
+    batch_size = cu_seqlens_k.size(0) - 1
+    q_block_size = q_block_size or block_size
+
+    assert q.dim() == k.dim() == v.dim() == 3
+    assert q.size(1) % k.size(1) == 0
+    assert q.size(2) == k.size(2)
+    # TODO(linxihui): allow k, v to have different head_size
+    assert k.shape == v.shape
+    assert cu_seqlens_k.dim() == 1
+
+    q_k_ratio = q.size(1) // k.size(1)
+
+    if cu_seqlens_q is None:
+        if q.size(0) == batch_size:  # decoding only
+            cu_seqlens_q = torch.arange(
+                0,
+                batch_size + 1,
+                dtype=cu_seqlens_k.dtype,
+                device=cu_seqlens_k.device,
+            )
+        elif q.size(0) == k.size(0):
+            cu_seqlens_q = cu_seqlens_k
+        else:
+            raise ValueError("cu_seqlens_q must be specified\
+                    if it mix of prefilling and decoding.")
+    else:
+        assert cu_seqlens_k.size(0) == cu_seqlens_q.size(0)
+
+    # switch to use cpu to avoid too many kernel launches when iterated over
+    q_lens = (cu_seqlens_q[1:] - cu_seqlens_q[:-1]).cpu()
+    k_lens = (cu_seqlens_k[1:] - cu_seqlens_k[:-1]).cpu()
+
+    assert torch.logical_or(q_lens == 1, k_lens == q_lens).all(), (
+        "length of q should either be 1 (decoding) or same as k (prefilling).")
+
+    if max_seqlen:
+        assert k_lens.max() <= max_seqlen
+
+    n_blocks = (q_lens + q_block_size - 1) // q_block_size
+
+    q_batch_ids = torch.tensor(
+        [i for i, n in enumerate(n_blocks) for _ in range(n)],
+        dtype=cu_seqlens_q.dtype,
+        device=cu_seqlens_q.device,
+    )
+    q_start_sids = torch.tensor(
+        [i * q_block_size for n in n_blocks for i in range(n)],
+        dtype=cu_seqlens_q.dtype,
+        device=cu_seqlens_q.device,
+    )
+
+    out = q.new_empty(q.shape)
+    cu_seqlens_q = cu_seqlens_q.contiguous()
+    cu_seqlens_k = cu_seqlens_k.contiguous()
+
+    layout_crow_indices, layout_col_indices = sparse_layout
+    block_d = triton.next_power_of_2(head_size)
+
+    decoding_only = (q_lens == 1).all().item()
+    grid = (len(q_start_sids), n_heads, 1)
+
+    _fwd_kernel_batch_inference[grid](
+        q,
+        k,
+        v,
+        out,
+        sm_scale,
+        cu_seqlens_q[:-1],
+        cu_seqlens_q[1:],
+        cu_seqlens_k[:-1],
+        cu_seqlens_k[1:],
+        q_batch_ids,
+        q_start_sids,
+        0,
+        *q.stride(),
+        0,
+        *k.stride(),
+        0,
+        *v.stride(),
+        0,
+        *out.stride(),
+        layout_crow_indices,
+        layout_col_indices,
+        *layout_crow_indices.stride(),
+        *layout_col_indices.stride(),
+        q_k_ratio,
+        HAS_BATCH_DIM=False,
+        D_HEAD=head_size,
+        BLOCK_M=q_block_size,
+        BLOCK_N=block_size,
+        BLOCK_D=block_d,
+        BLOCK_M_LOADING=(16 if decoding_only else
+                         q_block_size),  # smaller for decoding
+        EVEN_D=block_d == head_size,
+        num_warps=1 if decoding_only else 4,
+        num_stages=3)
+
+    return out
+
+
+@triton.jit
+def _fwd_kernel_inner(
+    acc,
+    l_i,
+    m_i,
+    q,
+    Q,
+    k_block_col_idx,
+    layout_col_ptr,
+    layout_col_stride_h,
+    layout_col_stride_m,
+    k_ptrs,
+    v_ptrs,
+    off_h,
+    offs_m,
+    offs_n,
+    offs_d,
+    stride_kt,
+    stride_vt,
+    sm_scale,
+    k_seqlen,
+    past_len,
+    LAST_K_BLOCK: tl.constexpr,
+    BLOCK_M_LOADING: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    D_HEAD: tl.constexpr,
+    EVEN_D: tl.constexpr,
+    M_LT_N: tl.constexpr,
+):
+    k_block_id = tl.load(layout_col_ptr + off_h * layout_col_stride_h +
+                         k_block_col_idx * layout_col_stride_m).to(tl.int32)
+    start_n = k_block_id * BLOCK_N
+    if LAST_K_BLOCK:
+        if EVEN_D:
+            k = tl.load(
+                k_ptrs + start_n * stride_kt,
+                mask=offs_n[None, :] + start_n < k_seqlen,
+            )
+        else:
+            k = tl.load(
+                k_ptrs + start_n * stride_kt,
+                mask=(offs_n[None, :] + start_n < k_seqlen) &
+                (offs_d[:, None] < D_HEAD),
+            )
+    else:
+        if EVEN_D:
+            k = tl.load(k_ptrs + start_n * stride_kt)
+        else:
+            k = tl.load(k_ptrs + start_n * stride_kt,
+                        mask=offs_d[:, None] < D_HEAD)
+
+    qk = tl.zeros([BLOCK_M_LOADING, BLOCK_N], dtype=tl.float32)
+    qk += tl.dot(q, k)
+    qk *= sm_scale
+
+    # the following is needed only when LAST_K_BLOCK or BLOCK_M < BLOCK_N
+    if LAST_K_BLOCK | M_LT_N:
+        qk += tl.where(
+            offs_m[:, None] + past_len >= (start_n + offs_n[None, :]),
+            0,
+            float("-inf"),
+        )
+
+    # flash-attn2
+    m_ij = tl.maximum(m_i, tl.max(qk, 1))
+    p = tl.math.exp2(qk - m_ij[:, None])
+    l_ij = tl.sum(p, 1)
+    alpha = tl.math.exp2(m_i - m_ij)
+    acc = acc * alpha[:, None]
+    # update m_i
+    m_i = m_ij
+    l_i = l_i * alpha + l_ij
+
+    p = p.to(Q.dtype.element_ty)
+    # update acc
+    if LAST_K_BLOCK:
+        if EVEN_D:
+            v = tl.load(
+                v_ptrs + start_n * stride_vt,
+                mask=offs_n[:, None] + start_n < k_seqlen,
+            )
+        else:
+            v = tl.load(
+                v_ptrs + start_n * stride_vt,
+                mask=(offs_n[:, None] + start_n < k_seqlen) &
+                (offs_d[None, :] < D_HEAD),
+            )
+    else:
+        if EVEN_D:
+            v = tl.load(v_ptrs + start_n * stride_vt)
+        else:
+            v = tl.load(v_ptrs + start_n * stride_vt,
+                        mask=offs_d[None, :] < D_HEAD)
+
+    acc += tl.dot(p, v)
+
+    return acc, l_i, m_i
+
+
+@triton.heuristics({
+    "M_LT_N":
+    lambda kwargs: kwargs["BLOCK_M"] < kwargs["BLOCK_N"],
+})
+@triton.jit
+def _fwd_kernel_batch_inference(
+    Q,
+    K,
+    V,
+    Out,
+    sm_scale,
+    q_batch_starts,
+    q_batch_ends,
+    k_batch_starts,
+    k_batch_ends,
+    q_batch_ids,
+    q_start_sids,
+    stride_qb,
+    stride_qt,
+    stride_qh,
+    stride_qd,
+    stride_kb,
+    stride_kt,
+    stride_kh,
+    stride_kd,
+    stride_vb,
+    stride_vt,
+    stride_vh,
+    stride_vd,
+    stride_ob,
+    stride_ot,
+    stride_oh,
+    stride_od,
+    layout_crow_ptr,
+    layout_col_ptr,
+    layout_crow_stride_h,
+    layout_crow_stride_m,
+    layout_col_stride_h,
+    layout_col_stride_m,
+    q_k_ratio,
+    HAS_BATCH_DIM: tl.constexpr,
+    D_HEAD: tl.constexpr,
+    BLOCK_M: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    BLOCK_D: tl.constexpr,
+    BLOCK_M_LOADING: tl.constexpr,
+    EVEN_D: tl.constexpr,
+    M_LT_N: tl.constexpr,
+):
+    """
+    NOTATION:
+    pid: position id
+    sid: storage id
+    sbid: storage block id
+    pbid: position block id
+    offs_m, offs_n: storage offsets of m-dim(q, row) and n-dim(k, col)
+
+    TODO(linxihui):
+    Optimize grouped-attn
+    """
+    off_zm = tl.program_id(0)
+    off_h = tl.program_id(1)
+
+    off_h_for_kv = off_h // q_k_ratio
+
+    if HAS_BATCH_DIM:
+        off_z = tl.program_id(2)
+        Q += off_z * stride_qb
+        K += off_z * stride_kb
+        V += off_z * stride_vb
+        Out += off_z * stride_ob
+        start_m = off_zm
+        q_start_sid = start_m * BLOCK_M  # always 0 for decoding
+    else:
+        off_z = tl.load(q_batch_ids + off_zm).to(tl.int32)  # [0, 0, 0, 1]
+        q_start_sid = tl.load(q_start_sids + off_zm)
+        start_m = q_start_sid // BLOCK_M  # q_sbid
+
+    offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M_LOADING)
+    offs_n = tl.arange(0, BLOCK_N)
+    offs_d = tl.arange(0, BLOCK_D)
+
+    q_cu_start = tl.load(q_batch_starts + off_z).to(tl.int32)
+    q_seqlen = tl.load(q_batch_ends + off_z).to(tl.int32) - q_cu_start
+    k_cu_start = tl.load(k_batch_starts + off_z).to(tl.int32)
+    k_seqlen = tl.load(k_batch_ends + off_z).to(tl.int32) - k_cu_start
+    past_len = k_seqlen - q_seqlen
+
+    Q += q_cu_start * stride_qt + off_h * stride_qh
+    K += k_cu_start * stride_kt + off_h_for_kv * stride_kh
+    V += k_cu_start * stride_vt + off_h_for_kv * stride_vh
+    Out += q_cu_start * stride_ot + off_h * stride_oh
+
+    q_pbid = (past_len + q_start_sid) // BLOCK_M
+
+    if EVEN_D:
+        q = tl.load(
+            Q + offs_m[:, None] * stride_qt + offs_d[None, :] * stride_qd,
+            mask=offs_m[:, None] < q_seqlen,
+        )
+    else:
+        q = tl.load(
+            Q + offs_m[:, None] * stride_qt + offs_d[None, :] * stride_qd,
+            mask=(offs_m[:, None] < q_seqlen) & (offs_d[None, :] < D_HEAD),
+            other=0,
+        )
+
+    sparse_crow_ptr = (layout_crow_ptr + off_h * layout_crow_stride_h +
+                       q_pbid * layout_crow_stride_m)
+
+    # TODO(linxihui): load at once, with any Triton version
+    # that supports `tl.split`, e.g., Triton 3.0
+    k_block_start = tl.load(sparse_crow_ptr).to(tl.int32)
+    k_block_end = tl.load(sparse_crow_ptr + 1).to(tl.int32)
+
+    m_i = tl.zeros([BLOCK_M_LOADING], dtype=tl.float32) - float("inf")
+    l_i = tl.zeros([BLOCK_M_LOADING], dtype=tl.float32)
+    acc = tl.zeros([BLOCK_M_LOADING, BLOCK_D], dtype=tl.float32)
+
+    k_ptrs = K + offs_n[None, :] * stride_kt + offs_d[:, None] * stride_kd
+    v_ptrs = V + offs_n[:, None] * stride_vt + offs_d[None, :] * stride_vd
+
+    sm_scale *= (
+        1.44269504  # 1/log2 as we use base2 for exponential and logarithm
+    )
+
+    for k_block_col_idx in range(k_block_start, k_block_end - 1):
+        acc, l_i, m_i = _fwd_kernel_inner(
+            acc,
+            l_i,
+            m_i,
+            q,
+            Q,
+            k_block_col_idx,
+            layout_col_ptr,
+            layout_col_stride_h,
+            layout_col_stride_m,
+            k_ptrs,
+            v_ptrs,
+            off_h,
+            offs_m,
+            offs_n,
+            offs_d,
+            stride_kt,
+            stride_vt,
+            sm_scale,
+            k_seqlen,
+            past_len,
+            False,
+            BLOCK_M_LOADING,
+            BLOCK_N,
+            D_HEAD,
+            EVEN_D,
+            M_LT_N,
+        )
+
+    acc, l_i, m_i = _fwd_kernel_inner(
+        acc,
+        l_i,
+        m_i,
+        q,
+        Q,
+        k_block_end - 1,
+        layout_col_ptr,
+        layout_col_stride_h,
+        layout_col_stride_m,
+        k_ptrs,
+        v_ptrs,
+        off_h,
+        offs_m,
+        offs_n,
+        offs_d,
+        stride_kt,
+        stride_vt,
+        sm_scale,
+        k_seqlen,
+        past_len,
+        True,
+        BLOCK_M_LOADING,
+        BLOCK_N,
+        D_HEAD,
+        EVEN_D,
+        M_LT_N,
+    )
+
+    # flash-attn 2
+    m_i += tl.math.log2(l_i)
+    acc = acc / l_i[:, None]
+
+    # write output
+    if EVEN_D:
+        tl.store(
+            Out + offs_m[:, None] * stride_ot + offs_d[None, :] * stride_od,
+            acc,
+            mask=offs_m[:, None] < q_seqlen,
+        )
+    else:
+        tl.store(
+            Out + offs_m[:, None] * stride_ot + offs_d[None, :] * stride_od,
+            acc,
+            mask=(offs_m[:, None] < q_seqlen) & (offs_d[None, :] < D_HEAD),
+        )

vllm-v0.6.2/vllm/attention/ops/blocksparse_attention/interface.py

@@ -0,0 +1,236 @@
+import math
+
+import torch
+
+from vllm.platforms import current_platform
+
+from .utils import (dense_to_crow_col, get_head_sliding_step,
+                    get_sparse_attn_mask)
+
+IS_COMPUTE_8_OR_ABOVE = current_platform.has_device_capability(80)
+
+if IS_COMPUTE_8_OR_ABOVE:
+    from .blocksparse_attention_kernel import blocksparse_flash_attn_varlen_fwd
+
+
+class LocalStridedBlockSparseAttn(torch.nn.Module):
+
+    def __init__(
+        self,
+        n_heads,
+        max_seqlen,
+        local_blocks,
+        vert_stride,
+        block_size,
+        device=None,
+        dtype=None,
+        homo_head=False,
+        active_head_range=None,
+        q_block_size=None,
+        use_spda=None,
+    ):
+        super().__init__()
+        if use_spda is None:
+            use_spda = current_platform.is_rocm() or \
+                        current_platform.is_cpu() or not \
+                        IS_COMPUTE_8_OR_ABOVE
+        device = device or (torch.cuda.current_device()
+                            if current_platform.is_cuda_alike() else "cpu")
+        device = torch.device(device)
+        # NOTE: vllm CPU backend support BF16 instead of FP16.
+        dtype = dtype or (torch.bfloat16 if IS_COMPUTE_8_OR_ABOVE
+                          or device.type == "cpu" else torch.half)
+
+        self.n_heads = n_heads
+        self.max_seqlen = max_seqlen
+        self.local_blocks = local_blocks
+        self.vert_stride = vert_stride
+        self.use_spda = use_spda
+        self.dtype = dtype
+        self.device = device
+        self.block_size = block_size
+        self.q_block_size = q_block_size
+        self.homo_head = homo_head
+        self.active_head_range = active_head_range
+        self.head_sliding_step = get_head_sliding_step(n_heads, vert_stride,
+                                                       homo_head)
+
+        sparse_layout, sparse_pattern, self.dense_attn_mask = (
+            self.get_attn_pattern(dtype, device))
+
+        if q_block_size is not None and q_block_size != block_size:
+            if q_block_size > block_size:
+                assert q_block_size % block_size == 0
+                blocks_to_merge = q_block_size // block_size
+                shape = sparse_pattern.shape
+                sparse_pattern = sparse_pattern.view(shape[0], -1,
+                                                     blocks_to_merge,
+                                                     shape[-1])
+                sparse_pattern = sparse_pattern.sum(2)
+                sparse_layout = dense_to_crow_col(sparse_pattern)
+            else:
+                raise ValueError(
+                    "Does not support smaller q_block_size. It will be slower."
+                )
+
+        self.sparse_layout = sparse_layout
+
+    def get_attn_pattern(self, dtype, device):
+        sparse_layout, sparse_pattern, dense_attn_mask = get_sparse_attn_mask(
+            self.n_heads,
+            self.max_seqlen,
+            self.max_seqlen,
+            dtype,
+            device,
+            block_size=self.block_size,
+            local_blocks=self.local_blocks,
+            vert_stride=self.vert_stride,
+            homo_head=self.homo_head,
+            return_dense=self.use_spda,
+            dense_mask_type="bias",
+        )
+        if (not self.homo_head) and (self.active_head_range is not None):
+            assert isinstance(self.active_head_range, tuple)
+            assert (len(self.active_head_range) == 2)
+            h_start, h_end = self.active_head_range
+            sparse_layout = tuple(x[h_start:h_end] for x in sparse_layout)
+            if self.use_spda:
+                dense_attn_mask = dense_attn_mask[h_start:h_end]
+        return sparse_layout, sparse_pattern, dense_attn_mask
+
+    def varlen_attn(self,
+                    q,
+                    k,
+                    v,
+                    cu_seqlens_k,
+                    cu_seqlens_q=None,
+                    sm_scale=None):
+        """
+        q, k, v: shape = (num_tokens, num_heads_q/kv, head_size).
+        Support grouped attention, with `q[:, i*r:(i*r + r)]`
+        is correspondent to `k[:, i]`, where `r` is the q/k ratio.
+        cu_seqlens_k: shape=(batch_size + 1,),
+        indicating segment of samples,
+        e.g., `k[cu_seqlen[i]:cu_seqlne[i+1]]` is q of sample i
+        cu_seqlens_q: shape=(batch_size + 1, ).
+        Default None: same as cu_seqlens_k for prefilling or
+        [0, 1, .., batch_size] for decoding.
+        The only case you need to specify is when q is a mix of
+        prefilling and decoding.
+        sm_scale: softmax scale, default to 1/sqrt(head_size).
+
+        return: tensor of shape as q.
+        """
+        assert (
+            IS_COMPUTE_8_OR_ABOVE
+        ), "Requires compute capability of 8 or above (Ampere or newer) to use \
+            Triton kernel."
+
+        sm_scale = sm_scale or 1.0 / math.sqrt(q.size(-1))
+
+        return blocksparse_flash_attn_varlen_fwd(
+            q,
+            k,
+            v,
+            cu_seqlens_k,
+            cu_seqlens_q,
+            sm_scale,
+            self.sparse_layout,
+            block_size=self.block_size,
+            q_block_size=self.q_block_size,
+            max_seqlen=self.max_seqlen,
+        )
+
+    @staticmethod
+    def transpose_and_pad(x, cu_seqlens, maxlen, head_repeats=1):
+        """
+        :param x: (total_tokens, n_heads, head_size)
+        :return: (batch, n_heads, length, head_size)
+        """
+        x_padded = x.new_empty(
+            len(cu_seqlens) - 1, x.size(1), head_repeats, maxlen, x.size(2))
+        cu_seqlens = cu_seqlens.cpu()
+        for i, (s, e) in enumerate(zip(cu_seqlens[:-1], cu_seqlens[1:])):
+            x_padded[i, :, :, :e - s].copy_(x[s:e].transpose(0,
+                                                             1).unsqueeze(1))
+        return x_padded.flatten(1, 2)
+
+    @staticmethod
+    def transpose_and_unpad(x_padded, cu_seqlens):
+        """
+        :param x_padded: (batch, n_heads, length, head_size)
+        :return: (total_tokens, n_heads, head_size)
+        """
+        cu_seqlens = cu_seqlens.cpu()
+        total_n_tokens = cu_seqlens[-1]
+        x = x_padded.new_empty(total_n_tokens, x_padded.size(1),
+                               x_padded.size(3))
+        for i, (s, e) in enumerate(zip(cu_seqlens[:-1], cu_seqlens[1:])):
+            x[s:e].copy_(x_padded[i, :, :e - s].transpose(0, 1))
+        return x
+
+    def spda(self, q, k, v, cu_seqlens_k, cu_seqlens_q=None, sm_scale=None):
+        """For CPU, V100 or other older GPUs.
+        NOTE: torch SPDA supports nested tensor,
+        but seems extremely slow. Choose to pad instead.
+        """
+        assert (cu_seqlens_q is None or
+                (cu_seqlens_q
+                 == cu_seqlens_k).all()), "Can only handle prompt with SPDA."
+        assert q.size(0) == k.size(0), "can only handle prompt with SPDA."
+
+        assert q.size(1) % k.size(1) == 0
+        q_k_ratio = q.size(1) // k.size(1)
+        sm_scale = sm_scale or 1.0 / math.sqrt(q.size(-1))
+        cu_seqlens = cu_seqlens_k.cpu()
+        maxlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
+
+        if (self.dense_attn_mask.dtype != q.dtype
+                or self.dense_attn_mask.device != q.device):
+            _, _, self.dense_attn_mask = self.get_attn_pattern(
+                q.dtype, q.device)
+        attn_mask = self.dense_attn_mask[None, :, :maxlen, :maxlen]
+
+        q2 = self.transpose_and_pad(q, cu_seqlens, maxlen, 1)
+        k2, v2 = (self.transpose_and_pad(x, cu_seqlens, maxlen, q_k_ratio)
+                  for x in [k, v])
+        spda_output = torch.nn.functional.scaled_dot_product_attention(
+            q2, k2, v2, attn_mask=attn_mask, scale=sm_scale)
+        return self.transpose_and_unpad(spda_output, cu_seqlens)
+
+    def forward(self, q, k, v, cu_seqlens_k, cu_seqlens_q=None, sm_scale=None):
+        """Dispatch to `varlen_attn` (Ampere or newer) or
+        `self.spda`(cpu, Volta, Turing or older)based on
+        the type of device used and cuda compute capability.
+
+        q, k, v: shape = (num_tokens, num_heads_q/kv, head_size).
+                Support grouped attention, with `q[:, i*r:(i*r + r)]`
+                is correspondent to `k[:, i]`, where `r` is the q/k ratio.
+        cu_seqlens_k: shape=(batch_size + 1,), indicating segment of samples,
+                    e.g., `k[cu_seqlen[i]:cu_seqlne[i+1]]` is q of sample i
+        cu_seqlens_q: shape=(batch_size + 1, ).
+                    Default None: same as cu_seqlens_k for prefilling or
+                    [0, 1, .., batch_size] for decoding.
+                    The only case you need to specify
+                    is when q is a mix of prefilling
+                    and decoding.
+        sm_scale: softmax scale, default to 1/sqrt(head_size).
+
+        return: tensor of shape as q.
+        """
+        assert k.dim() == 3
+        if self.use_spda:
+            return self.spda(
+                q,
+                k,
+                v,
+                cu_seqlens_k,
+                cu_seqlens_q=cu_seqlens_q,
+                sm_scale=sm_scale,
+            )
+        return self.varlen_attn(q,
+                                k,
+                                v,
+                                cu_seqlens_k,
+                                cu_seqlens_q=cu_seqlens_q,
+                                sm_scale=sm_scale)

vllm-v0.6.2/vllm/attention/ops/blocksparse_attention/utils.py

@@ -0,0 +1,242 @@
+# Helper functions for 3D sparse pattern
+# These function are not optimized and very inefficient.
+# Avoid calling them too frequent or use a cache mechanism.
+
+from functools import lru_cache
+
+import numpy as np
+import torch
+import triton
+
+
+class csr_matrix:
+    """Simple implementation of CSR matrix conversion without scipy.
+    This replaced scipy.sparse.csr_matrix() previously used."""
+
+    def __init__(self, input_array):
+        if not isinstance(input_array, np.ndarray):
+            raise ValueError("Input must be a NumPy array")
+
+        self.shape = input_array.shape
+        rows, cols = self.shape
+        data = []
+        indices = []
+        indptr = [0]
+
+        for i in range(rows):
+            for j in range(cols):
+                if input_array[i, j]:
+                    data.append(input_array[i, j])
+                    indices.append(j)
+            indptr.append(len(indices))
+
+        self.data = np.array(data)
+        self.indices = np.array(indices)
+        self.indptr = np.array(indptr)
+
+
+def dense_to_crow_col(x: torch.Tensor):
+    """Turning a 2D/3D torch tensor (x) to CSR rows/cols indexing.
+    NOTE: col_indices padded -1
+    """
+    device = x.device
+    pad = -1
+    dim = x.dim()
+    assert x.dim() in (2, 3)
+    if x.dim() == 2:
+        x = x[None]
+    x = [csr_matrix(xi.bool().cpu().numpy()) for xi in x]
+    crows = torch.vstack([torch.from_numpy(xi.indptr) for xi in x])
+    cols = [torch.from_numpy(xi.indices) for xi in x]
+    max_cols = max(len(xi) for xi in cols)
+    cols = [
+        torch.cat([xi, pad + xi.new_zeros(max_cols - xi.shape[0])])
+        for xi in cols
+    ]
+    cols = torch.vstack(cols)
+    if dim == 2:
+        crows = crows[0]
+        cols = cols[0]
+    return crows.to(device), cols.to(device)
+
+
+def crow_col_to_dense(crows: torch.Tensor,
+                      cols: torch.Tensor,
+                      dtype: torch.dtype = torch.float16):
+    dim = crows.dim()
+    if dim == 1:
+        crows = crows[None]
+        cols = cols[None]
+    device = crows.device
+    crows, cols = crows.cpu(), cols.cpu()  # faster in cpu
+    shape = (crows.shape[0], crows.shape[1] - 1, cols.max() + 1)
+    x = torch.zeros(shape, dtype=dtype)
+    for i in range(shape[0]):
+        for j in range(shape[1]):
+            x[i, j, cols[i, crows[i, j]:crows[i, j + 1]]] = 1
+    if dim == 1:
+        x = x[0]
+    return x.to(device)
+
+
+def dense_to_ccol_row(x: torch.Tensor):
+    """Similar, but to CSC format"""
+    x = x.transpose(-2, -1)
+    return dense_to_crow_col(x)
+
+
+def ccol_row_to_dense(ccol: torch.Tensor,
+                      rows: torch.Tensor,
+                      dtype: torch.dtype = torch.float16):
+    return crow_col_to_dense(ccol, rows, dtype).permute(0, 2, 1).contiguous()
+
+
+def _get_sparse_attn_mask_homo_head(
+    q_len: int,
+    max_seqlen: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    block_size: int = 128,
+    local_blocks: int = 4,
+    vert_stride: int = 4,
+    return_dense: bool = False,
+):
+    """
+    :return: a tuple of 3:
+        - tuple of crow_indices, col_indices representation
+            of CSR format.
+        - block dense mask
+        - all token dense mask (be aware that it can be
+            OOM if it is too big) if `return_dense==True`,
+            otherwise, None
+    """
+    with torch.no_grad():
+        num_blocks = triton.cdiv(max_seqlen, block_size)
+        q_pos = torch.arange(num_blocks)[:, None]
+        k_pos = torch.arange(num_blocks)[None]
+        mask_vert_strided = (torch.arange(num_blocks) + 1) % vert_stride == 0
+        block_mask_dense = (((q_pos >= k_pos)
+                             & ((q_pos - k_pos < local_blocks)
+                                | mask_vert_strided)).to(device).to(dtype))
+        num_blocks_q = triton.cdiv(q_len, block_size)
+        block_mask_dense_output = (dense_to_crow_col(
+            block_mask_dense[-num_blocks_q:].contiguous()))
+    if return_dense:
+        mask_dense = torch.kron(
+            block_mask_dense,
+            block_mask_dense.new_ones((block_size, block_size)),
+        )
+        causal_mask = torch.tril(torch.ones(
+            max_seqlen, max_seqlen)).type_as(mask_dense)[-q_len:]
+        mask_dense = mask_dense[-q_len:, :max_seqlen] * causal_mask
+        return (
+            block_mask_dense_output,
+            block_mask_dense,
+            mask_dense,
+        )
+    else:
+        return (
+            block_mask_dense_output,
+            block_mask_dense,
+            None,
+        )
+
+
+def binary_mask_to_bias(mask_dense: torch.Tensor):
+    mask_dense = 1 - mask_dense
+    mask_dense.masked_fill_(mask_dense.bool(), -torch.inf)
+    return mask_dense
+
+
+def get_head_sliding_step(n_heads: int,
+                          vert_stride: int,
+                          homo_head: bool = False):
+    if homo_head:
+        return 0
+    return max(1, int(vert_stride / n_heads))
+
+
+@lru_cache
+def get_sparse_attn_mask(
+    n_heads: int,
+    q_len: int,
+    max_seqlen: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    block_size: int = 64,
+    local_blocks: int = 4,
+    vert_stride: int = 4,
+    homo_head: bool = True,
+    return_dense: bool = False,
+    dense_mask_type: str = "binary",
+):
+    """
+    :param dense_mask_type: "binary" (0 for skip token, 1 for others)
+        or "bias" (-inf for skip token, 0 or others)
+    :return: a tuple of 3:
+        - tuple of crow_indices, col_indices representation
+            of CSR format.
+        - block dense mask
+        - all token dense mask (be aware that it can be OOM if it
+            is too big) if `return_dense==True`, otherwise, None
+    """
+    assert dense_mask_type in ("binary", "bias")
+    if homo_head:
+        with torch.no_grad():
+            (crow, col), block_mask_dense, mask_dense = (
+                _get_sparse_attn_mask_homo_head(
+                    q_len,
+                    max_seqlen,
+                    dtype,
+                    device,
+                    block_size,
+                    local_blocks,
+                    vert_stride,
+                    return_dense,
+                ))
+            crow = crow[None].expand(n_heads, crow.shape[0])
+            col = col[None].expand(n_heads, col.shape[0])
+            if return_dense:
+                mask_dense = mask_dense[None].expand(n_heads,
+                                                     *mask_dense.shape)
+                if dense_mask_type == "bias":
+                    mask_dense = binary_mask_to_bias(mask_dense)
+            return (crow, col), block_mask_dense, mask_dense
+
+    with torch.no_grad():
+        num_blocks = triton.cdiv(max_seqlen, block_size)
+        q_pos = torch.arange(num_blocks)[None, :, None]
+        k_pos = torch.arange(num_blocks)[None, None]
+        head_sliding_step = get_head_sliding_step(n_heads, vert_stride)
+        mask_vert_strided = [
+            (torch.arange(num_blocks) + h * head_sliding_step + 1) %
+            vert_stride == 0 for h in range(n_heads)
+        ]
+        mask_vert_strided = torch.vstack(mask_vert_strided).unsqueeze(1)
+        block_mask_dense = (((q_pos >= k_pos)
+                             & ((q_pos - k_pos < local_blocks)
+                                | mask_vert_strided)).to(device).to(dtype))
+        num_blocks_q = triton.cdiv(q_len, block_size)
+        block_mask_dense_output = block_mask_dense[:, -num_blocks_q:]
+    if return_dense:
+        mask_dense = torch.kron(
+            block_mask_dense,
+            block_mask_dense.new_ones((block_size, block_size)),
+        )
+        causal_mask = torch.tril(torch.ones(
+            max_seqlen, max_seqlen)).type_as(mask_dense)[-q_len:]
+        mask_dense = mask_dense[..., -q_len:, :max_seqlen] * causal_mask[None]
+        if dense_mask_type == "bias":
+            mask_dense = binary_mask_to_bias(mask_dense)
+
+        return (
+            dense_to_crow_col(block_mask_dense_output),
+            block_mask_dense,
+            mask_dense,
+        )
+    else:
+        return (
+            dense_to_crow_col(block_mask_dense_output),
+            block_mask_dense,
+            None,
+        )

vllm-v0.6.2/vllm/attention/ops/hpu_paged_attn.py

@@ -0,0 +1,103 @@
+###############################################################################
+# Copyright (C) 2024 Habana Labs, Ltd. an Intel Company
+###############################################################################
+
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Tuple
+
+import torch
+from vllm_hpu_extension import cache_ops, ops
+
+# Should be the same as PARTITION_SIZE in `paged_attention_v2_launcher`.
+_PARTITION_SIZE = 512
+
+
+@dataclass
+class HPUPagedAttentionMetadata:
+    """Metadata for PagedAttention."""
+    block_list: Optional[torch.Tensor]
+    block_mapping: Optional[torch.Tensor]
+    block_usage: Optional[torch.Tensor]
+    block_indices: Optional[torch.Tensor]
+    block_offsets: Optional[torch.Tensor]
+    block_scales: Optional[torch.Tensor]
+
+
+class HPUPagedAttention:
+
+    @staticmethod
+    def get_supported_head_sizes() -> List[int]:
+        return [64, 80, 96, 112, 128, 256]
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        return (num_blocks, block_size, num_kv_heads, head_size)
+
+    @staticmethod
+    def split_kv_cache(
+        kv_cache: torch.Tensor,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        key_cache = kv_cache[0]
+        value_cache = kv_cache[1]
+        return key_cache, value_cache
+
+    @staticmethod
+    def write_to_paged_cache(key: torch.Tensor, value: torch.Tensor,
+                             key_cache: torch.Tensor,
+                             value_cache: torch.Tensor,
+                             slot_mapping: torch.Tensor, kv_cache_dtype: str,
+                             is_prompt: bool) -> None:
+        cache_ops.reshape_and_cache(key, value, key_cache, value_cache,
+                                    slot_mapping, kv_cache_dtype, is_prompt)
+
+    @staticmethod
+    def forward_decode(**kwargs) -> torch.Tensor:
+        return ops.flat_pa(**kwargs)
+
+    @staticmethod
+    def forward_prefix(
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        key_cache: torch.Tensor,
+        value_cache: torch.Tensor,
+        block_tables: torch.Tensor,
+        subquery_start_loc: torch.Tensor,
+        seq_lens_tensor: torch.Tensor,
+        context_lens: torch.Tensor,
+        max_query_len: int,
+        alibi_slopes: Optional[torch.Tensor],
+        sliding_window: Optional[int],
+    ) -> torch.Tensor:
+        raise NotImplementedError(
+            "forward_prefix is not implemented for HPUPagedAttention")
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: Dict[int, int],
+    ) -> None:
+        src_key_cache = src_kv_cache[0]
+        dst_key_cache = dst_kv_cache[0]
+        cache_ops.swap_blocks(src_key_cache, dst_key_cache, src_to_dst)
+
+        src_value_cache = src_kv_cache[1]
+        dst_value_cache = dst_kv_cache[1]
+        cache_ops.swap_blocks(src_value_cache, dst_value_cache, src_to_dst)
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: Dict[int, List[int]],
+    ) -> None:
+        key_caches = [kv_cache[0] for kv_cache in kv_caches]
+        value_caches = [kv_cache[1] for kv_cache in kv_caches]
+        cache_ops.copy_blocks(key_caches, value_caches, src_to_dists)

vllm-v0.6.2/vllm/attention/ops/ipex_attn.py

@@ -0,0 +1,123 @@
+from typing import Dict, List, Optional, Tuple
+
+import intel_extension_for_pytorch.llm.modules as ipex_modules
+import torch
+
+from vllm import _custom_ops as ops
+
+
+class PagedAttention:
+
+    @staticmethod
+    def get_supported_head_sizes() -> List[int]:
+        return [32, 64, 80, 96, 112, 128, 256]
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+        *args,
+    ) -> Tuple[int, ...]:
+        return (2, num_blocks, block_size * num_kv_heads * head_size)
+
+    @staticmethod
+    def split_kv_cache(
+        kv_cache: torch.Tensor,
+        num_kv_heads: int,
+        head_size: int,
+        *args,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        num_blocks = kv_cache.shape[1]
+
+        key_cache = kv_cache[0]
+        key_cache = key_cache.view(num_blocks, num_kv_heads, -1, head_size)
+        value_cache = kv_cache[1]
+        value_cache = value_cache.view(num_blocks, num_kv_heads, -1, head_size)
+        return key_cache, value_cache
+
+    @staticmethod
+    def write_to_paged_cache(
+        key: torch.Tensor,
+        value: torch.Tensor,
+        key_cache: torch.Tensor,
+        value_cache: torch.Tensor,
+        slot_mapping: torch.Tensor,
+        kv_cache_dtype: str,
+        k_scale: float,
+        v_scale: float,
+        *args,
+    ) -> None:
+        ipex_modules.PagedAttention.reshape_and_cache(
+            key, value, key_cache, value_cache,
+            slot_mapping.flatten().int())
+
+    @staticmethod
+    def forward_decode(
+        query: torch.Tensor,
+        key_cache: torch.Tensor,
+        value_cache: torch.Tensor,
+        block_tables: torch.Tensor,
+        context_lens: torch.Tensor,
+        max_context_len: int,
+        kv_cache_dtype: str,
+        num_kv_heads: int,
+        scale: float,
+        alibi_slopes: Optional[torch.Tensor],
+        k_scale: float,
+        v_scale: float,
+        *args,
+    ) -> torch.Tensor:
+        output = torch.empty_like(query)
+        block_size = value_cache.shape[2]
+        head_mapping = torch.arange(
+            0,
+            num_kv_heads,
+            device="cpu",
+            dtype=torch.int32,
+        ).view(num_kv_heads,
+               1).repeat_interleave(query.size(1) // num_kv_heads).flatten()
+        ipex_modules.PagedAttention.single_query_cached_kv_attention(
+            output, query.contiguous(), key_cache, value_cache, head_mapping,
+            scale, block_tables, context_lens, block_size, max_context_len,
+            alibi_slopes)
+
+        return output
+
+    @staticmethod
+    def forward_prefix(
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache_dtype: str,
+        key_cache: torch.Tensor,
+        value_cache: torch.Tensor,
+        block_tables: torch.Tensor,
+        subquery_start_loc: torch.Tensor,
+        prompt_lens_tensor: torch.Tensor,
+        context_lens: torch.Tensor,
+        max_subquery_len: int,
+        alibi_slopes: Optional[torch.Tensor],
+        *args,
+    ) -> torch.Tensor:
+        raise NotImplementedError
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: Dict[int, int],
+        *args,
+    ) -> None:
+        raise NotImplementedError
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: Dict[int, List[int]],
+        *args,
+    ) -> None:
+        key_caches = [kv_cache[0] for kv_cache in kv_caches]
+        value_caches = [kv_cache[1] for kv_cache in kv_caches]
+        ops.copy_blocks(key_caches, value_caches, src_to_dists)

vllm-v0.6.2/vllm/attention/ops/paged_attn.py

@@ -0,0 +1,253 @@
+from dataclasses import dataclass
+from typing import List, Optional, Tuple
+
+import torch
+
+from vllm import _custom_ops as ops
+from vllm.triton_utils import HAS_TRITON
+
+if HAS_TRITON:
+    from vllm.attention.ops.prefix_prefill import context_attention_fwd
+
+# Should be the same as PARTITION_SIZE in `paged_attention_v2_launcher`.
+_PARTITION_SIZE = 512
+
+
+@dataclass
+class PagedAttentionMetadata:
+    """Metadata for PagedAttention."""
+    # (batch_size,). The length of sequences (entire tokens seen so far) per
+    # sequence.
+    seq_lens_tensor: Optional[torch.Tensor]
+    # Maximum sequence length in the batch. 0 if it is prefill-only batch.
+    max_decode_seq_len: int
+    # (batch_size, max_blocks_per_seq).
+    # Block addresses per sequence. (Seq id -> list of physical block)
+    # E.g., [0, 1, 2] means tokens are stored in 0th, 1st, and 2nd blocks
+    # in the kv cache. Each block can contain up to block_size tokens.
+    # 2nd dimensions are padded up to max_blocks_per_seq if it is cuda-graph
+    # captured.
+    block_tables: Optional[torch.Tensor]
+
+
+class PagedAttention:
+
+    @staticmethod
+    def get_supported_head_sizes() -> List[int]:
+        return [32, 64, 80, 96, 112, 120, 128, 192, 256]
+
+    @staticmethod
+    def get_kv_cache_shape(
+        num_blocks: int,
+        block_size: int,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[int, ...]:
+        return (2, num_blocks, block_size * num_kv_heads * head_size)
+
+    @staticmethod
+    def split_kv_cache(
+        kv_cache: torch.Tensor,
+        num_kv_heads: int,
+        head_size: int,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        x = 16 // kv_cache.element_size()
+        num_blocks = kv_cache.shape[1]
+
+        key_cache = kv_cache[0]
+        key_cache = key_cache.view(num_blocks, num_kv_heads, head_size // x,
+                                   -1, x)
+        value_cache = kv_cache[1]
+        value_cache = value_cache.view(num_blocks, num_kv_heads, head_size, -1)
+        return key_cache, value_cache
+
+    @staticmethod
+    def write_to_paged_cache(
+        key: torch.Tensor,
+        value: torch.Tensor,
+        key_cache: torch.Tensor,
+        value_cache: torch.Tensor,
+        slot_mapping: torch.Tensor,
+        kv_cache_dtype: str,
+        k_scale: float,
+        v_scale: float,
+    ) -> None:
+        ops.reshape_and_cache(
+            key,
+            value,
+            key_cache,
+            value_cache,
+            slot_mapping.flatten(),
+            kv_cache_dtype,
+            k_scale,
+            v_scale,
+        )
+
+    @staticmethod
+    def forward_decode(
+        query: torch.Tensor,
+        key_cache: torch.Tensor,
+        value_cache: torch.Tensor,
+        block_tables: torch.Tensor,
+        seq_lens: torch.Tensor,
+        max_seq_len: int,
+        kv_cache_dtype: str,
+        num_kv_heads: int,
+        scale: float,
+        alibi_slopes: Optional[torch.Tensor],
+        k_scale: float,
+        v_scale: float,
+        tp_rank: int = 0,
+        blocksparse_local_blocks: int = 0,
+        blocksparse_vert_stride: int = 0,
+        blocksparse_block_size: int = 64,
+        blocksparse_head_sliding_step: int = 0,
+    ) -> torch.Tensor:
+        if blocksparse_vert_stride is not None and blocksparse_vert_stride > 1:
+            # use blocksparse paged attention
+            block_size = value_cache.size(-1)
+            assert (blocksparse_block_size > 0 and
+                    blocksparse_block_size % block_size == 0), \
+                (f"{blocksparse_block_size=} needs to be a multiple of"
+                 f"{block_size=} used in block_tables.")
+
+        output = torch.empty_like(query)
+        block_size = value_cache.shape[3]
+        num_seqs, num_heads, head_size = query.shape
+        max_num_partitions = ((max_seq_len + _PARTITION_SIZE - 1) //
+                              _PARTITION_SIZE)
+        # NOTE(woosuk): We use a simple heuristic to decide whether to use
+        # PagedAttention V1 or V2. If the number of partitions is 1, we use
+        # V1 to avoid the overhead of reduction. Also, if the number of
+        # sequences or heads is large, we use V1 since there is enough work
+        # to parallelize.
+        # TODO(woosuk): Tune this heuristic.
+        # For context len > 8192, use V2 kernel to avoid shared memory shortage.
+        use_v1 = (max_seq_len <= 8192
+                  and (max_num_partitions == 1 or num_seqs * num_heads > 512))
+
+        if use_v1:
+            # Run PagedAttention V1.
+            ops.paged_attention_v1(
+                output,
+                query,
+                key_cache,
+                value_cache,
+                num_kv_heads,
+                scale,
+                block_tables,
+                seq_lens,
+                block_size,
+                max_seq_len,
+                alibi_slopes,
+                kv_cache_dtype,
+                k_scale,
+                v_scale,
+                tp_rank,
+                blocksparse_local_blocks,
+                blocksparse_vert_stride,
+                blocksparse_block_size,
+                blocksparse_head_sliding_step,
+            )
+        else:
+            # Run PagedAttention V2.
+            assert _PARTITION_SIZE % block_size == 0
+            tmp_output = torch.empty(
+                size=(num_seqs, num_heads, max_num_partitions, head_size),
+                dtype=output.dtype,
+                device=output.device,
+            )
+            exp_sums = torch.empty(
+                size=(num_seqs, num_heads, max_num_partitions),
+                dtype=torch.float32,
+                device=output.device,
+            )
+            max_logits = torch.empty_like(exp_sums)
+            ops.paged_attention_v2(
+                output,
+                exp_sums,
+                max_logits,
+                tmp_output,
+                query,
+                key_cache,
+                value_cache,
+                num_kv_heads,
+                scale,
+                block_tables,
+                seq_lens,
+                block_size,
+                max_seq_len,
+                alibi_slopes,
+                kv_cache_dtype,
+                k_scale,
+                v_scale,
+                tp_rank,
+                blocksparse_local_blocks,
+                blocksparse_vert_stride,
+                blocksparse_block_size,
+                blocksparse_head_sliding_step,
+            )
+        return output
+
+    @staticmethod
+    def forward_prefix(
+        query: torch.Tensor,
+        key: torch.Tensor,
+        value: torch.Tensor,
+        kv_cache_dtype: str,
+        key_cache: torch.Tensor,
+        value_cache: torch.Tensor,
+        block_tables: torch.Tensor,
+        query_start_loc: torch.Tensor,
+        seq_lens_tensor: torch.Tensor,
+        context_lens: torch.Tensor,
+        max_query_len: int,
+        alibi_slopes: Optional[torch.Tensor],
+        sliding_window: Optional[int],
+        k_scale: float,
+        v_scale: float,
+    ) -> torch.Tensor:
+        output = torch.empty_like(query)
+        context_attention_fwd(
+            query,
+            key,
+            value,
+            output,
+            kv_cache_dtype,
+            key_cache,
+            value_cache,
+            block_tables,
+            # query_start_loc is (batch_size + 1,)
+            query_start_loc[:-1],
+            seq_lens_tensor,
+            context_lens,
+            max_query_len,
+            k_scale,
+            v_scale,
+            alibi_slopes,
+            sliding_window,
+        )
+        return output
+
+    @staticmethod
+    def swap_blocks(
+        src_kv_cache: torch.Tensor,
+        dst_kv_cache: torch.Tensor,
+        src_to_dst: torch.Tensor,
+    ) -> None:
+        src_key_cache = src_kv_cache[0]
+        dst_key_cache = dst_kv_cache[0]
+        ops.swap_blocks(src_key_cache, dst_key_cache, src_to_dst)
+
+        src_value_cache = src_kv_cache[1]
+        dst_value_cache = dst_kv_cache[1]
+        ops.swap_blocks(src_value_cache, dst_value_cache, src_to_dst)
+
+    @staticmethod
+    def copy_blocks(
+        kv_caches: List[torch.Tensor],
+        src_to_dists: torch.Tensor,
+    ) -> None:
+        key_caches = [kv_cache[0] for kv_cache in kv_caches]
+        value_caches = [kv_cache[1] for kv_cache in kv_caches]
+        ops.copy_blocks(key_caches, value_caches, src_to_dists)

vllm-v0.6.2/vllm/attention/ops/prefix_prefill.py

@@ -0,0 +1,861 @@
+# The kernels in this file are adapted from LightLLM's context_attention_fwd:
+# https://github.com/ModelTC/lightllm/blob/main/lightllm/models/llama/triton_kernel/context_flashattention_nopad.py
+
+import torch
+import triton
+import triton.language as tl
+
+from vllm.platforms import current_platform
+
+if triton.__version__ >= "2.1.0":
+
+    @triton.jit
+    def _fwd_kernel(
+        Q,
+        K,
+        V,
+        K_cache,
+        V_cache,
+        B_Loc,
+        sm_scale,
+        k_scale,
+        v_scale,
+        B_Start_Loc,
+        B_Seqlen,
+        B_Ctxlen,
+        block_size,
+        x,
+        Out,
+        stride_b_loc_b,
+        stride_b_loc_s,
+        stride_qbs,
+        stride_qh,
+        stride_qd,
+        stride_kbs,
+        stride_kh,
+        stride_kd,
+        stride_vbs,
+        stride_vh,
+        stride_vd,
+        stride_obs,
+        stride_oh,
+        stride_od,
+        stride_k_cache_bs,
+        stride_k_cache_h,
+        stride_k_cache_d,
+        stride_k_cache_bl,
+        stride_k_cache_x,
+        stride_v_cache_bs,
+        stride_v_cache_h,
+        stride_v_cache_d,
+        stride_v_cache_bl,
+        num_queries_per_kv: int,
+        BLOCK_M: tl.constexpr,
+        BLOCK_DMODEL: tl.constexpr,  # head size
+        BLOCK_DMODEL_PADDED: tl.constexpr,  # head size padded to a power of 2
+        BLOCK_N: tl.constexpr,
+        SLIDING_WINDOW: tl.constexpr,
+    ):
+        cur_batch = tl.program_id(0)
+        cur_head = tl.program_id(1)
+        start_m = tl.program_id(2)
+
+        cur_kv_head = cur_head // num_queries_per_kv
+
+        cur_batch_ctx_len = tl.load(B_Ctxlen + cur_batch)
+        cur_batch_seq_len = tl.load(B_Seqlen + cur_batch)
+        cur_batch_in_all_start_index = tl.load(B_Start_Loc + cur_batch)
+        cur_batch_query_len = cur_batch_seq_len - cur_batch_ctx_len
+
+        # start position inside of the query
+        # generally, N goes over kv, while M goes over query_len
+        block_start_loc = BLOCK_M * start_m
+
+        # initialize offsets
+        # [N]; starts at 0
+        offs_n = tl.arange(0, BLOCK_N)
+        # [D]; starts at 0
+        offs_d = tl.arange(0, BLOCK_DMODEL_PADDED)
+        # [M]; starts at current position in query
+        offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
+        # [M,D]
+        off_q = (
+            (cur_batch_in_all_start_index + offs_m[:, None]) * stride_qbs +
+            cur_head * stride_qh + offs_d[None, :] * stride_qd)
+
+        dim_mask = tl.where(
+            tl.arange(0, BLOCK_DMODEL_PADDED) < BLOCK_DMODEL, 1,
+            0).to(tl.int1)  # [D]
+
+        q = tl.load(Q + off_q,
+                    mask=dim_mask[None, :] &
+                    (offs_m[:, None] < cur_batch_query_len),
+                    other=0.0)  # [M,D]
+
+        # initialize pointer to m and l
+        m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf")  # [M]
+        l_i = tl.zeros([BLOCK_M], dtype=tl.float32)  # [M]
+        acc = tl.zeros([BLOCK_M, BLOCK_DMODEL_PADDED],
+                       dtype=tl.float32)  # [M,D]
+
+        # compute query against context (no causal mask here)
+        for start_n in range(0, cur_batch_ctx_len, BLOCK_N):
+            start_n = tl.multiple_of(start_n, BLOCK_N)
+            # -- compute qk ----
+            bn = tl.load(B_Loc + cur_batch * stride_b_loc_b +
+                         ((start_n + offs_n) // block_size) * stride_b_loc_s,
+                         mask=(start_n + offs_n) < cur_batch_ctx_len,
+                         other=0)  # [N]
+            # [D,N]
+            off_k = (bn[None, :] * stride_k_cache_bs +
+                     cur_kv_head * stride_k_cache_h +
+                     (offs_d[:, None] // x) * stride_k_cache_d +
+                     ((start_n + offs_n[None, :]) % block_size) *
+                     stride_k_cache_bl +
+                     (offs_d[:, None] % x) * stride_k_cache_x)
+            # [N,D]
+            off_v = (
+                bn[:, None] * stride_v_cache_bs +
+                cur_kv_head * stride_v_cache_h +
+                offs_d[None, :] * stride_v_cache_d +
+                (start_n + offs_n[:, None]) % block_size * stride_v_cache_bl)
+            k_load = tl.load(K_cache + off_k,
+                             mask=dim_mask[:, None] &
+                             ((start_n + offs_n[None, :]) < cur_batch_ctx_len),
+                             other=0.0)  # [D,N]
+
+            if k_load.dtype.is_fp8():
+                k = (k_load.to(tl.float32) * k_scale).to(q.dtype)
+            else:
+                k = k_load
+
+            qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)  # [M,N]
+            qk += tl.dot(q, k)
+            qk = tl.where((start_n + offs_n[None, :]) < cur_batch_ctx_len, qk,
+                          float("-inf"))
+            qk *= sm_scale
+            if SLIDING_WINDOW > 0:
+                # (cur_batch_ctx_len + offs_m[:, None]) are the positions of
+                # Q entries in sequence
+                # (start_n + offs_n[None, :]) are the positions of
+                # KV entries in sequence
+                # So the condition makes sure each entry in Q only attends
+                # to KV entries not more than SLIDING_WINDOW away.
+                #
+                # We can't use -inf here, because the
+                # sliding window may lead to the entire row being masked.
+                # This then makes m_ij contain -inf, which causes NaNs in
+                # exp().
+                qk = tl.where((cur_batch_ctx_len + offs_m[:, None]) -
+                              (start_n + offs_n[None, :]) < SLIDING_WINDOW, qk,
+                              -10000)
+
+            # -- compute m_ij, p, l_ij
+            m_ij = tl.max(qk, 1)  # [M]
+            p = tl.exp(qk - m_ij[:, None])  # [M,N]
+            l_ij = tl.sum(p, 1)  # [M]
+            # -- update m_i and l_i
+            m_i_new = tl.maximum(m_i, m_ij)  # [M]
+            alpha = tl.exp(m_i - m_i_new)  # [M]
+            beta = tl.exp(m_ij - m_i_new)  # [M]
+            l_i_new = alpha * l_i + beta * l_ij  # [M]
+
+            # -- update output accumulator --
+            # scale p
+            p_scale = beta / l_i_new
+            p = p * p_scale[:, None]
+            # scale acc
+            acc_scale = l_i / l_i_new * alpha
+            acc = acc * acc_scale[:, None]
+            # update acc
+            v_load = tl.load(V_cache + off_v,
+                             mask=dim_mask[None, :] &
+                             ((start_n + offs_n[:, None]) < cur_batch_ctx_len),
+                             other=0.0)  # [N,D]
+            if v_load.dtype.is_fp8():
+                v = (v_load.to(tl.float32) * v_scale).to(q.dtype)
+            else:
+                v = v_load
+            p = p.to(v.dtype)
+
+            acc += tl.dot(p, v)
+            # # update m_i and l_i
+            l_i = l_i_new
+            m_i = m_i_new
+
+        off_k = (offs_n[None, :] * stride_kbs + cur_kv_head * stride_kh +
+                 offs_d[:, None] * stride_kd)
+        off_v = (offs_n[:, None] * stride_vbs + cur_kv_head * stride_vh +
+                 offs_d[None, :] * stride_vd)
+        k_ptrs = K + off_k
+        v_ptrs = V + off_v
+
+        # block_mask is 0 when we're already past the current query length
+        block_mask = tl.where(block_start_loc < cur_batch_query_len, 1, 0)
+
+        # compute query against itself (with causal mask)
+        for start_n in range(0, block_mask * (start_m + 1) * BLOCK_M, BLOCK_N):
+            start_n = tl.multiple_of(start_n, BLOCK_N)
+            # -- compute qk ----
+            k = tl.load(k_ptrs +
+                        (cur_batch_in_all_start_index + start_n) * stride_kbs,
+                        mask=dim_mask[:, None] &
+                        ((start_n + offs_n[None, :]) < cur_batch_query_len),
+                        other=0.0)
+
+            qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
+            qk += tl.dot(q, k)
+            qk *= sm_scale
+            # apply causal mask
+            qk = tl.where(offs_m[:, None] >= (start_n + offs_n[None, :]), qk,
+                          float("-inf"))
+            if SLIDING_WINDOW > 0:
+                qk = tl.where(
+                    offs_m[:, None] -
+                    (start_n + offs_n[None, :]) < SLIDING_WINDOW, qk, -10000)
+
+            # -- compute m_ij, p, l_ij
+            m_ij = tl.max(qk, 1)
+            p = tl.exp(qk - m_ij[:, None])
+            l_ij = tl.sum(p, 1)
+            # -- update m_i and l_i
+            m_i_new = tl.maximum(m_i, m_ij)
+            alpha = tl.exp(m_i - m_i_new)
+            beta = tl.exp(m_ij - m_i_new)
+            l_i_new = alpha * l_i + beta * l_ij
+            # -- update output accumulator --
+            # scale p
+            p_scale = beta / l_i_new
+            p = p * p_scale[:, None]
+            # scale acc
+            acc_scale = l_i / l_i_new * alpha
+            acc = acc * acc_scale[:, None]
+            # update acc
+            v = tl.load(v_ptrs +
+                        (cur_batch_in_all_start_index + start_n) * stride_vbs,
+                        mask=dim_mask[None, :] &
+                        ((start_n + offs_n[:, None]) < cur_batch_query_len),
+                        other=0.0)
+            p = p.to(v.dtype)
+
+            acc += tl.dot(p, v)
+            # update m_i and l_i
+            l_i = l_i_new
+            m_i = m_i_new
+        # initialize pointers to output
+        off_o = (
+            (cur_batch_in_all_start_index + offs_m[:, None]) * stride_obs +
+            cur_head * stride_oh + offs_d[None, :] * stride_od)
+        out_ptrs = Out + off_o
+        tl.store(out_ptrs,
+                 acc,
+                 mask=dim_mask[None, :] &
+                 (offs_m[:, None] < cur_batch_query_len))
+        return
+
+    @triton.jit
+    def _fwd_kernel_flash_attn_v2(
+        Q,
+        K,
+        V,
+        K_cache,
+        V_cache,
+        B_Loc,
+        sm_scale,
+        B_Start_Loc,
+        B_Seqlen,
+        B_Ctxlen,
+        block_size,
+        x,
+        Out,
+        stride_b_loc_b,
+        stride_b_loc_s,
+        stride_qbs,
+        stride_qh,
+        stride_qd,
+        stride_kbs,
+        stride_kh,
+        stride_kd,
+        stride_vbs,
+        stride_vh,
+        stride_vd,
+        stride_obs,
+        stride_oh,
+        stride_od,
+        stride_k_cache_bs,
+        stride_k_cache_h,
+        stride_k_cache_d,
+        stride_k_cache_bl,
+        stride_k_cache_x,
+        stride_v_cache_bs,
+        stride_v_cache_h,
+        stride_v_cache_d,
+        stride_v_cache_bl,
+        num_queries_per_kv: int,
+        BLOCK_M: tl.constexpr,
+        BLOCK_DMODEL: tl.constexpr,
+        BLOCK_N: tl.constexpr,
+    ):
+        cur_batch = tl.program_id(0)
+        cur_head = tl.program_id(1)
+        start_m = tl.program_id(2)
+
+        cur_kv_head = cur_head // num_queries_per_kv
+
+        cur_batch_ctx_len = tl.load(B_Ctxlen + cur_batch)
+        cur_batch_seq_len = tl.load(B_Seqlen + cur_batch)
+        cur_batch_in_all_start_index = tl.load(B_Start_Loc + cur_batch)
+
+        block_start_loc = BLOCK_M * start_m
+
+        # initialize offsets
+        offs_n = tl.arange(0, BLOCK_N)
+        offs_d = tl.arange(0, BLOCK_DMODEL)
+        offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
+        off_q = (
+            (cur_batch_in_all_start_index + offs_m[:, None]) * stride_qbs +
+            cur_head * stride_qh + offs_d[None, :] * stride_qd)
+
+        q = tl.load(
+            Q + off_q,
+            mask=offs_m[:, None] < cur_batch_seq_len - cur_batch_ctx_len,
+            other=0.0)
+
+        # # initialize pointer to m and l
+        m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf")
+        l_i = tl.zeros([BLOCK_M], dtype=tl.float32)
+        acc = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=tl.float32)
+
+        for start_n in range(0, cur_batch_ctx_len, BLOCK_N):
+            start_n = tl.multiple_of(start_n, BLOCK_N)
+            # -- compute qk ----
+            bn = tl.load(B_Loc + cur_batch * stride_b_loc_b +
+                         ((start_n + offs_n) // block_size) * stride_b_loc_s,
+                         mask=(start_n + offs_n) < cur_batch_ctx_len,
+                         other=0)
+            off_k = (bn[None, :] * stride_k_cache_bs +
+                     cur_kv_head * stride_k_cache_h +
+                     (offs_d[:, None] // x) * stride_k_cache_d +
+                     ((start_n + offs_n[None, :]) % block_size) *
+                     stride_k_cache_bl +
+                     (offs_d[:, None] % x) * stride_k_cache_x)
+            off_v = (
+                bn[:, None] * stride_v_cache_bs +
+                cur_kv_head * stride_v_cache_h +
+                offs_d[None, :] * stride_v_cache_d +
+                (start_n + offs_n[:, None]) % block_size * stride_v_cache_bl)
+            k = tl.load(K_cache + off_k,
+                        mask=(start_n + offs_n[None, :]) < cur_batch_ctx_len,
+                        other=0.0)
+            qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
+            qk += tl.dot(q, k)
+            qk = tl.where((start_n + offs_n[None, :]) < cur_batch_ctx_len, qk,
+                          float("-inf"))
+            qk *= sm_scale
+
+            # -- compute m_ij, p, l_ij
+            m_ij = tl.max(qk, 1)
+            m_i_new = tl.maximum(m_i, m_ij)
+            p = tl.math.exp(qk - m_i_new[:, None])
+            l_ij = tl.sum(p, 1)
+            # -- update m_i and l_i
+
+            alpha = tl.math.exp(m_i - m_i_new)
+            l_i_new = alpha * l_i + l_ij
+            # -- update output accumulator --
+            # scale p
+            # scale acc
+            acc_scale = alpha
+            # acc_scale = l_i / l_i_new * alpha
+            acc = acc * acc_scale[:, None]
+            # update acc
+            v = tl.load(V_cache + off_v,
+                        mask=(start_n + offs_n[:, None]) < cur_batch_ctx_len,
+                        other=0.0)
+
+            p = p.to(v.dtype)
+            acc += tl.dot(p, v)
+            # update m_i and l_i
+            l_i = l_i_new
+            m_i = m_i_new
+
+        off_k = (offs_n[None, :] * stride_kbs + cur_kv_head * stride_kh +
+                 offs_d[:, None] * stride_kd)
+        off_v = (offs_n[:, None] * stride_vbs + cur_kv_head * stride_vh +
+                 offs_d[None, :] * stride_vd)
+        k_ptrs = K + off_k
+        v_ptrs = V + off_v
+
+        block_mask = tl.where(
+            block_start_loc < cur_batch_seq_len - cur_batch_ctx_len, 1, 0)
+
+        for start_n in range(0, block_mask * (start_m + 1) * BLOCK_M, BLOCK_N):
+            start_n = tl.multiple_of(start_n, BLOCK_N)
+            # -- compute qk ----
+            k = tl.load(k_ptrs +
+                        (cur_batch_in_all_start_index + start_n) * stride_kbs,
+                        mask=(start_n + offs_n[None, :]) <
+                        cur_batch_seq_len - cur_batch_ctx_len,
+                        other=0.0)
+
+            qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
+            qk += tl.dot(q, k)
+            qk *= sm_scale
+            qk = tl.where(offs_m[:, None] >= (start_n + offs_n[None, :]), qk,
+                          float("-inf"))
+
+            # -- compute m_ij, p, l_ij
+            m_ij = tl.max(qk, 1)
+            m_i_new = tl.maximum(m_i, m_ij)
+            p = tl.math.exp(qk - m_i_new[:, None])
+            l_ij = tl.sum(p, 1)
+            # -- update m_i and l_i
+
+            alpha = tl.math.exp(m_i - m_i_new)
+            l_i_new = alpha * l_i + l_ij
+            # -- update output accumulator --
+            # scale p
+            # scale acc
+            acc_scale = alpha
+            # acc_scale = l_i / l_i_new * alpha
+            acc = acc * acc_scale[:, None]
+            # update acc
+            v = tl.load(v_ptrs +
+                        (cur_batch_in_all_start_index + start_n) * stride_vbs,
+                        mask=(start_n + offs_n[:, None]) <
+                        cur_batch_seq_len - cur_batch_ctx_len,
+                        other=0.0)
+
+            p = p.to(v.dtype)
+            acc += tl.dot(p, v)
+            # update m_i and l_i
+            l_i = l_i_new
+            m_i = m_i_new
+
+        # acc /= l_i[:, None]
+        # initialize pointers to output
+        off_o = (
+            (cur_batch_in_all_start_index + offs_m[:, None]) * stride_obs +
+            cur_head * stride_oh + offs_d[None, :] * stride_od)
+        out_ptrs = Out + off_o
+        tl.store(out_ptrs,
+                 acc,
+                 mask=offs_m[:, None] < cur_batch_seq_len - cur_batch_ctx_len)
+        return
+
+    @triton.jit
+    def _fwd_kernel_alibi(
+        Q,
+        K,
+        V,
+        K_cache,
+        V_cache,
+        B_Loc,
+        sm_scale,
+        k_scale,
+        v_scale,
+        B_Start_Loc,
+        B_Seqlen,
+        B_Ctxlen,
+        Alibi_slopes,
+        block_size,
+        x,
+        Out,
+        stride_b_loc_b,
+        stride_b_loc_s,
+        stride_qbs,
+        stride_qh,
+        stride_qd,
+        stride_kbs,
+        stride_kh,
+        stride_kd,
+        stride_vbs,
+        stride_vh,
+        stride_vd,
+        stride_obs,
+        stride_oh,
+        stride_od,
+        stride_k_cache_bs,
+        stride_k_cache_h,
+        stride_k_cache_d,
+        stride_k_cache_bl,
+        stride_k_cache_x,
+        stride_v_cache_bs,
+        stride_v_cache_h,
+        stride_v_cache_d,
+        stride_v_cache_bl,
+        num_queries_per_kv: int,
+        BLOCK_M: tl.constexpr,
+        BLOCK_DMODEL: tl.constexpr,  # head size
+        BLOCK_DMODEL_PADDED: tl.constexpr,  # head size padded to a power of 2
+        BLOCK_N: tl.constexpr,
+    ):
+        # attn_bias[]
+        cur_batch = tl.program_id(0)
+        cur_head = tl.program_id(1)
+        start_m = tl.program_id(2)
+
+        cur_kv_head = cur_head // num_queries_per_kv
+
+        # cur_batch_seq_len: the length of prompts
+        # cur_batch_ctx_len: the length of prefix
+        # cur_batch_in_all_start_index: the start id of the dim=0
+        cur_batch_ctx_len = tl.load(B_Ctxlen + cur_batch)
+        cur_batch_seq_len = tl.load(B_Seqlen + cur_batch)
+        cur_batch_in_all_start_index = tl.load(B_Start_Loc + cur_batch)
+
+        block_start_loc = BLOCK_M * start_m
+
+        # initialize offsets
+        offs_n = tl.arange(0, BLOCK_N)
+        offs_d = tl.arange(0, BLOCK_DMODEL_PADDED)
+        offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
+        off_q = (
+            (cur_batch_in_all_start_index + offs_m[:, None]) * stride_qbs +
+            cur_head * stride_qh + offs_d[None, :] * stride_qd)
+
+        dim_mask = tl.where(
+            tl.arange(0, BLOCK_DMODEL_PADDED) < BLOCK_DMODEL, 1, 0).to(tl.int1)
+
+        q = tl.load(Q + off_q,
+                    mask=dim_mask[None, :] &
+                    (offs_m[:, None] < cur_batch_seq_len - cur_batch_ctx_len),
+                    other=0.0)
+
+        # # initialize pointer to m and l
+        m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf")
+        l_i = tl.zeros([BLOCK_M], dtype=tl.float32)
+        acc = tl.zeros([BLOCK_M, BLOCK_DMODEL_PADDED], dtype=tl.float32)
+
+        alibi_slope = tl.load(Alibi_slopes + cur_head)
+        alibi_start_q = tl.arange(
+            0, BLOCK_M) + block_start_loc + cur_batch_ctx_len
+        alibi_start_k = 0
+        for start_n in range(0, cur_batch_ctx_len, BLOCK_N):
+            start_n = tl.multiple_of(start_n, BLOCK_N)
+            # -- compute qk ----
+            bn = tl.load(B_Loc + cur_batch * stride_b_loc_b +
+                         ((start_n + offs_n) // block_size) * stride_b_loc_s,
+                         mask=(start_n + offs_n) < cur_batch_ctx_len,
+                         other=0)
+            off_k = (bn[None, :] * stride_k_cache_bs +
+                     cur_kv_head * stride_k_cache_h +
+                     (offs_d[:, None] // x) * stride_k_cache_d +
+                     ((start_n + offs_n[None, :]) % block_size) *
+                     stride_k_cache_bl +
+                     (offs_d[:, None] % x) * stride_k_cache_x)
+            off_v = (
+                bn[:, None] * stride_v_cache_bs +
+                cur_kv_head * stride_v_cache_h +
+                offs_d[None, :] * stride_v_cache_d +
+                (start_n + offs_n[:, None]) % block_size * stride_v_cache_bl)
+            k_load = tl.load(K_cache + off_k,
+                             mask=dim_mask[:, None] &
+                             ((start_n + offs_n[None, :]) < cur_batch_ctx_len),
+                             other=0.0)  # [D,N]
+
+            if k_load.dtype.is_fp8():
+                k = (k_load.to(tl.float32) * k_scale).to(q.dtype)
+            else:
+                k = k_load
+
+            qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
+            qk += tl.dot(q, k)
+            qk = tl.where((start_n + offs_n[None, :]) < cur_batch_ctx_len, qk,
+                          float("-inf"))
+            qk *= sm_scale
+
+            # load alibi
+            alibi = (tl.arange(0, BLOCK_N)[None, :] + alibi_start_k -
+                     alibi_start_q[:, None]) * alibi_slope
+            alibi = tl.where(
+                (alibi <= 0) & (alibi_start_q[:, None] < cur_batch_seq_len),
+                alibi, float("-inf"))
+            qk += alibi
+            alibi_start_k += BLOCK_N
+
+            # -- compute m_ij, p, l_ij
+            m_ij = tl.max(qk, 1)
+            m_i_new = tl.maximum(m_i, m_ij)
+            p = tl.math.exp(qk - m_i_new[:, None])
+            l_ij = tl.sum(p, 1)
+            # -- update m_i and l_i
+
+            alpha = tl.math.exp(m_i - m_i_new)
+            l_i_new = alpha * l_i + l_ij
+            # -- update output accumulator --
+            # scale p
+            # scale acc
+            acc_scale = alpha
+            # acc_scale = l_i / l_i_new * alpha
+            acc = acc * acc_scale[:, None]
+            # update acc
+            v_load = tl.load(V_cache + off_v,
+                             mask=dim_mask[None, :] &
+                             ((start_n + offs_n[:, None]) < cur_batch_ctx_len),
+                             other=0.0)
+            if v_load.dtype.is_fp8():
+                v = (v_load.to(tl.float32) * v_scale).to(q.dtype)
+            else:
+                v = v_load
+            p = p.to(v.dtype)
+
+            acc += tl.dot(p, v, allow_tf32=False)
+            # update m_i and l_i
+            l_i = l_i_new
+            m_i = m_i_new
+
+        off_k = (offs_n[None, :] * stride_kbs + cur_kv_head * stride_kh +
+                 offs_d[:, None] * stride_kd)
+        off_v = (offs_n[:, None] * stride_vbs + cur_kv_head * stride_vh +
+                 offs_d[None, :] * stride_vd)
+        k_ptrs = K + off_k
+        v_ptrs = V + off_v
+
+        block_mask = tl.where(
+            block_start_loc < cur_batch_seq_len - cur_batch_ctx_len, 1, 0)
+
+        # init alibi
+        alibi_slope = tl.load(Alibi_slopes + cur_head)
+        alibi_start_q = tl.arange(
+            0, BLOCK_M) + block_start_loc + cur_batch_ctx_len
+        alibi_start_k = cur_batch_ctx_len
+        # # init debugger
+        # offset_db_q = tl.arange(0, BLOCK_M) + block_start_loc
+        # offset_db_k = tl.arange(0, BLOCK_N)
+        # calc q[BLOCK_M, BLOCK_MODEL] mul k[prefix_len: , BLOCK_DMODEL]
+        for start_n in range(0, block_mask * (start_m + 1) * BLOCK_M, BLOCK_N):
+            start_n = tl.multiple_of(start_n, BLOCK_N)
+            # -- compute qk ----
+            k = tl.load(k_ptrs +
+                        (cur_batch_in_all_start_index + start_n) * stride_kbs,
+                        mask=dim_mask[:, None] &
+                        ((start_n + offs_n[None, :]) <
+                         cur_batch_seq_len - cur_batch_ctx_len),
+                        other=0.0)
+
+            qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
+            qk += tl.dot(q, k, allow_tf32=False)
+            qk *= sm_scale
+            qk = tl.where(offs_m[:, None] >= (start_n + offs_n[None, :]), qk,
+                          float("-inf"))
+
+            # load alibi
+            alibi = (tl.arange(0, BLOCK_N)[None, :] + alibi_start_k -
+                     alibi_start_q[:, None]) * alibi_slope
+            alibi = tl.where(
+                (alibi <= 0) & (alibi_start_q[:, None] < cur_batch_seq_len),
+                alibi, float("-inf"))
+            qk += alibi
+            alibi_start_k += BLOCK_N
+
+            # -- compute m_ij, p, l_ij
+            m_ij = tl.max(qk, 1)
+            m_i_new = tl.maximum(m_i, m_ij)
+            p = tl.math.exp(qk - m_i_new[:, None])
+            l_ij = tl.sum(p, 1)
+            # -- update m_i and l_i
+
+            alpha = tl.math.exp(m_i - m_i_new)
+            l_i_new = alpha * l_i + l_ij
+            # -- update output accumulator --
+            # scale p
+            # scale acc
+            acc_scale = alpha
+            # acc_scale = l_i / l_i_new * alpha
+            acc = acc * acc_scale[:, None]
+            # update acc
+            v = tl.load(v_ptrs +
+                        (cur_batch_in_all_start_index + start_n) * stride_vbs,
+                        mask=dim_mask[None, :] &
+                        ((start_n + offs_n[:, None]) <
+                         cur_batch_seq_len - cur_batch_ctx_len),
+                        other=0.0)
+            p = p.to(v.dtype)
+
+            acc += tl.dot(p, v, allow_tf32=False)
+            # update m_i and l_i
+            l_i = l_i_new
+            m_i = m_i_new
+
+        acc = acc / l_i[:, None]
+
+        # initialize pointers to output
+        off_o = (
+            (cur_batch_in_all_start_index + offs_m[:, None]) * stride_obs +
+            cur_head * stride_oh + offs_d[None, :] * stride_od)
+        out_ptrs = Out + off_o
+        tl.store(out_ptrs,
+                 acc,
+                 mask=dim_mask[None, :] &
+                 (offs_m[:, None] < cur_batch_seq_len - cur_batch_ctx_len))
+        return
+
+    @torch.inference_mode()
+    def context_attention_fwd(q,
+                              k,
+                              v,
+                              o,
+                              kv_cache_dtype: str,
+                              k_cache,
+                              v_cache,
+                              b_loc,
+                              b_start_loc,
+                              b_seq_len,
+                              b_ctx_len,
+                              max_input_len,
+                              k_scale: float = 1.0,
+                              v_scale: float = 1.0,
+                              alibi_slopes=None,
+                              sliding_window=None):
+
+        BLOCK = 128 if current_platform.has_device_capability(80) else 64
+        NUM_WARPS = 8
+
+        # need to reduce num. blocks when using fp32
+        # due to increased use of GPU shared memory
+        if q.dtype is torch.float32:
+            BLOCK = BLOCK // 2
+
+        # Conversion of FP8 Tensor from uint8 storage to
+        # appropriate torch.dtype for interpretation by Triton
+        if "fp8" in kv_cache_dtype:
+            assert (k_cache.dtype == torch.uint8)
+            assert (v_cache.dtype == torch.uint8)
+
+            if kv_cache_dtype in ("fp8", "fp8_e4m3"):
+                target_dtype = torch.float8_e4m3fn
+            elif kv_cache_dtype == "fp8_e5m2":
+                target_dtype = torch.float8_e5m2
+            else:
+                raise ValueError("Unsupported FP8 dtype:", kv_cache_dtype)
+
+            k_cache = k_cache.view(target_dtype)
+            v_cache = v_cache.view(target_dtype)
+
+        if (k_cache.dtype == torch.uint8
+                or v_cache.dtype == torch.uint8 and kv_cache_dtype == "auto"):
+            raise ValueError("kv_cache_dtype='auto' unsupported for\
+                FP8 KV Cache prefill kernel")
+
+        # shape constraints
+        Lq, Lk, Lv = q.shape[-1], k.shape[-1], v.shape[-1]
+        assert Lq == Lk and Lk == Lv
+        # round up Lk to a power of 2 - this is required for Triton block size
+        Lk_padded = triton.next_power_of_2(Lk)
+
+        sm_scale = 1.0 / (Lq**0.5)
+        batch, head = b_seq_len.shape[0], q.shape[1]
+        num_queries_per_kv = q.shape[1] // k.shape[1]
+
+        grid = (batch, head, triton.cdiv(max_input_len, BLOCK))  # batch, head,
+
+        # 0 means "disable"
+        if sliding_window is None or sliding_window <= 0:
+            sliding_window = 0
+
+        if alibi_slopes is not None:
+            _fwd_kernel_alibi[grid](
+                q,
+                k,
+                v,
+                k_cache,
+                v_cache,
+                b_loc,
+                sm_scale,
+                k_scale,
+                v_scale,
+                b_start_loc,
+                b_seq_len,
+                b_ctx_len,
+                alibi_slopes,
+                v_cache.shape[3],
+                k_cache.shape[4],
+                o,
+                b_loc.stride(0),
+                b_loc.stride(1),
+                q.stride(0),
+                q.stride(1),
+                q.stride(2),
+                k.stride(0),
+                k.stride(1),
+                k.stride(2),
+                v.stride(0),
+                v.stride(1),
+                v.stride(2),
+                o.stride(0),
+                o.stride(1),
+                o.stride(2),
+                k_cache.stride(0),
+                k_cache.stride(1),
+                k_cache.stride(2),
+                k_cache.stride(3),
+                k_cache.stride(
+                    4
+                ),  #[num_blocks, num_kv_heads, head_size/x, block_size, x]
+                v_cache.stride(0),
+                v_cache.stride(1),
+                v_cache.stride(2),
+                v_cache.stride(
+                    3),  #[num_blocks, num_kv_heads, head_size, block_size]
+                num_queries_per_kv=num_queries_per_kv,
+                BLOCK_M=BLOCK,
+                BLOCK_DMODEL=Lk,
+                BLOCK_DMODEL_PADDED=Lk_padded,
+                BLOCK_N=BLOCK,
+                num_warps=NUM_WARPS,
+                num_stages=1,
+            )
+            return
+
+        _fwd_kernel[grid](
+            q,
+            k,
+            v,
+            k_cache,
+            v_cache,
+            b_loc,
+            sm_scale,
+            k_scale,
+            v_scale,
+            b_start_loc,
+            b_seq_len,
+            b_ctx_len,
+            v_cache.shape[3],
+            k_cache.shape[4],
+            o,
+            b_loc.stride(0),
+            b_loc.stride(1),
+            q.stride(0),
+            q.stride(1),
+            q.stride(2),
+            k.stride(0),
+            k.stride(1),
+            k.stride(2),
+            v.stride(0),
+            v.stride(1),
+            v.stride(2),
+            o.stride(0),
+            o.stride(1),
+            o.stride(2),
+            k_cache.stride(0),
+            k_cache.stride(1),
+            k_cache.stride(2),
+            k_cache.stride(3),
+            k_cache.stride(
+                4),  #[num_blocks, num_kv_heads, head_size/x, block_size, x]
+            v_cache.stride(0),
+            v_cache.stride(1),
+            v_cache.stride(2),
+            v_cache.stride(
+                3),  #[num_blocks, num_kv_heads, head_size, block_size]
+            num_queries_per_kv=num_queries_per_kv,
+            BLOCK_M=BLOCK,
+            BLOCK_DMODEL=Lk,
+            BLOCK_DMODEL_PADDED=Lk_padded,
+            BLOCK_N=BLOCK,
+            SLIDING_WINDOW=sliding_window,
+            num_warps=NUM_WARPS,
+            num_stages=1,
+        )
+        return

vllm-v0.6.2/vllm/attention/ops/triton_flash_attention.py

@@ -0,0 +1,820 @@
+#!/usr/bin/env python
+"""
+Fused Attention
+===============
+
+This is a Triton implementation of the Flash Attention v2 algorithm from Tri Dao
+(https://tridao.me/publications/flash2/flash2.pdf)
+Credits: OpenAI kernel team, AMD ML Frameworks Triton team
+
+Features supported:
+
+1) Fwd with causal masking
+2) Any sequence lengths without padding (currently fwd kernel only)
+3) Support for different sequence lengths for q and k
+4) Nested tensor API currently does not support dropout or bias.
+
+Not currently supported:
+
+1) Non power of two head dims
+
+"""
+
+import torch
+import triton
+import triton.language as tl
+
+torch_dtype: tl.constexpr = torch.float16
+
+
+@triton.jit
+def cdiv_fn(x, y):
+    return (x + y - 1) // y
+
+
+@triton.jit
+def max_fn(x, y):
+    return tl.math.max(x, y)
+
+
+@triton.jit
+def dropout_offsets(philox_seed, philox_offset, dropout_p, m, n, stride):
+    ms = tl.arange(0, m)
+    ns = tl.arange(0, n)
+    return philox_offset + ms[:, None] * stride + ns[None, :]
+
+
+@triton.jit
+def dropout_rng(philox_seed, philox_offset, dropout_p, m, n, stride):
+    rng_offsets = dropout_offsets(philox_seed, philox_offset, dropout_p, m, n,
+                                  stride).to(tl.uint32)
+    # TODO: use tl.randint for better performance
+    return tl.rand(philox_seed, rng_offsets)
+
+
+@triton.jit
+def dropout_mask(philox_seed, philox_offset, dropout_p, m, n, stride):
+    rng_output = dropout_rng(philox_seed, philox_offset, dropout_p, m, n,
+                             stride)
+    rng_keep = rng_output > dropout_p
+    return rng_keep
+
+
+@triton.jit
+def load_fn(block_ptr, first, second, pad):
+    if first and second:
+        tensor = tl.load(block_ptr, boundary_check=(0, 1), padding_option=pad)
+    elif first:
+        tensor = tl.load(block_ptr, boundary_check=(0, ), padding_option=pad)
+    elif second:
+        tensor = tl.load(block_ptr, boundary_check=(1, ), padding_option=pad)
+    else:
+        tensor = tl.load(block_ptr)
+    return tensor
+
+
+@triton.jit
+def _attn_fwd_inner(
+    acc,
+    l_i,
+    m_i,
+    q,
+    K_block_ptr,
+    V_block_ptr,
+    start_m,
+    actual_seqlen_k,
+    dropout_p,
+    philox_seed,
+    batch_philox_offset,
+    encoded_softmax_block_ptr,
+    block_min,
+    block_max,
+    offs_n_causal,
+    masked_blocks,
+    n_extra_tokens,
+    bias_ptr,
+    IS_CAUSAL: tl.constexpr,
+    BLOCK_M: tl.constexpr,
+    BLOCK_DMODEL: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    OFFS_M: tl.constexpr,
+    OFFS_N: tl.constexpr,
+    PRE_LOAD_V: tl.constexpr,
+    MASK_STEPS: tl.constexpr,
+    ENABLE_DROPOUT: tl.constexpr,
+    RETURN_ENCODED_SOFTMAX: tl.constexpr,
+    PADDED_HEAD: tl.constexpr,
+):
+    # loop over k, v, and update accumulator
+    for start_n in range(block_min, block_max, BLOCK_N):
+        # For padded blocks, we will overrun the tensor size if
+        # we load all BLOCK_N. For others, the blocks are all within range.
+        k = load_fn(
+            K_block_ptr,
+            PADDED_HEAD,
+            MASK_STEPS and (n_extra_tokens != 0),
+            "zero",
+        )
+        if PRE_LOAD_V:
+            v = load_fn(
+                V_block_ptr,
+                MASK_STEPS and (n_extra_tokens != 0),
+                PADDED_HEAD,
+                "zero",
+            )
+        qk = tl.zeros([BLOCK_M, BLOCK_N], dtype=tl.float32)
+        # We start from end of seqlen_k so only the first iteration would need
+        # to be checked for padding if it is not a multiple of block_n
+        # TODO: This can be optimized to only be true for the padded block.
+        if MASK_STEPS:  # noqa: SIM102
+            # If this is the last block / iteration, we want to
+            # mask if the sequence length is not a multiple of block size
+            # a solution is to always do BLOCK_M // BLOCK_N + 1 steps
+            # if not is_modulo_mn. last step might get wasted but that is okay.
+            # check if this masking works for that case.
+            if (start_n + BLOCK_N == block_max) and (n_extra_tokens != 0):
+                boundary_m = tl.full([BLOCK_M],
+                                     actual_seqlen_k,
+                                     dtype=tl.int32)
+                size_n = start_n + OFFS_N[None, :]
+                mask = size_n < boundary_m[:, None]
+                qk = tl.where(mask, qk, float("-inf"))
+        if IS_CAUSAL:
+            causal_boundary = start_n + offs_n_causal
+            causal_mask = OFFS_M[:, None] >= causal_boundary[None, :]
+            qk = tl.where(causal_mask, qk, float("-inf"))
+        # -- compute qk ----
+        qk += tl.dot(q, k)
+        if bias_ptr is not None:
+            bias = load_fn(bias_ptr, False, MASK_STEPS
+                           and (n_extra_tokens != 0), "zero")
+            # While bias is added after multiplying qk with sm_scale, our
+            # optimization to use 2^x instead of e^x results in an additional
+            # scale factor of log2(e) which we must also multiply the bias with.
+            qk += bias * 1.44269504089
+        m_ij = tl.maximum(m_i, tl.max(qk, 1))
+        qk = qk - m_ij[:, None]
+        p = tl.math.exp2(qk)
+
+        # CAVEAT: Must update l_ij before applying dropout
+        l_ij = tl.sum(p, 1)
+        if ENABLE_DROPOUT:
+            philox_offset = (batch_philox_offset +
+                             start_m * BLOCK_M * actual_seqlen_k + start_n -
+                             BLOCK_N)
+            keep = dropout_mask(
+                philox_seed,
+                philox_offset,
+                dropout_p,
+                BLOCK_M,
+                BLOCK_N,
+                actual_seqlen_k,
+            )
+            if RETURN_ENCODED_SOFTMAX:
+                tl.store(
+                    encoded_softmax_block_ptr,
+                    tl.where(keep, p,
+                             -p).to(encoded_softmax_block_ptr.type.element_ty),
+                )
+            p = tl.where(keep, p, 0.0)
+        elif RETURN_ENCODED_SOFTMAX:
+            tl.store(
+                encoded_softmax_block_ptr,
+                p.to(encoded_softmax_block_ptr.type.element_ty),
+            )
+        # -- update output accumulator --
+        alpha = tl.math.exp2(m_i - m_ij)
+        acc = acc * alpha[:, None]
+        if not PRE_LOAD_V:
+            v = load_fn(
+                V_block_ptr,
+                MASK_STEPS and (n_extra_tokens != 0),
+                PADDED_HEAD,
+                "zero",
+            )
+        # -- update m_i and l_i
+        l_i = l_i * alpha + l_ij
+        # update m_i and l_i
+        m_i = m_ij
+        acc += tl.dot(p.to(V_block_ptr.type.element_ty), v)
+        V_block_ptr = tl.advance(V_block_ptr, (BLOCK_N, 0))
+        K_block_ptr = tl.advance(K_block_ptr, (0, BLOCK_N))
+        if bias_ptr is not None:
+            bias_ptr = tl.advance(bias_ptr, (0, BLOCK_N))
+        if RETURN_ENCODED_SOFTMAX:
+            encoded_softmax_block_ptr = tl.advance(encoded_softmax_block_ptr,
+                                                   (0, BLOCK_N))
+    return acc, l_i, m_i
+
+
+@triton.autotune(
+    configs=[
+        triton.Config(
+            {
+                "BLOCK_M": 256,
+                "BLOCK_N": 64,
+                "waves_per_eu": 2,
+                "PRE_LOAD_V": False,
+            },
+            num_stages=1,
+            num_warps=8,
+        ),
+        triton.Config(
+            {
+                "BLOCK_M": 128,
+                "BLOCK_N": 128,
+                "waves_per_eu": 2,
+                "PRE_LOAD_V": False,
+            },
+            num_stages=1,
+            num_warps=4,
+        ),
+        triton.Config(
+            {
+                "BLOCK_M": 256,
+                "BLOCK_N": 128,
+                "waves_per_eu": 2,
+                "PRE_LOAD_V": False,
+            },
+            num_stages=1,
+            num_warps=8,
+        ),
+        triton.Config(
+            {
+                "BLOCK_M": 128,
+                "BLOCK_N": 64,
+                "waves_per_eu": 1,
+                "PRE_LOAD_V": False,
+            },
+            num_stages=1,
+            num_warps=4,
+        ),
+        triton.Config(
+            {
+                "BLOCK_M": 128,
+                "BLOCK_N": 64,
+                "waves_per_eu": 3,
+                "PRE_LOAD_V": True,
+            },
+            num_stages=1,
+            num_warps=4,
+        ),
+        triton.Config(
+            {
+                "BLOCK_M": 128,
+                "BLOCK_N": 64,
+                "waves_per_eu": 3,
+                "PRE_LOAD_V": False,
+            },
+            num_stages=1,
+            num_warps=4,
+        ),
+        triton.Config(
+            {
+                "BLOCK_M": 64,
+                "BLOCK_N": 64,
+                "waves_per_eu": 4,
+                "PRE_LOAD_V": False,
+            },
+            num_stages=1,
+            num_warps=8,
+        ),
+        triton.Config(
+            {
+                "BLOCK_M": 32,
+                "BLOCK_N": 32,
+                "waves_per_eu": 4,
+                "PRE_LOAD_V": False,
+            },
+            num_stages=1,
+            num_warps=8,
+        ),
+        # TODO: This config fails with head_size not pow2 with data mismatches.
+        #    triton.Config({'BLOCK_M': 32, 'BLOCK_N': 16, 'waves_per_eu': 1,
+        #                   'PRE_LOAD_V': False}, num_stages=1, num_warps=4),
+        triton.Config(
+            {
+                "BLOCK_M": 16,
+                "BLOCK_N": 16,
+                "waves_per_eu": 1,
+                "PRE_LOAD_V": False,
+            },
+            num_stages=1,
+            num_warps=4,
+        ),
+    ],
+    key=['IS_CAUSAL', 'dropout_p', 'BLOCK_DMODEL'],
+)
+@triton.jit
+def attn_fwd(
+    Q,
+    K,
+    V,
+    bias,
+    sm_scale,
+    L,
+    Out,
+    stride_qz,
+    stride_qh,
+    stride_qm,
+    stride_qk,
+    stride_kz,
+    stride_kh,
+    stride_kn,
+    stride_kk,
+    stride_vz,
+    stride_vh,
+    stride_vk,
+    stride_vn,
+    stride_oz,
+    stride_oh,
+    stride_om,
+    stride_on,
+    stride_bz,
+    stride_bh,
+    stride_bm,
+    stride_bn,
+    cu_seqlens_q,
+    cu_seqlens_k,
+    dropout_p,
+    philox_seed,
+    philox_offset_base,
+    encoded_softmax,
+    HQ: tl.constexpr,
+    HK: tl.constexpr,
+    ACTUAL_BLOCK_DMODEL: tl.constexpr,
+    MAX_SEQLENS_Q: tl.constexpr,
+    MAX_SEQLENS_K: tl.constexpr,
+    VARLEN: tl.constexpr,
+    IS_CAUSAL: tl.constexpr,
+    BLOCK_M: tl.constexpr,
+    BLOCK_DMODEL: tl.constexpr,
+    BLOCK_N: tl.constexpr,
+    PRE_LOAD_V: tl.constexpr,
+    BIAS_TYPE: tl.constexpr,
+    ENABLE_DROPOUT: tl.constexpr,
+    RETURN_ENCODED_SOFTMAX: tl.constexpr,
+):
+    start_m = tl.program_id(0)
+    off_h_q = tl.program_id(1)
+    off_z = tl.program_id(2)
+    offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
+    offs_n = tl.arange(0, BLOCK_N)
+    if VARLEN:
+        cu_seqlens_q_start = tl.load(cu_seqlens_q + off_z)
+        cu_seqlens_q_end = tl.load(cu_seqlens_q + off_z + 1)
+        seqlen_q = cu_seqlens_q_end - cu_seqlens_q_start
+        # We have a one-size-fits-all grid in id(0). Some seqlens might be too
+        # small for all start_m so for those we return early.
+        if start_m * BLOCK_M > seqlen_q:
+            return
+        cu_seqlens_k_start = tl.load(cu_seqlens_k + off_z)
+        cu_seqlens_k_end = tl.load(cu_seqlens_k + off_z + 1)
+        seqlen_k = cu_seqlens_k_end - cu_seqlens_k_start
+    else:
+        cu_seqlens_q_start = 0
+        cu_seqlens_k_start = 0
+        seqlen_q = MAX_SEQLENS_Q
+        seqlen_k = MAX_SEQLENS_K
+
+    # Now we compute whether we need to exit early due to causal masking.
+    # This is because for seqlen_q > seqlen_k, M rows of the attn scores
+    # are completely masked, resulting in 0s written to the output, and
+    # inf written to LSE. We don't need to do any GEMMs in this case.
+    # This block of code determines what N is, and if this WG is operating
+    # on those M rows.
+    n_blocks = cdiv_fn(seqlen_k, BLOCK_N)
+    if IS_CAUSAL:
+        # If seqlen_q == seqlen_k, the attn scores are a square matrix.
+        # If seqlen_q != seqlen_k, attn scores are rectangular which means
+        # the causal mask boundary is bottom right aligned, and ends at either
+        # the top edge (seqlen_q < seqlen_k) or left edge.
+        # This captures the decrease in n_blocks if we have a rectangular attn
+        # matrix
+        n_blocks_seqlen = cdiv_fn(
+            (start_m + 1) * BLOCK_M + seqlen_k - seqlen_q, BLOCK_N)
+        # This is what adjusts the block_max for the current WG, only
+        # if IS_CAUSAL. Otherwise we want to always iterate through all n_blocks
+        n_blocks = min(n_blocks, n_blocks_seqlen)
+        # If we have no blocks after adjusting for seqlen deltas, this WG is
+        # part of the blocks that are all 0. We exit early.
+        if n_blocks <= 0:
+            o_offset = (off_z * stride_oz + cu_seqlens_q_start * stride_om +
+                        off_h_q * stride_oh)
+            O_block_ptr = tl.make_block_ptr(
+                base=Out + o_offset,
+                shape=(seqlen_q, BLOCK_DMODEL),
+                strides=(stride_om, stride_on),
+                offsets=(start_m * BLOCK_M, 0),
+                block_shape=(BLOCK_M, BLOCK_DMODEL),
+                order=(1, 0),
+            )
+            acc = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=Out.type.element_ty)
+            # We still need to write 0s to the result
+            # tl.store(O_block_ptr,
+            # acc.to(Out.type.element_ty), boundary_check=(0,1))
+            # l_ptrs = L + off_z * HQ * MAX_SEQLENS_Q + off_h_q * MAX_SEQLENS_Q
+            #          + offs_m
+            # We store inf to LSE, not -inf because in the bwd pass,
+            # we subtract this
+            # from qk which makes it -inf, such that exp(qk - inf) = 0
+            # for these masked blocks.
+            # l = tl.full([BLOCK_M], value=float("inf"), dtype=tl.float32)
+            # tl.store(l_ptrs, l)
+            # TODO: Should dropout and return encoded softmax be handled here?
+            return
+
+    # If MQA / GQA, set the K and V head offsets appropriately.
+    GROUP_SIZE: tl.constexpr = HQ // HK
+    off_h_k = off_h_q // GROUP_SIZE if GROUP_SIZE != 1 else off_h_q
+
+    n_extra_tokens = 0
+    if seqlen_k < BLOCK_N:
+        n_extra_tokens = BLOCK_N - seqlen_k
+    elif seqlen_k % BLOCK_N:
+        n_extra_tokens = seqlen_k % BLOCK_N
+    padded_head = ACTUAL_BLOCK_DMODEL != BLOCK_DMODEL
+
+    # Compute pointers for all the tensors used in this kernel.
+    q_offset = (off_z * stride_qz + off_h_q * stride_qh +
+                cu_seqlens_q_start * stride_qm)
+    Q_block_ptr = tl.make_block_ptr(
+        base=Q + q_offset,
+        shape=(seqlen_q, ACTUAL_BLOCK_DMODEL),
+        strides=(stride_qm, stride_qk),
+        offsets=(start_m * BLOCK_M, 0),
+        block_shape=(BLOCK_M, BLOCK_DMODEL),
+        order=(1, 0),
+    )
+    k_offset = (off_z * stride_kz + off_h_k * stride_kh +
+                cu_seqlens_k_start * stride_kn)
+    K_block_ptr = tl.make_block_ptr(
+        base=K + k_offset,
+        shape=(ACTUAL_BLOCK_DMODEL, seqlen_k),
+        strides=(stride_kk, stride_kn),
+        offsets=(0, 0),
+        block_shape=(BLOCK_DMODEL, BLOCK_N),
+        order=(0, 1),
+    )
+    v_offset = (off_z * stride_vz + off_h_k * stride_vh +
+                cu_seqlens_k_start * stride_vk)
+    V_block_ptr = tl.make_block_ptr(
+        base=V + v_offset,
+        shape=(seqlen_k, ACTUAL_BLOCK_DMODEL),
+        strides=(stride_vk, stride_vn),
+        offsets=(0, 0),
+        block_shape=(BLOCK_N, BLOCK_DMODEL),
+        order=(1, 0),
+    )
+    if BIAS_TYPE != 0:
+        bias_ptr = tl.make_block_ptr(
+            base=bias + off_h_q * stride_bh,
+            shape=(seqlen_q, seqlen_k),
+            strides=(stride_bm, stride_bn),
+            offsets=(start_m * BLOCK_M, 0),
+            block_shape=(BLOCK_M, BLOCK_N),
+            order=(1, 0),
+        )
+    else:
+        bias_ptr = None
+    if ENABLE_DROPOUT:
+        batch_philox_offset = philox_offset_base \
+                              + (off_z * HQ + off_h_q) \
+                              * seqlen_q * seqlen_k
+    else:
+        batch_philox_offset = 0
+    # We can ask to return the dropout mask without actually doing any dropout.
+    # In this case, we return an invalid pointer so indicate the mask is not i
+    # valid.
+    # TODO: Fix encoded softmax. It currently uses just h_q in the base offset.
+    if RETURN_ENCODED_SOFTMAX:
+        encoded_softmax_block_ptr = tl.make_block_ptr(
+            base=encoded_softmax + off_h_q * seqlen_q * seqlen_k,
+            shape=(seqlen_q, seqlen_k),
+            strides=(seqlen_k, 1),
+            offsets=(start_m * BLOCK_M, 0),
+            block_shape=(BLOCK_M, BLOCK_N),
+            order=(1, 0),
+        )
+    else:
+        encoded_softmax_block_ptr = 0
+    # initialize pointer to m and l
+    m_i = tl.full([BLOCK_M], float("-inf"), dtype=tl.float32)
+    l_i = tl.full([BLOCK_M], 1.0, dtype=tl.float32)
+    acc = tl.zeros([BLOCK_M, BLOCK_DMODEL], dtype=tl.float32)
+    # scale sm_scale by log_2(e) and use 2^x in the loop as we do not
+    # have native e^x support in HW.
+    qk_scale = sm_scale * 1.44269504089
+    # Q is loaded once at the beginning and shared by all N blocks.
+    q = load_fn(Q_block_ptr, True, padded_head, "zero")
+    q = (q * qk_scale).to(Q_block_ptr.type.element_ty)
+
+    # Here we compute how many full and masked blocks we have.
+    padded_block_k = n_extra_tokens != 0
+    is_modulo_mn = not padded_block_k and (seqlen_q % BLOCK_M == 0)
+    if IS_CAUSAL:
+        # There are always at least BLOCK_M // BLOCK_N masked blocks.
+        # Additionally there might be one more due to dissimilar seqlens.
+        masked_blocks = BLOCK_M // BLOCK_N + (not is_modulo_mn)
+    else:
+        # Padding on Q does not need to be masked in the FA loop.
+        masked_blocks = padded_block_k
+    # if IS_CAUSAL, not is_modulo_mn does not always result in an additional
+    # block. In this case we might exceed n_blocks so pick the min.
+    masked_blocks = min(masked_blocks, n_blocks)
+    n_full_blocks = n_blocks - masked_blocks
+    block_min = 0
+    block_max = n_blocks * BLOCK_N
+    # Compute for full blocks. Here we set causal to false regardless of its
+    # value because there is no masking. Similarly we do not need padding.
+    if n_full_blocks > 0:
+        block_max = (n_blocks - masked_blocks) * BLOCK_N
+        acc, l_i, m_i = _attn_fwd_inner(
+            acc,
+            l_i,
+            m_i,
+            q,
+            K_block_ptr,
+            V_block_ptr,
+            start_m,
+            seqlen_k,
+            dropout_p,
+            philox_seed,
+            batch_philox_offset,
+            encoded_softmax_block_ptr,
+            # _, _, offs_n_causal, masked_blocks, n_extra_tokens, _
+            block_min,
+            block_max,
+            0,
+            0,
+            0,
+            bias_ptr,
+            # IS_CAUSAL, ....
+            False,
+            BLOCK_M,
+            BLOCK_DMODEL,
+            BLOCK_N,
+            offs_m,
+            offs_n,
+            # _, MASK_STEPS, ...
+            PRE_LOAD_V,
+            False,
+            ENABLE_DROPOUT,
+            RETURN_ENCODED_SOFTMAX,
+            padded_head,
+        )
+        block_min = block_max
+        block_max = n_blocks * BLOCK_N
+
+    tl.debug_barrier()
+    # Remaining blocks, if any, are full / not masked.
+    if masked_blocks > 0:
+        offs_n_causal = offs_n + (seqlen_q - seqlen_k) if IS_CAUSAL else 0
+        K_block_ptr = tl.advance(K_block_ptr, (0, n_full_blocks * BLOCK_N))
+        V_block_ptr = tl.advance(V_block_ptr, (n_full_blocks * BLOCK_N, 0))
+        if bias_ptr is not None:
+            bias_ptr = tl.advance(bias_ptr, (0, n_full_blocks * BLOCK_N))
+        if RETURN_ENCODED_SOFTMAX:
+            encoded_softmax_block_ptr = tl.advance(encoded_softmax_block_ptr,
+                                                   (0, n_full_blocks))
+        acc, l_i, m_i = _attn_fwd_inner(
+            acc,
+            l_i,
+            m_i,
+            q,
+            K_block_ptr,
+            V_block_ptr,
+            start_m,
+            seqlen_k,
+            dropout_p,
+            philox_seed,
+            batch_philox_offset,
+            encoded_softmax_block_ptr,
+            block_min,
+            block_max,
+            offs_n_causal,
+            masked_blocks,
+            n_extra_tokens,
+            bias_ptr,
+            IS_CAUSAL,
+            BLOCK_M,
+            BLOCK_DMODEL,
+            BLOCK_N,
+            offs_m,
+            offs_n,
+            # _, MASK_STEPS, ...
+            PRE_LOAD_V,
+            True,
+            ENABLE_DROPOUT,
+            RETURN_ENCODED_SOFTMAX,
+            padded_head,
+        )
+    # epilogue
+    acc = acc / l_i[:, None]
+    if ENABLE_DROPOUT:
+        acc = acc / (1 - dropout_p)
+    # If seqlen_q > seqlen_k but the delta is not a multiple of BLOCK_M,
+    # then we have one block with a row of all NaNs which come from computing
+    # softmax over a row of all -infs (-inf - inf = NaN). We check for that here
+    # and store 0s where there are NaNs as these rows should've been zeroed out.
+    end_m_idx = (start_m + 1) * BLOCK_M
+    start_m_idx = start_m * BLOCK_M
+    causal_start_idx = seqlen_q - seqlen_k
+    acc = acc.to(Out.type.element_ty)
+    if IS_CAUSAL:  # noqa: SIM102
+        if causal_start_idx > start_m_idx and causal_start_idx < end_m_idx:
+            out_mask_boundary = tl.full((BLOCK_DMODEL, ),
+                                        causal_start_idx,
+                                        dtype=tl.int32)
+            mask_m_offsets = start_m_idx + tl.arange(0, BLOCK_M)
+            out_ptrs_mask = (mask_m_offsets[:, None] >=
+                             out_mask_boundary[None, :])
+            z = 0.0
+            acc = tl.where(out_ptrs_mask, acc, z.to(acc.type.element_ty))
+    # write back LSE
+    # l_ptrs = L + off_z * HQ * MAX_SEQLENS_Q + off_h_q * MAX_SEQLENS_Q + offs_m
+    # If seqlen_q not multiple of BLOCK_M, we need to mask out the last
+    # few rows. This is only true for the last M block. For others,
+    # overflow_size will be -ve
+    # overflow_size = end_m_idx - seqlen_q
+    # if overflow_size > 0:
+    #    boundary = tl.full((BLOCK_M,), BLOCK_M - overflow_size, dtype=tl.int32)
+    #    # This is a > check because mask being 0 blocks the store.
+    #    l_ptrs_mask = boundary > tl.arange(0, BLOCK_M)
+    #    tl.store(l_ptrs, m_i + tl.math.log2(l_i), mask=l_ptrs_mask)
+    # else:
+    #    tl.store(l_ptrs, m_i + tl.math.log2(l_i))
+
+    # write back O
+    o_offset = (off_z * stride_oz + cu_seqlens_q_start * stride_om +
+                off_h_q * stride_oh)
+    O_block_ptr = tl.make_block_ptr(
+        base=Out + o_offset,
+        shape=(seqlen_q, ACTUAL_BLOCK_DMODEL),
+        strides=(stride_om, stride_on),
+        offsets=(start_m * BLOCK_M, 0),
+        block_shape=(BLOCK_M, BLOCK_DMODEL),
+        order=(1, 0),
+    )
+    # Need boundary check on this to make sure the padding from the
+    # Q and KV tensors in both dims are not part of what we store back.
+    # TODO: Do the boundary check optionally.
+    tl.store(O_block_ptr, acc, boundary_check=(0, 1))
+
+
+def check_args(
+    q,
+    k,
+    v,
+    o,
+    varlen=True,
+    max_seqlens=None,
+    cu_seqlens_q=None,
+    cu_seqlens_k=None,
+):
+    assert q.dim() == k.dim() and q.dim() == v.dim()
+    if varlen:
+        assert q.dim() == 3
+        total_q, nheads_q, head_size = q.shape
+        total_k, nheads_k, _ = k.shape
+        assert cu_seqlens_q is not None
+        assert cu_seqlens_k is not None
+        assert len(cu_seqlens_q) == len(cu_seqlens_k)
+    else:
+        assert q.dim() == 4
+        batch, nheads_q, seqlen_q, head_size = q.shape
+        _, nheads_k, seqlen_k, _ = k.shape
+        assert max_seqlens > 0
+    assert k.shape == v.shape
+    assert q.shape[-1] == k.shape[-1] and q.shape[-1] == v.shape[-1]
+    # TODO: Change assert if we support qkl f8 and v f16
+    assert q.dtype == k.dtype and q.dtype == v.dtype
+    assert head_size <= 256
+    assert o.shape == q.shape
+    assert (nheads_q % nheads_k) == 0
+
+
+class _attention(torch.autograd.Function):
+
+    @staticmethod
+    def forward(
+        ctx,
+        q,
+        k,
+        v,
+        o,
+        cu_seqlens_q,
+        cu_seqlens_k,
+        max_seqlens_q,
+        max_seqlens_k,
+        causal=False,
+        sm_scale=1.0,
+        bias=None,
+    ):
+        if o is None:
+            o = torch.empty_like(q, dtype=v.dtype)
+
+        check_args(
+            q,
+            k,
+            v,
+            o,
+            varlen=True,
+            cu_seqlens_q=cu_seqlens_q,
+            cu_seqlens_k=cu_seqlens_k,
+        )
+        if True:  # varlen
+            total_q, nheads_q, head_size = q.shape
+            total_k, nheads_k, _ = k.shape
+            batch = len(cu_seqlens_q) - 1
+            q_strides = (0, q.stride(1), q.stride(0), q.stride(2))
+            k_strides = (0, k.stride(1), k.stride(0), k.stride(2))
+            v_strides = (0, v.stride(1), v.stride(0), v.stride(2))
+            o_strides = (0, o.stride(1), o.stride(0), o.stride(2))
+        else:
+            batch, seqlen_q, nheads_q, head_size = q.shape
+            _, seqlen_k, nheads_k, _ = k.shape
+            q_strides = (q.stride(0), q.stride(2), q.stride(1), q.stride(3))
+            k_strides = (k.stride(0), k.stride(2), k.stride(1), k.stride(3))
+            v_strides = (v.stride(0), v.stride(2), v.stride(1), v.stride(3))
+            o_strides = (o.stride(0), o.stride(2), o.stride(1), o.stride(3))
+
+        # Get closest power of 2 over or equal to 32.
+        unpadded_head_dims = {32, 64, 128, 256}
+        if head_size not in unpadded_head_dims:
+            padded_d_model = None
+            for i in unpadded_head_dims:
+                if i > head_size:
+                    padded_d_model = i
+                    break
+            assert padded_d_model is not None
+        else:
+            padded_d_model = head_size
+
+        grid = lambda META: (
+            triton.cdiv(max_seqlens_q, META["BLOCK_M"]),
+            nheads_q,
+            batch,
+        )
+
+        encoded_softmax = None
+
+        # Seed the RNG so we get reproducible results for testing.
+        philox_seed = 0x1BF52
+        philox_offset = 0x1D4B42
+
+        if bias is not None:
+            bias_strides = (
+                bias.stride(0),
+                bias.stride(1),
+                bias.stride(2),
+                bias.stride(3),
+            )
+        else:
+            bias_strides = (0, 0, 0, 0)
+
+        attn_fwd[grid](
+            q,
+            k,
+            v,
+            bias,
+            sm_scale,
+            None,
+            o,
+            *q_strides,
+            *k_strides,
+            *v_strides,
+            *o_strides,
+            *bias_strides,
+            cu_seqlens_q,
+            cu_seqlens_k,
+            dropout_p=0.0,
+            philox_seed=philox_seed,
+            philox_offset_base=philox_offset,
+            encoded_softmax=encoded_softmax,
+            HQ=nheads_q,
+            HK=nheads_k,
+            ACTUAL_BLOCK_DMODEL=head_size,
+            MAX_SEQLENS_Q=max_seqlens_q,
+            MAX_SEQLENS_K=max_seqlens_k,
+            IS_CAUSAL=causal,
+            VARLEN=True,
+            BLOCK_DMODEL=padded_d_model,
+            BIAS_TYPE=0 if bias is None else 1,
+            ENABLE_DROPOUT=False,
+            RETURN_ENCODED_SOFTMAX=False,
+        )
+
+        ctx.grid = grid
+        ctx.sm_scale = sm_scale
+        ctx.BLOCK_DMODEL = head_size
+        ctx.causal = causal
+        ctx.dropout_p = 0.0
+        ctx.philox_seed = philox_seed
+        ctx.philox_offset = philox_offset
+        ctx.encoded_softmax = encoded_softmax
+        ctx.return_encoded_softmax = False
+        return o, encoded_softmax
+
+
+triton_attention = _attention.apply

vllm-v0.6.2/vllm/attention/selector.py

@@ -0,0 +1,347 @@
+import enum
+import os
+from contextlib import contextmanager
+from functools import lru_cache
+from typing import Generator, Optional, Type
+
+import torch
+
+import vllm.envs as envs
+from vllm.attention.backends.abstract import AttentionBackend
+from vllm.logger import init_logger
+from vllm.platforms import current_platform
+from vllm.utils import STR_BACKEND_ENV_VAR
+
+logger = init_logger(__name__)
+
+
+class _Backend(enum.Enum):
+    FLASH_ATTN = enum.auto()
+    FLASH_ATTN_VLLM_V1 = enum.auto()
+    XFORMERS = enum.auto()
+    ROCM_FLASH = enum.auto()
+    TORCH_SDPA = enum.auto()
+    OPENVINO = enum.auto()
+    FLASHINFER = enum.auto()
+    HPU_ATTN = enum.auto()
+    PALLAS = enum.auto()
+    IPEX = enum.auto()
+    MLU_FLASH_ATTN = enum.auto()
+    NO_ATTENTION = enum.auto()
+
+
+def backend_name_to_enum(backend_name: str) -> _Backend:
+    assert backend_name is not None
+
+    backend_members = _Backend.__members__
+    if backend_name not in backend_members:
+        raise ValueError(f"Invalid attention backend '{backend_name}'. "
+                         f"Available backends: {', '.join(backend_members)} "
+                         "(case-sensitive).")
+
+    return _Backend[backend_name]
+
+
+def get_env_variable_attn_backend() -> Optional[_Backend]:
+    '''
+    Get the backend override specified by the vLLM attention
+    backend environment variable, if one is specified.
+
+    Returns:
+
+    * _Backend enum value if an override is specified
+    * None otherwise
+    '''
+    backend_name = os.environ.get(STR_BACKEND_ENV_VAR)
+    return (None
+            if backend_name is None else backend_name_to_enum(backend_name))
+
+
+# Global state allows a particular choice of backend
+# to be forced, overriding the logic which auto-selects
+# a backend based on system & workload configuration
+# (default behavior if this variable is None)
+#
+# THIS SELECTION TAKES PRECEDENCE OVER THE
+# VLLM ATTENTION BACKEND ENVIRONMENT VARIABLE
+forced_attn_backend: Optional[_Backend] = None
+
+
+def global_force_attn_backend(attn_backend: Optional[_Backend]) -> None:
+    '''
+    Force all attention operations to use a specified backend.
+
+    Passing `None` for the argument re-enables automatic
+    backend selection.,
+
+    Arguments:
+
+    * attn_backend: backend selection (None to revert to auto)
+    '''
+    global forced_attn_backend
+    forced_attn_backend = attn_backend
+
+
+def get_global_forced_attn_backend() -> Optional[_Backend]:
+    '''
+    Get the currently-forced choice of attention backend,
+    or None if auto-selection is currently enabled.
+    '''
+    return forced_attn_backend
+
+
+def get_attn_backend(
+    head_size: int,
+    dtype: torch.dtype,
+    kv_cache_dtype: Optional[str],
+    block_size: int,
+    is_attention_free: bool,
+    is_blocksparse: bool = False,
+) -> Type[AttentionBackend]:
+    """Selects which attention backend to use and lazily imports it."""
+    # Accessing envs.* behind an @lru_cache decorator can cause the wrong
+    # value to be returned from the cache if the value changes between calls.
+    # To avoid this, we read envs.VLLM_USE_V1 here and pass it explicitly to the
+    # private function.
+    return _cached_get_attn_backend(
+        head_size=head_size,
+        dtype=dtype,
+        kv_cache_dtype=kv_cache_dtype,
+        block_size=block_size,
+        is_attention_free=is_attention_free,
+        is_blocksparse=is_blocksparse,
+        use_v1=envs.VLLM_USE_V1,
+    )
+
+
+@lru_cache(maxsize=None)
+def _cached_get_attn_backend(
+    head_size: int,
+    dtype: torch.dtype,
+    kv_cache_dtype: Optional[str],
+    block_size: int,
+    is_attention_free: bool,
+    is_blocksparse: bool = False,
+    use_v1: bool = False,
+) -> Type[AttentionBackend]:
+    if is_blocksparse:
+        logger.info("Using BlocksparseFlashAttention backend.")
+        from vllm.attention.backends.blocksparse_attn import (
+            BlocksparseFlashAttentionBackend)
+        return BlocksparseFlashAttentionBackend
+
+    backend = which_attn_to_use(head_size, dtype, kv_cache_dtype, block_size,
+                                is_attention_free, use_v1)
+    if backend == _Backend.FLASH_ATTN:
+        logger.info("Using Flash Attention backend.")
+        from vllm.attention.backends.flash_attn import (  # noqa: F401
+            FlashAttentionBackend)
+        return FlashAttentionBackend
+    if backend == _Backend.FLASH_ATTN_VLLM_V1:
+        from vllm.v1.attention.backends.flash_attn import (  # noqa: F401
+            FlashAttentionBackend as FlashAttentionBackendV1)
+        return FlashAttentionBackendV1
+    if backend == _Backend.XFORMERS:
+        logger.info("Using XFormers backend.")
+        from vllm.attention.backends.xformers import (  # noqa: F401
+            XFormersBackend)
+        return XFormersBackend
+    elif backend == _Backend.ROCM_FLASH:
+        logger.info("Using ROCmFlashAttention backend.")
+        from vllm.attention.backends.rocm_flash_attn import (  # noqa: F401
+            ROCmFlashAttentionBackend)
+        return ROCmFlashAttentionBackend
+    elif backend == _Backend.TORCH_SDPA:
+        assert current_platform.is_cpu(), RuntimeError(
+            "Torch SDPA backend is only used for the CPU device.")
+        logger.info("Using Torch SDPA backend.")
+        from vllm.attention.backends.torch_sdpa import TorchSDPABackend
+        return TorchSDPABackend
+    elif backend == _Backend.OPENVINO:
+        logger.info("Using OpenVINO Attention backend.")
+        from vllm.attention.backends.openvino import OpenVINOAttentionBackend
+        return OpenVINOAttentionBackend
+    elif backend == _Backend.IPEX:
+        assert current_platform.is_xpu(), RuntimeError(
+            "IPEX attention backend is only used for the XPU device.")
+        logger.info("Using IPEX attention backend.")
+        from vllm.attention.backends.ipex_attn import IpexAttnBackend
+        return IpexAttnBackend
+    elif backend == _Backend.FLASHINFER:
+        logger.info("Using Flashinfer backend.")
+        from vllm.attention.backends.flashinfer import FlashInferBackend
+        return FlashInferBackend
+    elif backend == _Backend.HPU_ATTN:
+        logger.info("Using HPUAttention backend.")
+        from vllm.attention.backends.hpu_attn import HPUAttentionBackend
+        return HPUAttentionBackend
+    elif backend == _Backend.PALLAS:
+        logger.info("Using Pallas backend.")
+        from vllm.attention.backends.pallas import PallasAttentionBackend
+        return PallasAttentionBackend
+    elif backend == _Backend.MLU_FLASH_ATTN:
+        logger.info("Using MLUFlashAttention backend.")
+        from vllm.attention.backends.mlu_attn import MLUFlashAttentionBackend
+        return MLUFlashAttentionBackend
+    elif backend == _Backend.NO_ATTENTION:
+        from vllm.attention.backends.placeholder_attn import (
+            PlaceholderAttentionBackend)
+        return PlaceholderAttentionBackend
+    else:
+        raise ValueError("Invalid attention backend.")
+
+
+def which_attn_to_use(head_size: int,
+                      dtype: torch.dtype,
+                      kv_cache_dtype: Optional[str],
+                      block_size: int,
+                      is_attention_free: bool,
+                      use_v1: bool = False) -> _Backend:
+    """Returns which flash attention backend to use."""
+    # Default case.
+    selected_backend = _Backend.FLASH_ATTN
+
+    # If there are no attention layers (e.g. we are running Mamba),
+    # use the placeholder NO_ATTENTION
+    if is_attention_free:
+        return _Backend.NO_ATTENTION
+
+    # Check whether a particular choice of backend was
+    # previously forced.
+    #
+    # THIS SELECTION OVERRIDES THE VLLM_ATTENTION_BACKEND
+    # ENVIRONMENT VARIABLE.
+    backend_by_global_setting: Optional[_Backend] = (
+        get_global_forced_attn_backend())
+    if backend_by_global_setting is not None:
+        selected_backend = backend_by_global_setting
+    else:
+        # Check the environment variable and override if specified
+        backend_by_env_var: Optional[str] = envs.VLLM_ATTENTION_BACKEND
+        if backend_by_env_var is not None:
+            selected_backend = backend_name_to_enum(backend_by_env_var)
+
+    if current_platform.is_cpu():
+        if selected_backend != _Backend.TORCH_SDPA:
+            logger.info("Cannot use %s backend on CPU.", selected_backend)
+        return _Backend.TORCH_SDPA
+
+    if current_platform.is_openvino():
+        if selected_backend != _Backend.OPENVINO:
+            logger.info("Cannot use %s backend on OpenVINO.", selected_backend)
+        return _Backend.OPENVINO
+
+    if current_platform.is_xpu():
+        if selected_backend != _Backend.IPEX:
+            logger.info("Cannot use %s backend on XPU.", selected_backend)
+        return _Backend.IPEX
+
+    if current_platform.is_tpu():
+        if selected_backend != _Backend.PALLAS:
+            logger.info("Cannot use %s backend on TPU.", selected_backend)
+        return _Backend.PALLAS
+    
+    if current_platform.is_mlu():
+        if selected_backend != _Backend.MLU_FLASH_ATTN:
+            logger.debug("Cannot use %s backend on MLU.", selected_backend)
+        return _Backend.MLU_FLASH_ATTN
+
+    if current_platform.is_rocm():
+        # AMD GPUs.
+        selected_backend = (_Backend.ROCM_FLASH if selected_backend
+                            == _Backend.FLASH_ATTN else selected_backend)
+        if selected_backend == _Backend.ROCM_FLASH:
+            if not current_platform.has_device_capability(90):
+                # not Instinct series GPUs.
+                logger.info("flash_attn is not supported on NAVI GPUs.")
+        else:
+            logger.info("%s is not supported in AMD GPUs.", selected_backend)
+        return _Backend.ROCM_FLASH
+
+    if current_platform.is_hpu():
+        return _Backend.HPU_ATTN
+
+    if use_v1:
+        return _Backend.FLASH_ATTN_VLLM_V1
+
+    # FlashAttn in NVIDIA GPUs.
+    if selected_backend == _Backend.FLASH_ATTN:
+        if not current_platform.has_device_capability(80):
+            # Volta and Turing NVIDIA GPUs.
+            logger.info(
+                "Cannot use FlashAttention-2 backend for Volta and Turing "
+                "GPUs.")
+            selected_backend = _Backend.XFORMERS
+        elif dtype not in (torch.float16, torch.bfloat16):
+            logger.info(
+                "Cannot use FlashAttention-2 backend for dtype other than "
+                "torch.float16 or torch.bfloat16.")
+            selected_backend = _Backend.XFORMERS
+        elif kv_cache_dtype is not None and kv_cache_dtype.startswith("fp8"):
+            logger.info(
+                "Cannot use FlashAttention-2 backend for FP8 KV cache.")
+            logger.warning(
+                "Please use FlashInfer backend with FP8 KV Cache for "
+                "better performance by setting environment variable  "
+                "VLLM_ATTENTION_BACKEND=FLASHINFER")
+            selected_backend = _Backend.XFORMERS
+        elif block_size % 16 != 0:
+            logger.info(
+                "Cannot use FlashAttention-2 backend for block size not "
+                "divisible by 16.")
+            selected_backend = _Backend.XFORMERS
+
+    # FlashAttn is valid for the model, checking if the package is installed.
+    if selected_backend == _Backend.FLASH_ATTN:
+        try:
+            import vllm.vllm_flash_attn  # noqa: F401
+            from vllm.attention.backends.flash_attn import (  # noqa: F401
+                FlashAttentionBackend)
+
+            supported_sizes = FlashAttentionBackend.get_supported_head_sizes()
+            if head_size not in supported_sizes:
+                logger.info(
+                    "Cannot use FlashAttention-2 backend for head size %d.",
+                    head_size)
+                selected_backend = _Backend.XFORMERS
+        except ImportError:
+            logger.info(
+                "Cannot use FlashAttention-2 backend because the "
+                "vllm.vllm_flash_attn package is not found. "
+                "Make sure that vllm_flash_attn was built and installed "
+                "(on by default).")
+            selected_backend = _Backend.XFORMERS
+
+    return selected_backend
+
+
+@contextmanager
+def global_force_attn_backend_context_manager(
+        attn_backend: _Backend) -> Generator[None, None, None]:
+    '''
+    Globally force a vLLM attention backend override within a
+    context manager, reverting the global attention backend
+    override to its prior state upon exiting the context
+    manager.
+
+    Arguments:
+
+    * attn_backend: attention backend to force
+
+    Returns:
+
+    * Generator
+    '''
+
+    # Save the current state of the global backend override (if any)
+    original_value = get_global_forced_attn_backend()
+
+    # Globally force the new backend override
+    global_force_attn_backend(attn_backend)
+
+    # Yield control back to the enclosed code block
+    try:
+        yield
+    finally:
+        # Revert the original global backend override, if any
+        global_force_attn_backend(original_value)

vllm-v0.6.2/vllm/beam_search.py

@@ -0,0 +1,71 @@
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Union
+
+from vllm.sequence import Logprob
+
+if TYPE_CHECKING:
+    from vllm.multimodal import MultiModalDataDict
+
+
+@dataclass
+class BeamSearchSequence:
+    """A sequence for beam search.
+    It keeps track of the tokens and the log probability of the sequence.
+    The text field is optional and will only be filled when the sequence is
+    about to be returned to the user.
+    """
+    # The tokens includes the prompt.
+    tokens: List[int]
+    logprobs: List[Dict[int, Logprob]]
+    cum_logprob: float = 0.0
+    text: Optional[str] = None
+    finish_reason: Optional[str] = None
+    stop_reason: Union[int, str, None] = None
+    multi_modal_data: Optional["MultiModalDataDict"] = None
+    mm_processor_kwargs: Optional[Dict[str, Any]] = None
+
+
+@dataclass
+class BeamSearchOutput:
+    """The output of beam search.
+    It contains the list of the best beam search sequences.
+    The length of the list is equal to the beam width.
+    """
+    sequences: List[BeamSearchSequence]
+
+
+class BeamSearchInstance:
+
+    def __init__(self, prompt_tokens: List[int]):
+        self.beams: List[BeamSearchSequence] = [
+            BeamSearchSequence(tokens=prompt_tokens, logprobs=[])
+        ]
+        self.completed: List[BeamSearchSequence] = []
+
+
+def get_beam_search_score(
+    tokens: List[int],
+    cumulative_logprob: float,
+    eos_token_id: int,
+    length_penalty: float = 1.0,
+) -> float:
+    """Calculate the beam search score with length penalty.
+
+    Adapted from
+
+    https://github.com/huggingface/transformers/blob/ccb92be23def445f2afdea94c31286f84b89eb5b/src/transformers/generation/beam_search.py#L938
+    """
+    seq_len = len(tokens)
+    if tokens[-1] == eos_token_id:
+        seq_len -= 1
+
+    return cumulative_logprob / (seq_len**length_penalty)
+
+
+def create_sort_beams_key_function(eos_token_id: int, length_penalty: float):
+
+    def sort_beams_key(x: BeamSearchSequence) -> float:
+        return get_beam_search_score(x.tokens, x.cum_logprob, eos_token_id,
+                                     length_penalty)
+
+    return sort_beams_key

vllm-v0.6.2/vllm/block.py

@@ -0,0 +1,88 @@
+"""Token blocks."""
+from typing import TYPE_CHECKING, Iterator, List, Optional
+
+from vllm.utils import Device
+
+DEFAULT_LAST_ACCESSED_TIME: float = -1
+
+
+class PhysicalTokenBlock:
+    """Represents the state of a block in the KV cache."""
+
+    def __init__(
+        self,
+        device: Device,
+        block_number: int,
+        block_size: int,
+        block_hash: int,
+        num_hashed_tokens: int,
+    ) -> None:
+        self.device = device
+        self.block_number = block_number
+        self.block_size = block_size
+        self.block_hash = block_hash
+        self.num_hashed_tokens = num_hashed_tokens
+
+        self.ref_count = 0
+        self.last_accessed = DEFAULT_LAST_ACCESSED_TIME
+
+        self.computed = False
+
+    def __repr__(self) -> str:
+        return (f'PhysicalTokenBlock(device={self.device}, '
+                f'block_number={self.block_number}, '
+                f'num_hashed_tokens={self.num_hashed_tokens}, '
+                f'ref_count={self.ref_count}, '
+                f'last_accessed={self.last_accessed}, '
+                f'computed={self.computed})')
+
+
+class BlockTable:
+    """Holds a list of blocks with caching of their associated block_ids 
+    """
+
+    def __init__(self, blocks: Optional[List[PhysicalTokenBlock]] = None):
+        self._blocks: List[PhysicalTokenBlock] = []
+        self._block_ids: List[int] = []
+
+        if blocks is not None:
+            for block in blocks:
+                self.append(block)
+
+    def append(self, block: PhysicalTokenBlock):
+        self._blocks.append(block)
+        self._block_ids.append(block.block_number)
+
+    def __len__(self) -> int:
+        return len(self._blocks)
+
+    def __getitem__(self, key):
+        return self._blocks[key]
+
+    if TYPE_CHECKING:
+
+        def __iter__(self) -> Iterator[PhysicalTokenBlock]:
+            raise RuntimeError("Method should be automatically generated")
+
+    def __setitem__(self, key, value):
+        if isinstance(key, slice):
+            blocks = value
+            self._blocks[key] = blocks
+            self._block_ids[key] = [b.block_number for b in blocks]
+        else:
+            block = value
+            self._blocks[key] = block
+            self._block_ids[key] = block.block_number
+
+    def reset(self):
+        self._blocks = []
+        self._block_ids = []
+
+    def copy(self) -> "BlockTable":
+        return BlockTable(self._blocks)
+
+    def list(self) -> List[PhysicalTokenBlock]:
+        return self._blocks
+
+    def ids(self) -> List[int]:
+        return self._block_ids

vllm-v0.6.2/vllm/compilation/backends.py

@@ -0,0 +1,691 @@
+import copy
+import dataclasses
+import operator
+from contextlib import ExitStack
+from typing import (Any, Callable, Dict, List, Optional, Sequence, Set, Tuple,
+                    Union)
+from unittest.mock import patch
+
+import torch
+import torch.fx as fx
+
+import vllm.envs as envs
+from vllm.logger import init_logger
+from vllm.utils import combine_fx_passes, weak_ref_tensors
+
+from .config import CompilationConfig
+from .counter import compilation_counter
+from .fusion import FusionPass
+from .levels import CompilationLevel
+from .reshapes import RedundantReshapesPass
+
+logger = init_logger(__name__)
+
+
+def fix_functionalization(graph: fx.Graph):
+    """
+    Rewrite the graph module to replace the pattern involving
+    torch._higher_order_ops.auto_functionalize.auto_functionalized
+    with a direct call to the inplace custom op.
+
+    # TODO: check if PyTorch nightly has fixed this issue
+    """
+
+    # debug code, if we want to see the graph before the transformation
+    # with open("before.py", "w") as f:
+    #     print(graph.python_code(root_module="self", verbose=True).src, file=f)
+
+    nodes_to_remove = []
+
+    for node in graph.nodes:
+        # Identify the auto_functionalized node
+        if node.op == 'call_function' and node.target == torch._higher_order_ops.auto_functionalize.auto_functionalized:  # noqa
+            if node.args[0] == torch.ops._C.rotary_embedding.default:
+                # manual replace for rotary_embedding
+
+                # Now, collect the arguments
+                kwargs = node.kwargs
+
+                query = kwargs['query']
+                mm_node = query.args[0].args[0]
+
+                # Create a new call to torch.ops._C.rotary_embedding.default
+                with graph.inserting_before(node):
+                    # just insert the call to the custom op
+                    # NOTE: don't run dead code elimination,
+                    # otherwise this op will be removed
+                    graph.call_function(torch.ops._C.rotary_embedding.default,
+                                        kwargs=kwargs)
+
+                # Remove the auto_functionalized node
+                # Since the node may have outputs, we need to handle its users
+                # Replace uses of the outputs (getitem nodes) with mm_node
+                for user in list(node.users):
+                    if user.op == 'call_function' and user.target == operator.getitem:  # noqa
+                        # Remove the getitem node
+                        for getitem_user in list(user.users):
+                            if (getitem_user.op == 'call_function'
+                                    and getitem_user.target
+                                    == torch.ops.aten.slice_scatter.default):
+                                # Replace the uses of slice_scatter node
+                                # with mm_node
+                                getitem_user.replace_all_uses_with(mm_node)
+                                nodes_to_remove.append(getitem_user)
+                        nodes_to_remove.append(user)
+                nodes_to_remove.append(node)
+
+            elif node.args[0] == torch.ops._C.fused_add_rms_norm.default:
+                # manual replace for fused_add_rms_norm
+                # this is the most effective optimization for llama
+                # failing to do this will result in many unnecessary copies
+
+                kwargs = node.kwargs
+
+                input = kwargs['input']
+                residual = kwargs['residual']
+
+                # Create a new call to torch.ops._C.rotary_embedding.default
+                with graph.inserting_before(node):
+                    # just insert the call to the custom op
+                    # NOTE: don't run dead code elimination,
+                    # otherwise this op will be removed
+                    graph.call_function(
+                        torch.ops._C.fused_add_rms_norm.default, kwargs=kwargs)
+
+                for user in list(node.users):
+                    if user.op == 'call_function' and user.target == operator.getitem:  # noqa
+                        # Remove the getitem node
+                        if user.args[1] == 1:
+                            replace_node = input
+                        elif user.args[1] == 2:
+                            replace_node = residual
+                        user.replace_all_uses_with(replace_node)
+                        nodes_to_remove.append(user)
+                nodes_to_remove.append(node)
+            elif (node.args[0] ==
+                  torch.ops._C.fused_add_rms_norm_static_fp8_quant.default):
+                # manual replace for fused_add_rms_norm_static_fp8_quant
+                # this is the most effective optimization for llama
+                # failing to do this will result in many unnecessary copies
+
+                kwargs = node.kwargs
+
+                result = kwargs['result']
+                residual = kwargs['residual']
+
+                # Create a new call to
+                # torch.ops._C.fused_add_rms_norm_static_fp8_quant.default
+                with graph.inserting_before(node):
+                    # just insert the call to the custom op
+                    # NOTE: don't run dead code elimination,
+                    # otherwise this op will be removed
+                    graph.call_function(
+                        torch.ops._C.fused_add_rms_norm_static_fp8_quant.
+                        default,
+                        kwargs=kwargs)
+
+                for user in list(node.users):
+                    if user.op == 'call_function' and user.target == operator.getitem:  # noqa
+                        # Remove the getitem node
+                        if user.args[1] == 1:
+                            replace_node = result
+                        elif user.args[1] == 2:
+                            replace_node = residual
+                        user.replace_all_uses_with(replace_node)
+                        nodes_to_remove.append(user)
+                nodes_to_remove.append(node)
+
+            elif node.args[0] == torch.ops._C.rms_norm.default:
+                # manual replace for rms_norm
+
+                kwargs = node.kwargs
+
+                replace_node = kwargs['result']
+                # Create a new call to torch.ops._C.rms_norm.default
+                with graph.inserting_before(node):
+                    # just insert the call to the custom op
+                    # NOTE: don't run dead code elimination,
+                    # otherwise this op will be removed
+                    graph.call_function(torch.ops._C.rms_norm.default,
+                                        kwargs=kwargs)
+
+                for user in list(node.users):
+                    if user.op == 'call_function' and user.target == operator.getitem:  # noqa
+                        user.replace_all_uses_with(replace_node)
+                        nodes_to_remove.append(user)
+                nodes_to_remove.append(node)
+
+            elif node.args[
+                    0] == torch.ops._C.rms_norm_static_fp8_quant.default:  # noqa
+                # manual replace for rms_norm_static_fp8_quant
+
+                kwargs = node.kwargs
+
+                replace_node = kwargs['result']
+                # Create a new call to torch.ops._C.rms_norm_static_fp8_quant.default  # noqa
+                with graph.inserting_before(node):
+                    # just insert the call to the custom op
+                    # NOTE: don't run dead code elimination,
+                    # otherwise this op will be removed
+                    graph.call_function(
+                        torch.ops._C.rms_norm_static_fp8_quant.default,
+                        kwargs=kwargs)
+
+                for user in list(node.users):
+                    if user.op == 'call_function' and user.target == operator.getitem:  # noqa
+                        user.replace_all_uses_with(replace_node)
+                        nodes_to_remove.append(user)
+                nodes_to_remove.append(node)
+
+            elif node.args[0] == torch.ops._C.silu_and_mul.default:
+                # manual replace for silu_and_mul
+
+                kwargs = node.kwargs
+
+                input = kwargs['input']
+                out = kwargs['out']
+
+                # Create a new call to torch.ops._C.silu_and_mul.default
+                # cannot use kwargs, because we have an `out`, see https://github.com/pytorch/pytorch/blob/a00faf440888ffb724bad413f329a49e2b6388e7/torch/_inductor/lowering.py#L351 # noqa
+                with graph.inserting_before(node):
+                    # just insert the call to the custom op
+                    # NOTE: don't run dead code elimination,
+                    # otherwise this op will be removed
+                    graph.call_function(
+                        torch.ops._C.silu_and_mul.default,
+                        args=(out, input),
+                    )
+                replace_node = out
+
+                for user in list(node.users):
+                    if user.op == 'call_function' and user.target == operator.getitem:  # noqa
+                        user.replace_all_uses_with(replace_node)
+                        nodes_to_remove.append(user)
+                nodes_to_remove.append(node)
+
+    # Remove the nodes all at once
+    for node in nodes_to_remove:
+        graph.erase_node(node)
+
+    # debug code, if we want to see the graph after the transformation
+    # with open("after.py", "w") as f:
+    #     print(graph.python_code(root_module="self", verbose=True).src, file=f)
+
+
+def wrap_inductor(graph,
+                  example_inputs,
+                  additional_inductor_config,
+                  do_logging=False,
+                  runtime_shape: Optional[int] = None,
+                  use_inductor: bool = True):
+    if not use_inductor:
+        return graph
+
+    compilation_counter.num_inductor_compilations += 1
+
+    if do_logging:
+        if runtime_shape is None:
+            logger.info("Compiling a graph for general shape")
+        else:
+            logger.info("Compiling a graph for shape %s", runtime_shape)
+
+    from torch._inductor import config
+    current_config = config.shallow_copy_dict()
+    from torch._inductor.compile_fx import compile_fx
+
+    if additional_inductor_config is not None:
+        current_config.update(additional_inductor_config)
+
+    # inductor can inplace modify the graph, so we need to copy it
+    # see https://github.com/pytorch/pytorch/issues/138980
+    graph = copy.deepcopy(graph)
+    return compile_fx(graph, example_inputs, config_patches=current_config)
+
+
+@dataclasses.dataclass
+class SplitItem:
+    submod_name: str
+    graph_id: int
+    is_splitting_graph: bool
+    graph: fx.GraphModule
+
+
+def split_graph(graph: fx.GraphModule,
+                ops: List[str]) -> Tuple[fx.GraphModule, List[SplitItem]]:
+    # split graph by ops
+    subgraph_id = 0
+    node_to_subgraph_id = {}
+    split_op_graphs = []
+    for node in graph.graph.nodes:
+        if node.op in ("output", "placeholder"):
+            continue
+        if node.op == 'call_function' and str(node.target) in ops:
+            subgraph_id += 1
+            node_to_subgraph_id[node] = subgraph_id
+            split_op_graphs.append(subgraph_id)
+            subgraph_id += 1
+        else:
+            node_to_subgraph_id[node] = subgraph_id
+
+    # `keep_original_order` is important!
+    # otherwise pytorch might reorder the nodes and
+    # the semantics of the graph will change when we
+    # have mutations in the graph
+    split_gm = torch.fx.passes.split_module.split_module(
+        graph,
+        None,
+        lambda node: node_to_subgraph_id[node],
+        keep_original_order=True)
+
+    outputs = []
+
+    names = [name for (name, module) in split_gm.named_modules()]
+
+    for name in names:
+        if "." in name or name == "":
+            # recursive child module or the root module
+            continue
+
+        module = getattr(split_gm, name)
+
+        graph_id = int(name.replace("submod_", ""))
+        outputs.append(
+            SplitItem(name, graph_id, (graph_id in split_op_graphs), module))
+
+    # sort by intetger graph_id, rather than string name
+    outputs.sort(key=lambda x: x.graph_id)
+
+    return split_gm, outputs
+
+
+# we share the global graph pool among all the backends
+global_graph_pool = None
+
+
+class PiecewiseCompileInterpreter(torch.fx.Interpreter):
+    """Code adapted from `torch.fx.passes.shape_prop.ShapeProp`.
+    It runs the given graph with fake inputs, and compile some
+    submodules specified by `compile_submod_names` with the given
+    compilation configs.
+
+    NOTE: the order in `compile_submod_names` matters, because
+    it will be used to determine the order of the compiled piecewise
+    graphs. The first graph will handle logging, and the last graph
+    has some special cudagraph output handling.
+    """
+
+    def __init__(self, module: torch.fx.GraphModule,
+                 compile_submod_names: List[str],
+                 compilation_configs: CompilationConfig, graph_pool):
+        super().__init__(module)
+        from torch._guards import detect_fake_mode
+        self.fake_mode = detect_fake_mode()
+        self.compile_submod_names = compile_submod_names
+        self.compilation_configs = compilation_configs
+        self.graph_pool = graph_pool
+
+    def run(self, *args):
+        fake_args = [
+            self.fake_mode.from_tensor(t) if isinstance(t, torch.Tensor) else t
+            for t in args
+        ]
+        with self.fake_mode:
+            return super().run(*fake_args)
+
+    def call_module(self, target: torch.fx.node.Target,
+                    args: Tuple[torch.fx.node.Argument,
+                                ...], kwargs: Dict[str, Any]) -> Any:
+        assert isinstance(target, str)
+        output = super().call_module(target, args, kwargs)
+
+        if target in self.compile_submod_names:
+            index = self.compile_submod_names.index(target)
+            submod = self.fetch_attr(target)
+            sym_shape_indices = [
+                i for i, x in enumerate(args) if isinstance(x, torch.SymInt)
+            ]
+            compiled_graph_for_general_shape = wrap_inductor(
+                submod,
+                args,
+                self.compilation_configs.inductor_compile_config,
+                runtime_shape=None,
+                do_logging=index == 0,
+                use_inductor=self.compilation_configs.use_inductor)
+
+            self.module.__dict__[target] = PiecewiseBackend(
+                submod, self.compilation_configs, self.graph_pool, index,
+                len(self.compile_submod_names), sym_shape_indices,
+                compiled_graph_for_general_shape)
+
+            compilation_counter.num_piecewise_capturable_graphs_seen += 1
+
+        return output
+
+
+class VllmBackend:
+    """The compilation backend for `torch.compile` with VLLM.
+    It is used for compilation level of `CompilationLevel.PIECEWISE`,
+    where we customize the compilation.
+
+    The major work of this backend is to split the graph into
+    piecewise graphs, and pass them to the piecewise backend.
+
+    This backend also handles custom passes and adds them to Inductor config.
+    The order of the post-grad post-passes is:
+    1. post_grad_passes (constructor parameter)
+    2. config["post_grad_custom_post_pass"]
+    3. fix_functionalization
+    This way, all passes operate on a functionalized graph.
+    """
+
+    compilation_configs: CompilationConfig
+    graph_pool: Any
+    _called: bool = False
+    # the graph we compiled
+    graph: fx.GraphModule
+    # the stiching graph module for all the piecewise graphs
+    split_gm: fx.GraphModule
+    piecewise_graphs: List[SplitItem]
+    returned_callable: Callable
+    # Inductor passes to run on the graph pre-defunctionalization
+    post_grad_passes: Sequence[Callable]
+    sym_tensor_indices: List[int]
+    input_buffers: List[torch.Tensor]
+
+    def __init__(self, post_grad_passes: Sequence[Callable] = ()):
+        global global_graph_pool
+        if global_graph_pool is None:
+            global_graph_pool = torch.cuda.graph_pool_handle()
+
+        # TODO: in the future, if we want to use multiple
+        # streams, it might not be safe to share a global pool.
+        # only investigate this when we use multiple streams
+        self.graph_pool = global_graph_pool
+        self.post_grad_passes = post_grad_passes
+
+        self.sym_tensor_indices = []
+        self.input_buffers = []
+
+        # `torch.compile` is JIT compiled, so we don't need to
+        # do anything here
+
+    def add_passes_to_config(self):
+        config = self.compilation_configs
+        passes = list(self.post_grad_passes)
+
+        passes = passes + [RedundantReshapesPass(config)]
+
+        if config.enable_fusion:
+            passes = passes + [FusionPass.instance(config)]
+
+        inductor_config = config.inductor_compile_config
+        if "post_grad_custom_post_pass" in inductor_config:
+            passes = passes + [inductor_config["post_grad_custom_post_pass"]]
+
+        # add the fix_functionalization pass last, so that all other
+        # passes operate on a functionalized graph
+        passes = passes + [fix_functionalization]
+        combined_pass = combine_fx_passes(passes)
+        inductor_config["post_grad_custom_post_pass"] = combined_pass
+
+    def __call__(self, graph: fx.GraphModule, example_inputs) -> Callable:
+
+        compilation_counter.num_graphs_seen += 1
+
+        # we control the compilation process, each instance can only be
+        # called once
+        assert not self._called, "VllmBackend can only be called once"
+
+        self.graph = graph
+        # config is read now, because only here can
+        # we get the sizes to capture for cudagraph
+        # from compilation context
+        self.compilation_configs = CompilationConfig.select_and_init_config()
+        self.add_passes_to_config()
+
+        self.split_gm, self.piecewise_graphs = split_graph(
+            graph, self.compilation_configs.non_cudagraph_ops)
+
+        from torch._dynamo.utils import lazy_format_graph_code
+        logger.debug("%s", lazy_format_graph_code("before split", self.graph))
+        logger.debug("%s", lazy_format_graph_code("after split",
+                                                  self.split_gm))
+
+        compilation_counter.num_piecewise_graphs_seen += len(
+            self.piecewise_graphs)
+        submod_names_to_compile = [
+            item.submod_name for item in self.piecewise_graphs
+            if not item.is_splitting_graph
+        ]
+
+        # propagate the split graph to the piecewise backend,
+        # compile submodules with symbolic shapes
+        PiecewiseCompileInterpreter(self.split_gm, submod_names_to_compile,
+                                    self.compilation_configs,
+                                    self.graph_pool).run(*example_inputs)
+
+        self._called = True
+
+        if not self.compilation_configs.use_cudagraph or \
+            not self.compilation_configs.cudagraph_copy_inputs:
+            return self.split_gm
+
+        # if we need to copy input buffers for cudagraph
+        from torch._guards import detect_fake_mode
+        fake_mode = detect_fake_mode()
+        fake_args = [
+            fake_mode.from_tensor(t) if isinstance(t, torch.Tensor) else t
+            for t in example_inputs
+        ]
+
+        # index of tensors that have symbolic shapes (batch size)
+        self.sym_tensor_indices = [
+            i for i, x in enumerate(fake_args)
+            if isinstance(x, torch._subclasses.fake_tensor.FakeTensor)
+        ]
+
+        # compiler managed cudagraph input buffers
+        # we assume the first run with symbolic shapes
+        # has the maximum size among all the tensors
+        self.input_buffers = [
+            example_inputs[x].clone() for x in self.sym_tensor_indices
+        ]
+
+        def copy_and_call(*args):
+            list_args = list(args)
+            for i, index in enumerate(self.sym_tensor_indices):
+                runtime_tensor = list_args[index]
+                runtime_shape = runtime_tensor.shape[0]
+                static_tensor = self.input_buffers[i][:runtime_shape]
+
+                # copy the tensor to the static buffer
+                static_tensor.copy_(runtime_tensor)
+
+                # replace the tensor in the list_args to the static buffer
+                list_args[index] = static_tensor
+            return self.split_gm(*list_args)
+
+        return copy_and_call
+
+
+@dataclasses.dataclass
+class ConcreteSizeEntry:
+    runtime_shape: int
+    need_to_compile: bool  # the size is in compile_sizes
+    use_cudagraph: bool  # the size is in capture_sizes
+
+    compiled: bool = False
+    runnable: Callable = None  # type: ignore
+    num_finished_warmup: int = 0
+    cudagraph: Optional[torch.cuda.CUDAGraph] = None
+    output: Optional[Any] = None
+
+    # for cudagraph debugging, track the input addresses
+    # during capture, and check if they are the same during replay
+    input_addresses: Optional[List[int]] = None
+
+
+class PiecewiseBackend:
+
+    def __init__(self, graph: fx.GraphModule,
+                 compilation_configs: CompilationConfig, graph_pool: Any,
+                 piecewise_compile_index: int, total_piecewise_compiles: int,
+                 sym_shape_indices: List[int],
+                 compiled_graph_for_general_shape: Callable):
+        """
+        The backend for piecewise compilation.
+        It mainly handles the compilation and cudagraph capturing.
+
+        We will compile `self.graph` once for the general shape,
+        and then compile for different shapes specified in
+        `compilation_configs.compile_sizes`.
+
+        Independently, we will capture cudagraph for different shapes.
+
+        If a shape needs both compilation and cudagraph, we will
+        compile it first, and then capture cudagraph.
+        """
+        self.graph = graph
+        self.compilation_configs = compilation_configs
+        self.graph_pool = graph_pool
+        self.piecewise_compile_index = piecewise_compile_index
+        self.total_piecewise_compiles = total_piecewise_compiles
+
+        self.is_first_graph = piecewise_compile_index == 0
+        self.is_last_graph = (
+            piecewise_compile_index == total_piecewise_compiles - 1)
+
+        self.compile_sizes: Set[int] = set(
+            self.compilation_configs.compile_sizes)
+        self.capture_sizes: Set[int] = set(
+            self.compilation_configs.capture_sizes
+        ) if self.compilation_configs.use_cudagraph else set()
+
+        self.first_run_finished = False
+
+        self.compiled_graph_for_general_shape = compiled_graph_for_general_shape  # noqa
+
+        self.sym_shape_indices = sym_shape_indices
+
+        self.is_debugging_mode = envs.VLLM_LOGGING_LEVEL == "DEBUG"
+
+        # the entries for different shapes that we need to either
+        # compile or capture cudagraph
+        self.concrete_size_entries: Dict[int, ConcreteSizeEntry] = {}
+        for shape in self.compile_sizes.union(self.capture_sizes):
+            self.concrete_size_entries[shape] = ConcreteSizeEntry(
+                runtime_shape=shape,
+                need_to_compile=shape in self.compile_sizes,
+                use_cudagraph=shape in self.capture_sizes,
+            )
+
+    def __call__(self, *args) -> Any:
+        if not self.first_run_finished:
+            self.first_run_finished = True
+            return self.compiled_graph_for_general_shape(*args)
+
+        runtime_shape = args[self.sym_shape_indices[0]]
+        if runtime_shape not in self.concrete_size_entries:
+            # we don't need to do anything for this shape
+            return self.compiled_graph_for_general_shape(*args)
+
+        entry = self.concrete_size_entries[runtime_shape]
+
+        if entry.runnable is None:
+            entry.runnable = self.compiled_graph_for_general_shape
+
+        if entry.need_to_compile and not entry.compiled:
+            entry.compiled = True
+            # args are real arguments
+            entry.runnable = wrap_inductor(
+                self.graph,
+                args,
+                self.compilation_configs.inductor_compile_config,
+                runtime_shape=runtime_shape,
+                do_logging=self.is_first_graph,
+                use_inductor=self.compilation_configs.use_inductor)
+
+        if not entry.use_cudagraph:
+            return entry.runnable(*args)
+
+        if entry.cudagraph is None:
+            if entry.num_finished_warmup < self.compilation_configs.cudagraph_num_of_warmups:  # noqa
+                entry.num_finished_warmup += 1
+                if self.is_first_graph:
+                    logger.debug(
+                        "Warming up %s/%s for shape %s",
+                        entry.num_finished_warmup,
+                        self.compilation_configs.cudagraph_num_of_warmups,
+                        runtime_shape)
+                return entry.runnable(*args)
+
+            if self.is_first_graph:
+                # Since we capture cudagraph for many different shapes and
+                # capturing is fast, we don't need to log it for every shape.
+                # We only log it in the debug mode.
+                logger.debug("Capturing a cudagraph for shape %s",
+                             runtime_shape)
+
+            input_addresses = [
+                x.data_ptr() for x in args if isinstance(x, torch.Tensor)
+            ]
+            entry.input_addresses = input_addresses
+            cudagraph = torch.cuda.CUDAGraph()
+
+            with ExitStack() as stack:
+                if not self.is_first_graph:
+                    # during every model forward, we will capture
+                    # many pieces of cudagraphs (roughly one per layer).
+                    # running gc again and again across layers will
+                    # make the cudagraph capture very slow.
+                    # therefore, we only run gc for the first graph,
+                    # and disable gc for the rest of the graphs.
+                    stack.enter_context(patch("gc.collect", lambda: None))
+                    stack.enter_context(
+                        patch("torch.cuda.empty_cache", lambda: None))
+
+                # mind-exploding: carefully manage the reference and memory.
+                with torch.cuda.graph(cudagraph, pool=self.graph_pool):
+                    # `output` is managed by pytorch's cudagraph pool
+                    output = entry.runnable(*args)
+                    if self.is_last_graph:
+                        # by converting it to weak ref,
+                        # the original `output` will immediately be released
+                        # to save memory. It is only safe to do this for
+                        # the last graph, because the output of the last graph
+                        # will not be used by any other cuda graph.
+                        output = weak_ref_tensors(output)
+
+            # here we always use weak ref for the output
+            # to save memory
+            entry.output = weak_ref_tensors(output)
+            entry.cudagraph = cudagraph
+
+            compilation_counter.num_cudagraph_caputured += 1
+
+            # important: we need to return the output, rather than
+            # the weak ref of the output, so that pytorch can correctly
+            # manage the memory during cuda graph capture
+            return output
+
+        if self.is_debugging_mode:
+            # check if the input addresses are the same
+            new_input_addresses = [
+                x.data_ptr() for x in args if isinstance(x, torch.Tensor)
+            ]
+            assert new_input_addresses == entry.input_addresses, (
+                "Input addresses for cudagraphs are different during replay."
+                f" Expected {entry.input_addresses}, got {new_input_addresses}"
+            )
+
+        entry.cudagraph.replay()
+        return entry.output
+
+
+def select_default_backend(level: int) -> Union[str, Callable]:
+    if level in [CompilationLevel.DYNAMO_AS_IS, CompilationLevel.DYNAMO_ONCE]:
+        backend_str = "eager"
+        return backend_str
+    assert level == CompilationLevel.PIECEWISE
+
+    return VllmBackend()

vllm-v0.6.2/vllm/compilation/compile_context.py

@@ -0,0 +1,23 @@
+from contextlib import contextmanager
+from typing import Any
+
+_compile_context: Any = None
+
+
+def get_compile_context() -> Any:
+    """Get the current compile context."""
+    return _compile_context
+
+
+@contextmanager
+def set_compile_context(context: Any):
+    """A context manager that stores the current compile context,
+    usually it is a list of sizes to specialize.
+    """
+    global _compile_context
+    prev_context = _compile_context
+    _compile_context = context
+    try:
+        yield
+    finally:
+        _compile_context = prev_context

vllm-v0.6.2/vllm/compilation/config.py

@@ -0,0 +1,159 @@
+import copy
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+from pydantic import BaseModel, Field, PrivateAttr
+
+import vllm.envs as envs
+from vllm.logger import init_logger
+
+from .compile_context import get_compile_context
+
+logger = init_logger(__name__)
+
+
+class CompilationConfig(BaseModel):
+    """
+    Configuration for compilation.
+    It has two parts:
+    - CudaGraph capture:
+        - use_cudagraph: whether to use cudagraph inside compilation.
+            - False: cudagraph inside compilation is not used.
+            - True: cudagraph inside compilation is used. It requires
+                that all input buffers have fixed addresses.
+            Note that this is orthogonal to the cudagraph capture out
+            side of compilation.
+            TODO: move outside cudagraph logic into compilation.
+            torch.compile will handle cudagraph capture logic in the future.
+        - cudagraph_capture_sizes: sizes to capture cudagraph.
+            - None: capture sizes are inferred from compilation context.
+            - List[int]: capture sizes are specified.
+        - cudagraph_num_of_warmups: number of warmup runs for cudagraph.
+            It means the first several runs will be treated as warmup runs.
+            Only after that, the execution will be recorded, and the recorded
+            cudagraph will be used for subsequent runs.
+        - cudagraph_copy_inputs: whether to copy input tensors for
+            cudagraph. If the caller can guarantee that the same input buffers
+            are always used, it can set this to False. Otherwise, it should
+            set this to True, and the compiler will copy the input to an
+            internally managed buffer. Default is False.
+    - Inductor compilation:
+        - use_inductor: whether to use inductor compilation.
+            - False: inductor compilation is not used. graph runs in eager.
+            - True: inductor compilation is used. one graph for symbolic shape
+                is compiled. In addition, compile for different sizes specified
+                in inductor_compile_sizes, using configurations
+                in inductor_compile_config.
+        - inductor_compile_sizes: sizes to compile for inductor.
+        - inductor_specialize_for_cudagraph_no_more_than: an optional integer
+            to specialize inductor for cudagraph sizes no more than the
+            specified size. It is useful when we want to specialize inductor
+            with a subset of cudagraph sizes.
+        - inductor_compile_config: additional configurations for inductor.
+            - None: use default configurations.
+        - inductor_passes: additional passes for inductor. It is a dictionary
+            from pass name to pass function qualified name. We use function
+            name because the config uses json format. If we pass the config
+            from Python, functions can also be passed directly via Python object
+            constructor, e.g. `CompilationConfig(inductor_passes={"a": func})`
+    - Custom inductor passes:
+        - dump_graph_stages: list of stages for which we want to dump the graph.
+            Each pass defines its own stages (before, after, maybe in-between).
+        - dump_graph_dir: directory to dump the graph. Default is .
+        - enable_fusion: whether to enable the custom fusion pass.
+            TODO better pass enabling system.
+    
+    Why we have different sizes for cudagraph and inductor:
+    - cudagraph: a cudagraph captured for a specific size can only be used
+        for the same size. We need to capture all the sizes we want to use.
+    - inductor: a graph compiled by inductor for a general shape can be used
+        for different sizes. Inductor can also compile for specific sizes,
+        where it can have more information to optimize the graph with fully
+        static shapes. However, we find the general shape compilation is
+        sufficient for most cases. It might be beneficial to compile for
+        certain small batchsizes, where inductor is good at optimizing.
+    """
+    use_inductor: bool = True
+    inductor_specialize_for_cudagraph_no_more_than: Optional[int] = None
+    inductor_compile_sizes: Optional[List[int]] = Field(default_factory=dict)
+    inductor_compile_config: Dict = Field(default_factory=dict)
+    inductor_passes: Dict[str, str] = Field(default_factory=dict)
+
+    use_cudagraph: bool = False
+    non_cudagraph_ops: List[str] = Field(default_factory=list)
+    cudagraph_num_of_warmups: int = 0
+    cudagraph_capture_sizes: Optional[List[int]] = None
+    cudagraph_copy_inputs: bool = False
+
+    dump_graph_stages: List[str] = Field(default_factory=list)
+    dump_graph_dir: Path = Field(default=Path("."))
+    enable_fusion: bool = True
+
+    # not configurable, computed after init
+    compile_sizes: List[int] = PrivateAttr
+    capture_sizes: List[int] = PrivateAttr
+
+    def model_post_init(self, __context: Any) -> None:
+        for k, v in self.inductor_passes.items():
+            if not isinstance(v, str):
+                assert callable(v), (
+                    f"pass {k} should be a function or a qualified name")
+                self.inductor_compile_config[k] = v
+                continue
+
+            # resolve function from qualified name
+            names = v.split(".")
+            module = ".".join(names[:-1])
+            func_name = names[-1]
+            func = __import__(module).__dict__[func_name]
+            self.inductor_compile_config[k] = func
+
+    def init_during_runtime(self):
+        """To complete the initialization of config,
+        we need to know the compile context, which is only available
+        during the first run of the model.
+        """
+        context = get_compile_context()
+        context = copy.deepcopy(context) if context is not None else []
+        sizes_to_specialize: List[int] = context
+        if self.cudagraph_capture_sizes is None:
+            self.capture_sizes = sizes_to_specialize
+        else:
+            self.capture_sizes = self.cudagraph_capture_sizes
+            logger.info(("cudagraph sizes specified by model runner"
+                         " %s is overridden by config %s"),
+                        sizes_to_specialize, self.cudagraph_capture_sizes)
+        if self.inductor_specialize_for_cudagraph_no_more_than is not None:
+            assert self.inductor_compile_sizes is None, (
+                "inductor_compile_sizes should be None when "
+                "inductor_specialize_for_cudagraph_no_more_than is not None")
+            self.compile_sizes = [
+                x for x in self.capture_sizes
+                if x <= self.inductor_specialize_for_cudagraph_no_more_than
+            ]
+        else:
+            assert self.inductor_compile_sizes is not None, (
+                "inductor_compile_sizes should not be None when "
+                "inductor_specialize_for_cudagraph_no_more_than is None")
+            self.compile_sizes = self.inductor_compile_sizes
+
+    @staticmethod
+    def select_and_init_config() -> "CompilationConfig":
+        """The order of selecting config is:
+        1. Use the config specified in environment variable.
+        2. Use the config specified in plugins.
+        3. Use the default config.
+        """
+        config_path = envs.VLLM_TORCH_COMPILE_CONFIG
+        if config_path is not None:
+            with open(config_path) as json_file:
+                config = CompilationConfig.model_validate_json(
+                    json_file.read())
+        else:
+            from vllm.plugins import get_compilation_config
+            predefined_config = get_compilation_config()
+            config = predefined_config if predefined_config is not None else (
+                CompilationConfig())
+
+        config.init_during_runtime()
+        return config

vllm-v0.6.2/vllm/compilation/counter.py

@@ -0,0 +1,30 @@
+import copy
+import dataclasses
+from contextlib import contextmanager
+
+
+@dataclasses.dataclass
+class CompilationCounter:
+    num_graphs_seen: int = 0
+    # including the splitting ops
+    num_piecewise_graphs_seen: int = 0
+    # not including the splitting ops
+    num_piecewise_capturable_graphs_seen: int = 0
+    num_inductor_compilations: int = 0
+    num_cudagraph_caputured: int = 0
+
+    def clone(self) -> "CompilationCounter":
+        return copy.deepcopy(self)
+
+    @contextmanager
+    def expect(self, **kwargs):
+        old = self.clone()
+        yield
+        for k, v in kwargs.items():
+            assert getattr(self, k) - getattr(old, k) == v, (
+                f"{k} not as expected, before it is {getattr(old, k)}"
+                f", after it is {getattr(self, k)}, "
+                f"expected diff is {v}")
+
+
+compilation_counter = CompilationCounter()

vllm-v0.6.2/vllm/compilation/decorators.py

@@ -0,0 +1,182 @@
+import inspect
+from typing import Dict, List, Optional, Union
+
+import torch
+
+import vllm.envs as envs
+from vllm.compilation.levels import CompilationLevel
+from vllm.compilation.wrapper import TorchCompileWrapperWithCustomDispatcher
+from vllm.config import VllmConfig
+from vllm.logger import init_logger
+from vllm.sequence import IntermediateTensors
+from vllm.utils import supports_dynamo
+
+logger = init_logger(__name__)
+
+
+def support_torch_compile(
+        cls: Optional[type] = None,
+        dynamic_arg_dims: Optional[Dict[str, Union[int, List[int]]]] = None):
+    """
+    A decorator to add support for compiling the forward method of a class.
+
+    Usage 1: use directly as a decorator without arguments:
+
+    ```python
+    @support_torch_compile
+    class MyModel(nn.Module):
+        def forward(self, x: torch.Tensor, y: Optional[torch.Tensor]):
+            ...
+    ```
+
+    Usage 2: use as a decorator with arguments:
+
+    ```python
+    @support_torch_compile(dynamic_arg_dims={"x": 0, "y": 0})
+    class MyModel(nn.Module):
+        def forward(self, x: torch.Tensor, y: Optional[torch.Tensor]):
+            ...
+    ```
+
+    `dynamic_arg_dims` is a dictionary that maps argument names to the dynamic
+    dimensions of the argument. The dynamic dimensions can be either a single
+    integer or a list of integers.
+
+    if `dynamic_arg_dims` is `None`, it is inferred from the type annotation
+    of the `forward` method, based on the following default rules:
+
+    - if the argument is annotated as `torch.Tensor` or
+        `Optional[torch.Tensor]`, the first dimension will be
+        marked as dynamic.
+    - if the argument is annotated as `IntermediateTensors`, the first
+        dimension of all the tensors in the intermediate tensors
+        will be marked as dynamic.
+
+    During runtime, when we actually mark dimensions of tensors,
+     it depends on the value of arguments:
+
+    - if it is a single integer, the corresponding dimension of the argument
+        will be marked as dynamic.
+    - if it is `None`, ignored.
+    - if it is `IntermediateTensors`, all the tensors in the intermediate
+        tensors will be marked as dynamic.
+    - otherwise, it will raise an error.
+
+    NOTE: if an argument is `None`, it should always be passed as `None` during
+    the lifetime of the model, otherwise, it cannot be captured as a single
+    computation graph.
+    """
+
+    def cls_decorator_helper(cls: type):
+        # helper to pass `dynamic_arg_dims`` to `_support_torch_compile``
+        # to avoid too much indentation for `_support_torch_compile``
+        if not hasattr(cls, 'forward'):
+            raise TypeError("decorated class should have a forward method.")
+        sig = inspect.signature(cls.forward)
+        inferred_dynamic_arg_dims = dynamic_arg_dims
+        if inferred_dynamic_arg_dims is None:
+            inferred_dynamic_arg_dims = {}
+            for k, v in sig.parameters.items():
+                if v.annotation in [
+                        torch.Tensor, Optional[torch.Tensor],
+                        IntermediateTensors, Optional[IntermediateTensors]
+                ]:
+                    inferred_dynamic_arg_dims[k] = 0
+
+            logger.debug(("Inferred dynamic dimensions for "
+                          "forward method of %s: %s"), cls,
+                         list(inferred_dynamic_arg_dims.keys()))
+
+        if len(inferred_dynamic_arg_dims) == 0:
+            raise ValueError(
+                "No dynamic dimensions found in the forward method of "
+                f"{cls}. Please provide dynamic_arg_dims explicitly.")
+
+        for k in inferred_dynamic_arg_dims:
+            if k not in sig.parameters:
+                raise ValueError(
+                    f"Argument {k} not found in the forward method of {cls}")
+        return _support_torch_compile(cls, inferred_dynamic_arg_dims)
+
+    if cls is not None:
+        # use `support_torch_compile` as a decorator without arguments
+        assert isinstance(cls, type)
+        return cls_decorator_helper(cls)
+
+    return cls_decorator_helper
+
+
+def _support_torch_compile(cls: type,
+                           dynamic_arg_dims: Dict[str, Union[int, List[int]]]):
+    """
+    A decorator to add support for compiling the forward method of a class.
+    """
+    if TorchCompileWrapperWithCustomDispatcher in cls.__bases__:
+        # support decorating multiple times
+        return cls
+
+    # take care of method resolution order
+    # make sure super().__init__ is called on the base class
+    #  other than TorchCompileWrapperWithCustomDispatcher
+    cls.__bases__ = cls.__bases__ + (TorchCompileWrapperWithCustomDispatcher, )
+
+    old_init = cls.__init__  # type: ignore
+
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = '', **kwargs):
+        old_init(self, vllm_config=vllm_config, prefix=prefix, **kwargs)
+        # for CompilationLevel.DYNAMO_AS_IS , the upper level model runner
+        # will handle the compilation, so we don't need to do anything here.
+        self.do_not_compile = envs.VLLM_TORCH_COMPILE_LEVEL in [
+            CompilationLevel.NO_COMPILATION, CompilationLevel.DYNAMO_AS_IS
+        ] or not supports_dynamo()
+        if self.do_not_compile:
+            return
+        TorchCompileWrapperWithCustomDispatcher.__init__(self)
+
+    cls.__init__ = __init__  # type: ignore
+
+    def __call__(self, *args, **kwargs):
+        # torch.compiler.is_compiling() means we are inside the compilation
+        # e.g. TPU has the compilation logic in model runner, so we don't
+        # need to compile the model inside.
+        if self.do_not_compile or torch.compiler.is_compiling():
+            return self.forward(*args, **kwargs)
+
+        # the first compilation needs to have dynamic shapes marked
+        if len(self.compiled_codes) < 1:
+            sig = inspect.signature(self.__class__.forward)
+            bound_args = sig.bind(self, *args, **kwargs)
+            bound_args.apply_defaults()
+            for k, dims in dynamic_arg_dims.items():
+                arg = bound_args.arguments.get(k)
+                if arg is not None:
+                    if isinstance(arg, torch.Tensor):
+                        torch._dynamo.mark_dynamic(arg, dims)
+                    elif isinstance(arg, IntermediateTensors):
+                        for tensor in arg.tensors.values():
+                            torch._dynamo.mark_dynamic(tensor, dims)
+                    else:
+                        raise ValueError(
+                            "Unsupported dynamic dimensions"
+                            f" {dims} for argument {k} with type {type(arg)}.")
+
+        # if we don't use custom dispatcher, we can directly call the
+        # compiled function and let torch.compile handle the dispatching,
+        # with the overhead of guard evaluation and recompilation.
+        if len(self.compiled_codes) < 1 or not self.use_custom_dispatcher:
+            # it seems Dynamo reuse the compilation across instances,
+            # while we need to make sure the compiled code is not reused.
+            # we need to control all the compilation of the model.
+            torch._dynamo.eval_frame.remove_from_cache(
+                self.original_code_object)
+            return self.compiled_callable(*args, **kwargs)
+
+        # usually, capturing the model once is enough, and then we can
+        # dispatch to the compiled code directly, without going through
+        # the Dynamo guard mechanism.
+        with self.dispatch_to_code(0):
+            model_output = self.forward(*args, **kwargs)
+            return model_output
+
+    cls.__call__ = __call__  # type: ignore
+    return cls

vllm-v0.6.2/vllm/compilation/fusion.py

@@ -0,0 +1,291 @@
+import operator
+from typing import Iterable, List, Optional
+
+import torch
+from torch._higher_order_ops.auto_functionalize import auto_functionalized
+from torch._inductor.pattern_matcher import (Match, PatternMatcherPass,
+                                             fwd_only, register_replacement)
+
+from vllm.compilation.config import CompilationConfig
+from vllm.compilation.inductor_pass import InductorPass
+from vllm.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+def rms_pattern_static(result: torch.Tensor, result_rms: torch.Tensor,
+                       input: torch.Tensor, weight: torch.Tensor,
+                       scale: torch.Tensor):
+    at1 = auto_functionalized(torch.ops._C.rms_norm.default,
+                              result=result_rms,
+                              input=input,
+                              weight=weight,
+                              epsilon=1e-5)
+    at2 = auto_functionalized(torch.ops._C.static_scaled_fp8_quant.default,
+                              result=result,
+                              input=at1[1],
+                              scale=scale)
+
+    # result
+    return at2[1]
+
+
+def rms_replacement_static(result: torch.Tensor, result_rms: torch.Tensor,
+                           input: torch.Tensor, weight: torch.Tensor,
+                           scale: torch.Tensor):
+    at = auto_functionalized(torch.ops._C.rms_norm_static_fp8_quant.default,
+                             result=result,
+                             input=input,
+                             weight=weight,
+                             scale=scale,
+                             epsilon=1e-5)
+
+    # result
+    return at[1]
+
+
+def rms_pattern_residual_static(result: torch.Tensor, input: torch.Tensor,
+                                residual: torch.Tensor, weight: torch.Tensor,
+                                scale: torch.Tensor):
+    at = auto_functionalized(torch.ops._C.fused_add_rms_norm.default,
+                             input=input,
+                             residual=residual,
+                             weight=weight,
+                             epsilon=1e-5)
+    at1 = auto_functionalized(torch.ops._C.static_scaled_fp8_quant.default,
+                              result=result,
+                              input=at[1],
+                              scale=scale)
+
+    # result, residual
+    return at1[1], at[2]
+
+
+def rms_replacement_residual_static(result: torch.Tensor, input: torch.Tensor,
+                                    residual: torch.Tensor,
+                                    weight: torch.Tensor, scale: torch.Tensor):
+    at = auto_functionalized(
+        torch.ops._C.fused_add_rms_norm_static_fp8_quant.default,
+        result=result,
+        input=input,
+        residual=residual,
+        weight=weight,
+        scale=scale,
+        epsilon=1e-5)
+    # result, residual
+    return at[1], at[2]
+
+
+def empty_bf16(*args, **kwargs):
+    return torch.empty(*args, **kwargs, dtype=torch.bfloat16, device="cuda")
+
+
+def empty_fp8(*args, **kwargs):
+    fp8 = torch.float8_e4m3fn
+    return torch.empty(*args, **kwargs, dtype=fp8, device="cuda")
+
+
+def empty_fp32(*args, **kwargs):
+    return torch.empty(*args, **kwargs, dtype=torch.float32, device="cuda")
+
+
+# Utilities for post-processing multi-output matches
+def is_func(node: torch.fx.Node, target) -> bool:
+    return node.op == "call_function" and node.target == target
+
+
+# Returns the first auto_functionalized node with the given op (if it exists)
+def find_auto_fn_maybe(nodes: Iterable[torch.fx.Node],
+                       op) -> Optional[torch.fx.Node]:
+    for node in nodes:
+        if is_func(node, auto_functionalized) and node.args[0] == op:  # noqa
+            return node
+    return None
+
+
+# Returns the first auto_functionalized node with the given op
+def find_auto_fn(nodes: Iterable[torch.fx.Node], op) -> torch.fx.Node:
+    node = find_auto_fn_maybe(nodes, op)
+    assert node is not None, f"Could not find {op} in nodes {nodes}"
+    return node
+
+
+# Returns the getitem node that extracts the idx-th element from node
+# (if it exists)
+def find_getitem_maybe(node: torch.fx.Node,
+                       idx: int) -> Optional[torch.fx.Node]:
+    for user in node.users:
+        if is_func(user, operator.getitem) and user.args[1] == idx:
+            return user
+    return None
+
+
+# Returns the getitem node that extracts the idx-th element from node
+def find_getitem(node: torch.fx.Node, idx: int) -> torch.fx.Node:
+    ret = find_getitem_maybe(node, idx)
+    assert ret is not None, f"Could not find getitem {idx} in node {node}"
+    return ret
+
+
+class FusionPass(InductorPass):
+    """
+    This pass fuses a pre-defined set of custom ops into fused ops.
+    It uses the torch pattern matcher to find the patterns and replace them.
+    It also manually processes multi-output matches, as those are broken in
+    the torch pattern matcher.
+
+    Because patterns can only be registered once, the pass is a singleton.
+    This will be addressed in a future version of PyTorch:
+    https://github.com/pytorch/pytorch/pull/139321#issuecomment-2452354980
+    """
+
+    _instance: 'Optional[FusionPass]' = None
+
+    @classmethod
+    def instance(cls, config: CompilationConfig):
+        """
+        Get the singleton instance of the FusionPass.
+        If the instance exists, the config is updated but
+        initialization is not repeated.
+        """
+        if cls._instance is None:
+            cls._instance = FusionPass(config)
+        else:
+            cls._instance.config = config
+        return cls._instance
+
+    def __init__(self, config: CompilationConfig):
+        assert self.__class__._instance is None, \
+            "FusionPass singleton instance already exists"
+        super().__init__(config)
+
+        self.matches: List[Match] = []
+        self.patterns: PatternMatcherPass = PatternMatcherPass(
+            pass_name="fusion_pass")
+
+        # Fuse rms_norm + static_scaled_fp8_quant into
+        # rms_norm_static_fp8_quant
+        inputs = [
+            empty_fp8(5, 4),
+            empty_bf16(5, 4),
+            empty_bf16(5, 4),
+            empty_bf16(1, 5),
+            empty_fp32(1, 1)
+        ]
+        register_replacement(rms_pattern_static, rms_replacement_static,
+                             inputs, fwd_only, self.patterns)
+
+        # Fuse fused_add_rms_norm + static_scaled_fp8_quant into
+        # fused_add_rms_norm_static_fp8_quant
+        # Because pattern has 2 outputs, we need to manually process the match
+        # (see process_matches)
+        inputs = [
+            empty_fp8(5, 4),
+            empty_bf16(5, 4),
+            empty_bf16(5, 4),
+            empty_bf16(1, 5),
+            empty_fp32(1, 1)
+        ]
+        register_replacement(rms_pattern_residual_static,
+                             rms_replacement_residual_static,
+                             inputs,
+                             fwd_only,
+                             self.patterns,
+                             extra_check=lambda m: self.record_match(m))
+
+    def record_match(self, match: Match) -> bool:
+        # Hijack the extra_check to record the match and
+        # save it for post-processing.
+        self.matches.append(match)
+
+        # Return False to prevent automatic replacement.
+        return False
+
+    def process_matches(self, graph: torch.fx.Graph):
+        """
+        Manually process multi-output matches and replace them with fused nodes.
+        This is necessary because the automatic replacement for multi-output
+        matches is broken: https://github.com/pytorch/pytorch/issues/137280
+        """
+        for match in self.matches:
+            # To avoid use-before-definition errors, insert replacement nodes
+            # after the last node in the match.
+            # match.nodes is not guaranteed to be sorted.
+            # Find the last node in the match.
+            for last_node_in_match in reversed(graph.nodes):
+                if last_node_in_match in match.nodes:
+                    break
+            else:
+                raise ValueError("No nodes in graph")
+
+            # Insert a new auto_functionalized node for the fused operation,
+            # as well as getitem nodes to extract the result and residual.
+            # The auto_functionalized node returns a tuple of
+            # (None, result, residual) - None is the function return value.
+            # The resulting graph looks like this:
+            # at = auto_functionalized(torch.ops._C.fused_add_rms_norm_static_fp8_quant.default, ...)  # noqa
+            # result_node_new = at[1]
+            # residual_node_new = at[2]
+            with graph.inserting_after(last_node_in_match):
+                kwargs = match.kwargs
+                kwargs["epsilon"] = 1e-5  # Currently hard-coded in RMSNorm
+
+                fused_node = graph.call_function(
+                    auto_functionalized,
+                    (torch.ops._C.fused_add_rms_norm_static_fp8_quant.default,
+                     ),
+                    kwargs=kwargs)
+
+                graph.inserting_after(fused_node)
+                result_node_new = graph.call_function(operator.getitem,
+                                                      (fused_node, 1))
+                residual_node_new = graph.call_function(
+                    operator.getitem, (fused_node, 2))
+
+            # Last part of replacement is rebinding the users of nodes in the
+            # match to use the new nodes.
+
+            # Find the nodes in the match that we need to rebind
+            rms_node = find_auto_fn(match.nodes,
+                                    torch.ops._C.fused_add_rms_norm.default)
+            quant_node = find_auto_fn(
+                match.nodes, torch.ops._C.static_scaled_fp8_quant.default)
+
+            assert len(rms_node.users) == 2
+            assert len(quant_node.users) == 1
+
+            # meta["val"] is used by de-functionalization and has to contain the
+            # value of the node (tuple of tensors) that would be returned by the
+            # functionalized node during tracing.
+
+            rms_tup = rms_node.meta["val"]
+            quant_tup = quant_node.meta["val"]
+
+            # The result of fused_node must be a tuple with the first element
+            # None (the function return value) and the remaining elements
+            # representing the mutated inputs.
+            fused_tup = (None, quant_tup[1], rms_tup[1], rms_tup[2])
+            fused_node.meta["val"] = fused_tup
+
+            # Find the getitem nodes and replace their uses with the new nodes.
+            # The old nodes will be removed by DCE at the end of the pass.
+            find_getitem(rms_node, 2).replace_all_uses_with(residual_node_new)
+            find_getitem(quant_node, 1).replace_all_uses_with(result_node_new)
+
+        # Finally, remove matched nodes
+        graph.eliminate_dead_code()
+        assert all(node not in graph.nodes for match in self.matches
+                   for node in match.nodes)
+
+    def __call__(self, graph: torch.fx.Graph):
+        self.dump_graph(graph, "before_fusion")
+
+        count = self.patterns.apply(graph)
+        logger.debug("Replaced %s patterns", count)
+        self.dump_graph(graph, "after_pattern_match")
+
+        # Manually process multi-output matches (and run DCE)
+        self.process_matches(graph)
+        logger.debug("Post-processed %s matches", len(self.matches))
+        self.dump_graph(graph, "after_fusion")
+        self.matches.clear()

vllm-v0.6.2/vllm/compilation/inductor_pass.py

@@ -0,0 +1,38 @@
+from abc import ABC, abstractmethod
+
+import torch
+
+from vllm.compilation.config import CompilationConfig
+# yapf: disable
+from vllm.distributed import get_tensor_model_parallel_rank as get_tp_rank
+from vllm.distributed import (
+    get_tensor_model_parallel_world_size as get_tp_world_size)
+from vllm.distributed import model_parallel_is_initialized as p_is_init
+# yapf: enable
+from vllm.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+class InductorPass(ABC):
+
+    @abstractmethod
+    def __call__(self, graph: torch.fx.Graph):
+        raise NotImplementedError
+
+    def __init__(self, config: CompilationConfig):
+        self.config = config
+
+    def dump_graph(self, graph: torch.fx.Graph, stage: str):
+        if stage in self.config.dump_graph_stages:
+            # Make sure filename includes rank in the distributed setting
+            parallel = p_is_init() and get_tp_world_size() > 1
+            rank = f"-{get_tp_rank()}" if parallel else ""
+            filepath = self.config.dump_graph_dir / f"{stage}{rank}.py"
+
+            logger.info("Printing graph to %s", filepath)
+            with open(filepath, "w") as f:
+                src = graph.python_code(root_module="self", verbose=True).src
+                # Add imports so it's not full of errors
+                print("import torch; from torch import device", file=f)
+                print(src, file=f)

vllm-v0.6.2/vllm/compilation/levels.py

@@ -0,0 +1,8 @@
+# constants for the levels of the compilation process
+
+
+class CompilationLevel:
+    NO_COMPILATION = 0
+    DYNAMO_AS_IS = 1
+    DYNAMO_ONCE = 2
+    PIECEWISE = 3

vllm-v0.6.2/vllm/compilation/reshapes.py

@@ -0,0 +1,85 @@
+from typing import Union
+
+import torch.fx
+from torch import SymInt
+
+from vllm.compilation.fusion import is_func
+from vllm.compilation.inductor_pass import InductorPass
+from vllm.logger import init_logger
+
+logger = init_logger(__name__)
+
+
+class RedundantReshapesPass(InductorPass):
+    """
+    This is an inductor pass that removes redundant reshape operations.
+    It is required for RMSNorm-quant fusion to work properly.
+    That's because apply_fp8_linear adds a reshape, which is redundant
+    in the 2D-case.
+
+    Example graph:
+
+    getitem_1: "f16[s0, 4096]" = ...
+    view_1: "f16[s0, 4096]" = torch.reshape(getitem_1, [-1, 4096])
+    at = auto_functionalized(static_scaled_fp8_quant, input = view_1, ...)
+    out: "f8e4m3fn[s0, 4096]" = at[1]
+
+    Can be replaced with:
+    getitem_1: "f16[s0, 4096]" = ...
+    at = auto_functionalized(static_scaled_fp8_quant, input = getitem_1, ...)
+    out: "f8e4m3fn[s0, 4096]" = at[1]
+    """
+
+    def __call__(self, graph: torch.fx.Graph):
+        self.dump_graph(graph, "before_reshapes")
+        count = 0
+        # Remove no-op reshapes/views:
+        for node in graph.nodes:
+            if is_func(node, torch.ops.aten.reshape.default):
+                input, shape = node.args[:2]
+                input_shape = input.meta["val"].shape
+                if len(shape) != len(input_shape):
+                    # Reshape changing rank, skip
+                    continue
+
+                if shape.count(-1) > 1:
+                    # Invalid reshape args, skip
+                    continue
+
+                if all(
+                        self.dims_equivalent(s, i_s)
+                        for s, i_s in zip(shape, input_shape)):
+                    node.replace_all_uses_with(input)
+                    graph.erase_node(node)
+                    count += 1
+
+        logger.debug("Removed %s no-op reshapes", count)
+
+        self.dump_graph(graph, "after_reshapes")
+
+    def dims_equivalent(self, dim: Union[int, torch.fx.Node],
+                        i_dim: Union[int, SymInt]) -> bool:
+        """
+        This function checks if two dimensions are equivalent.
+        :param dim: The dimension arg to reshape
+        :param i_dim: The corresponding dimension in the input tensor
+        :return: Are the dimensions equivalent?
+
+        There are three cases in which the dimensions are equivalent:
+        1. The dimensions are equal (both integers)
+        2. The reshape dimension is -1 (i.e. inferred)
+        3. The dimensions both correspond to the same SymInt
+
+        While case 2 does not guarantee the dimensions are equal,
+        they are equal if all other dimensions are equal.
+
+        In case 3, the reshape dimension is a torch.fx.Node,
+        and its value is a SymInt. That value is equal to the
+        input dimension.
+
+        """
+        # Case 1 and 2
+        if dim == i_dim or dim == -1:
+            return True
+        # Case 3
+        return isinstance(dim, torch.fx.Node) and dim.meta["val"] == i_dim

vllm-v0.6.2/vllm/compilation/wrapper.py

@@ -0,0 +1,102 @@
+import os
+import sys
+from abc import abstractmethod
+from contextlib import contextmanager
+from types import CodeType
+from typing import Callable, List, Optional
+
+import torch
+
+import vllm.envs as envs
+
+from .levels import CompilationLevel
+
+
+class TorchCompileWrapperWithCustomDispatcher:
+    """
+    A wrapper class for torch.compile, with a custom dispatch logic.
+    Subclasses should:
+    1. Implement the forward method
+    2. Implement the dispatch logic in the __call__ method
+        It can use `self.compiled_codes` to access the compiled bytecode,
+        and `with self.dispatch_to_code(index):` to dispatch to
+        the compiled code.
+    3. Implement the `__init__` method to determine how to call
+        `torch.compile` over the forward method.
+    """
+
+    def __init__(self, compiled_callable: Optional[Callable] = None):
+
+        if compiled_callable is None:
+            # default compilation settings
+            # compiling the forward method
+
+            # choose the compile backend
+
+            # if the user has set the backend, use it
+            from vllm.plugins import get_torch_compile_backend
+            backend = get_torch_compile_backend()
+            if backend is None:
+                from vllm.compilation.backends import select_default_backend
+                backend = select_default_backend(envs.VLLM_TORCH_COMPILE_LEVEL)
+
+            compiled_callable = torch.compile(
+                self.forward,
+                fullgraph=envs.VLLM_TEST_DYNAMO_FULLGRAPH_CAPTURE,
+                backend=backend)
+
+        self.compiled_callable = compiled_callable
+        self.original_code_object = self.__class__.forward.__code__
+        self.compiled_codes: List[CodeType] = []
+        torch._dynamo.convert_frame.register_bytecode_hook(self.bytecode_hook)
+
+        # read the env var to determine whether to use the custom dispatcher
+        # subclasses can use this to switch between the custom dispatcher
+        # and the default Dynamo guard mechanism.
+        self.use_custom_dispatcher: bool = \
+            envs.VLLM_TORCH_COMPILE_LEVEL >= CompilationLevel.DYNAMO_ONCE
+
+    def __call__(self, *args, **kwargs):
+        """Implement the dispatch logic here, beyond the torch.compile level.
+        NOTE: this function can have additional arguments beyond the forward
+         method, for directly dispatching to the compiled code.
+        """
+        return self.compiled_callable(*args, **kwargs)
+
+    @abstractmethod
+    def forward(self, *args, **kwargs):
+        ...
+
+    def bytecode_hook(self, old_code: CodeType, new_code: CodeType):
+        """Hook to save the compiled bytecode for direct execution."""
+        if old_code is not self.original_code_object:
+            return
+        # code borrowed from https://github.com/thuml/depyf/blob/f4ad79fadee27ea113b4c75202db1eb1a11c0dbc/depyf/explain/enable_debugging.py#L25
+        frame = sys._getframe()
+        while frame and frame.f_back:
+            frame = frame.f_back
+            code_name = frame.f_code.co_name
+            file_name = frame.f_code.co_filename.split(os.path.sep)[-1]
+            if code_name == "_compile" and file_name == "convert_frame.py":
+                break
+        frame = frame.f_locals["frame"]
+        assert frame.f_code == old_code
+
+        if frame.f_locals["self"] is not self:
+            return
+
+        self.compiled_codes.append(new_code)
+
+    @contextmanager
+    def dispatch_to_code(self, index: int):
+        """Context manager to dispatch to the compiled code.
+        Why does this work? Because Dynamo guarantees that the compiled
+        bytecode has exactly the same arguments, cell variables, and free
+        variables as the original code. Therefore we can directly switch
+        the code object in the function and call it.
+
+        See https://dev-discuss.pytorch.org/t/what-is-the-relationship-requirement-among-original-bytecode-transformed-bytecode-and-bytecode-returned-by-hooks-in-dynamo/1693/7 for more details.
+        """ # noqa
+        self.__class__.forward.__code__ = self.compiled_codes[index]
+        yield
+        self.__class__.forward.__code__ = self.original_code_object

vllm-v0.6.2/vllm/connections.py

@@ -0,0 +1,167 @@
+from pathlib import Path
+from typing import Mapping, MutableMapping, Optional
+from urllib.parse import urlparse
+
+import aiohttp
+import requests
+
+from vllm.version import __version__ as VLLM_VERSION
+
+
+class HTTPConnection:
+    """Helper class to send HTTP requests."""
+
+    def __init__(self, *, reuse_client: bool = True) -> None:
+        super().__init__()
+
+        self.reuse_client = reuse_client
+
+        self._sync_client: Optional[requests.Session] = None
+        self._async_client: Optional[aiohttp.ClientSession] = None
+
+    def get_sync_client(self) -> requests.Session:
+        if self._sync_client is None or not self.reuse_client:
+            self._sync_client = requests.Session()
+
+        return self._sync_client
+
+    # NOTE: We intentionally use an async function even though it is not
+    # required, so that the client is only accessible inside async event loop
+    async def get_async_client(self) -> aiohttp.ClientSession:
+        if self._async_client is None or not self.reuse_client:
+            self._async_client = aiohttp.ClientSession()
+
+        return self._async_client
+
+    def _validate_http_url(self, url: str):
+        parsed_url = urlparse(url)
+
+        if parsed_url.scheme not in ("http", "https"):
+            raise ValueError("Invalid HTTP URL: A valid HTTP URL "
+                             "must have scheme 'http' or 'https'.")
+
+    def _headers(self, **extras: str) -> MutableMapping[str, str]:
+        return {"User-Agent": f"vLLM/{VLLM_VERSION}", **extras}
+
+    def get_response(
+        self,
+        url: str,
+        *,
+        stream: bool = False,
+        timeout: Optional[float] = None,
+        extra_headers: Optional[Mapping[str, str]] = None,
+    ):
+        self._validate_http_url(url)
+
+        client = self.get_sync_client()
+        extra_headers = extra_headers or {}
+
+        return client.get(url,
+                          headers=self._headers(**extra_headers),
+                          stream=stream,
+                          timeout=timeout)
+
+    async def get_async_response(
+        self,
+        url: str,
+        *,
+        timeout: Optional[float] = None,
+        extra_headers: Optional[Mapping[str, str]] = None,
+    ):
+        self._validate_http_url(url)
+
+        client = await self.get_async_client()
+        extra_headers = extra_headers or {}
+
+        return client.get(url,
+                          headers=self._headers(**extra_headers),
+                          timeout=timeout)
+
+    def get_bytes(self, url: str, *, timeout: Optional[float] = None) -> bytes:
+        with self.get_response(url, timeout=timeout) as r:
+            r.raise_for_status()
+
+            return r.content
+
+    async def async_get_bytes(
+        self,
+        url: str,
+        *,
+        timeout: Optional[float] = None,
+    ) -> bytes:
+        async with await self.get_async_response(url, timeout=timeout) as r:
+            r.raise_for_status()
+
+            return await r.read()
+
+    def get_text(self, url: str, *, timeout: Optional[float] = None) -> str:
+        with self.get_response(url, timeout=timeout) as r:
+            r.raise_for_status()
+
+            return r.text
+
+    async def async_get_text(
+        self,
+        url: str,
+        *,
+        timeout: Optional[float] = None,
+    ) -> str:
+        async with await self.get_async_response(url, timeout=timeout) as r:
+            r.raise_for_status()
+
+            return await r.text()
+
+    def get_json(self, url: str, *, timeout: Optional[float] = None) -> str:
+        with self.get_response(url, timeout=timeout) as r:
+            r.raise_for_status()
+
+            return r.json()
+
+    async def async_get_json(
+        self,
+        url: str,
+        *,
+        timeout: Optional[float] = None,
+    ) -> str:
+        async with await self.get_async_response(url, timeout=timeout) as r:
+            r.raise_for_status()
+
+            return await r.json()
+
+    def download_file(
+        self,
+        url: str,
+        save_path: Path,
+        *,
+        timeout: Optional[float] = None,
+        chunk_size: int = 128,
+    ) -> Path:
+        with self.get_response(url, timeout=timeout) as r:
+            r.raise_for_status()
+
+            with save_path.open("wb") as f:
+                for chunk in r.iter_content(chunk_size):
+                    f.write(chunk)
+
+        return save_path
+
+    async def async_download_file(
+        self,
+        url: str,
+        save_path: Path,
+        *,
+        timeout: Optional[float] = None,
+        chunk_size: int = 128,
+    ) -> Path:
+        async with await self.get_async_response(url, timeout=timeout) as r:
+            r.raise_for_status()
+
+            with save_path.open("wb") as f:
+                async for chunk in r.content.iter_chunked(chunk_size):
+                    f.write(chunk)
+
+        return save_path
+
+
+global_http_connection = HTTPConnection()
+"""The global :class:`HTTPConnection` instance used by vLLM."""