init src 0.9.2

2026-01-09 15:09:53 +08:00
parent 0eb2c0a4b3
commit 41d98d4359
1438 changed files with 417605 additions and 683 deletions
--- a/vllm/attention/backends/triton_mla.py
+++ b/vllm/attention/backends/triton_mla.py
@@ -0,0 +1,135 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+import os
+
+from typing import Any, Dict, List, Optional, Type
+from .triton_config import get_nearest_config, get_attention_mla_configs, get_config, get_attention_mla_configs_json
+
+import torch
+
+from vllm.attention.backends.abstract import (AttentionType,
+                                              is_quantized_kv_cache)
+from vllm.attention.backends.mla.common import (MLACommonBackend,
+                                                MLACommonImpl,
+                                                MLACommonMetadata)
+from vllm.attention.ops.triton_decode_attention import decode_attention_fwd
+import vllm.envs as envs
+
+from vllm.logger import init_logger
+logger = init_logger(__name__)
+
+
+class TritonMLABackend(MLACommonBackend):
+
+    @staticmethod
+    def get_name() -> str:
+        return "TRITON_MLA"
+
+    @staticmethod
+    def get_impl_cls() -> Type["TritonMLAImpl"]:
+        return TritonMLAImpl
+
+
+class TritonMLAImpl(MLACommonImpl[MLACommonMetadata]):
+
+    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]],
+            logits_soft_cap: Optional[float],
+            attn_type: str,
+            kv_sharing_target_layer_name: Optional[str],
+            # MLA Specific Arguments
+            **mla_args) -> None:
+        super().__init__(num_heads, head_size, scale, num_kv_heads,
+                         alibi_slopes, sliding_window, kv_cache_dtype,
+                         blocksparse_params, logits_soft_cap, attn_type,
+                         kv_sharing_target_layer_name, **mla_args)
+
+        unsupported_features = [
+            alibi_slopes, sliding_window, blocksparse_params, logits_soft_cap
+        ]
+        if any(unsupported_features):
+            raise NotImplementedError(
+                "TritonMLAImpl does not support one of the following: "
+                "alibi_slopes, sliding_window, blocksparse_params, "
+                "logits_soft_cap")
+
+        if attn_type != AttentionType.DECODER:
+            raise NotImplementedError("Encoder self-attention and "
+                                      "encoder/decoder cross-attention "
+                                      "are not implemented for "
+                                      "TritonMLAImpl")
+            
+        if envs.VLLM_USE_TRITON_OPT_MLA:
+            self.attn_configs = get_attention_mla_configs_json(self.num_heads, 1, self.kv_lora_rank + self.qk_rope_head_dim, self.kv_lora_rank, "fp16")
+
+        if is_quantized_kv_cache(self.kv_cache_dtype):
+            raise NotImplementedError(
+                "TritonMLA with FP8 KV cache not yet supported")
+
+    def _forward_decode(
+        self,
+        q_nope: torch.Tensor,
+        q_pe: torch.Tensor,
+        kv_c_and_k_pe_cache: torch.Tensor,
+        attn_metadata: MLACommonMetadata,
+    ) -> torch.Tensor:
+        assert kv_c_and_k_pe_cache.numel() > 0
+
+        decode_meta = attn_metadata.decode_metadata
+        assert decode_meta is not None
+        B = q_nope.shape[0]
+
+        q = torch.cat([q_nope, q_pe], dim=-1)
+        o = torch.zeros(B,
+                        self.num_heads,
+                        self.kv_lora_rank,
+                        dtype=q.dtype,
+                        device=q.device)
+
+        num_kv_splits = 4  # TODO: heuristic
+
+        # TODO(lucas) Allocate ahead of time
+        attn_logits = torch.empty(
+            (
+                B,
+                self.num_heads,
+                num_kv_splits,
+                # NOTE(lucas) idk why the +1 is here but sglang has it so we
+                # just mirror that
+                self.kv_lora_rank + 1,
+            ),
+            dtype=torch.float32,
+            device=q.device,
+        )
+
+        # Add a head dim of 1
+        kv_c_and_k_pe_cache = kv_c_and_k_pe_cache.unsqueeze(2)
+        kv_c_cache = kv_c_and_k_pe_cache[..., :self.kv_lora_rank]
+        PAGE_SIZE = kv_c_and_k_pe_cache.size(1)
+        
+        # TODO
+        max_seq_len = torch.max(decode_meta.seq_lens_tensor).item()
+        if os.environ.get('PA_MATCH_USE_MEAN_SEQ') == '1':
+            match_seq_len = int((decode_meta.seq_lens_tensor.sum()/ max(1, B)).item())
+        else:
+            match_seq_len = max_seq_len
+        
+        if envs.VLLM_USE_TRITON_OPT_MLA:
+            best_config = self.attn_configs[min(self.attn_configs.keys(), key=lambda x: abs(int(x) - match_seq_len))]
+                
+        # Run MQA
+        decode_attention_fwd(q, kv_c_and_k_pe_cache, kv_c_cache, o,
+                             decode_meta.block_tables,
+                             decode_meta.seq_lens_tensor, attn_logits,
+                             num_kv_splits, self.scale, best_config, PAGE_SIZE)
+
+        return self._v_up_proj(o)