feat: Add FlexAttention Backend for Efficient Sparse Attention (#9947)

Co-authored-by: Baizhou Zhang <sobereddiezhang@gmail.com>

python/sglang/srt/layers/attention/torch_flex_backend.py

@@ -0,0 +1,325 @@
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import torch
+from torch.nn.attention.flex_attention import create_block_mask, flex_attention
+
+from sglang.srt.layers.attention.base_attn_backend import AttentionBackend
+from sglang.srt.layers.radix_attention import AttentionType
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
+
+if TYPE_CHECKING:
+    from sglang.srt.layers.radix_attention import RadixAttention
+    from sglang.srt.model_executor.model_runner import ModelRunner
+
+
+class TorchFlexAttnBackend(AttentionBackend):
+    def __init__(self, model_runner: ModelRunner):
+        super().__init__()
+        self.forward_metadata = None
+        self.device = model_runner.device
+        self.flex_attention = torch.compile(flex_attention, dynamic=True)
+        torch._dynamo.config.cache_size_limit = 1024
+        torch._dynamo.config.accumulated_cache_size_limit = 1024
+
+    def init_forward_metadata(self, forward_batch: ForwardBatch):
+        """Init the metadata for a forward pass."""
+        # TODO: find a more elegant way to save memory
+        # Currently maintain the same memory as torch_native_backend
+        torch.cuda.empty_cache()
+
+        # Provide two block_mask Lists per seq_idx for lower latency, later will support per layer level mask generation
+        self.extend_block_masks = []
+        self.decode_block_masks = []
+
+        if forward_batch.forward_mode.is_extend():
+            for seq_idx in range(forward_batch.seq_lens.shape[0]):
+                seq_len_kv = forward_batch.seq_lens[seq_idx]
+                seq_len_q = seq_len_kv
+                self.extend_block_masks.append(
+                    create_block_mask(
+                        self._causal_mask,
+                        None,
+                        None,
+                        seq_len_q,
+                        seq_len_kv,
+                        device=self.device,
+                        _compile=False,
+                    )
+                )
+
+        elif forward_batch.forward_mode.is_decode():
+            for seq_idx in range(forward_batch.seq_lens.shape[0]):
+                seq_len_q = 1
+                seq_len_kv = forward_batch.seq_lens[seq_idx]
+
+                self.decode_block_masks.append(
+                    create_block_mask(
+                        self._decode_mask,
+                        None,
+                        None,
+                        seq_len_q,
+                        seq_len_kv,
+                        device=self.device,
+                        _compile=False,
+                    )
+                )
+
+    def _causal_mask(self, b, h, q_idx, kv_idx):
+        return q_idx >= kv_idx
+
+    def _decode_mask(self, b, h, q_idx, kv_idx):
+        return q_idx <= kv_idx
+
+    def _run_flex_forward_extend(
+        self,
+        query: torch.Tensor,
+        output: torch.Tensor,
+        k_cache: torch.Tensor,
+        v_cache: torch.Tensor,
+        req_to_token: torch.Tensor,
+        req_pool_indices: torch.Tensor,
+        seq_lens: torch.Tensor,
+        extend_prefix_lens: torch.Tensor,
+        extend_seq_lens: torch.Tensor,
+        scaling=None,
+        enable_gqa=False,
+        causal=False,
+    ):
+        """Run the extend forward by using torch flex attention op.
+
+        Args:
+            query: [num_tokens, num_heads, head_size]
+            output: [num_tokens, num_heads, head_size]
+            k_cache: [max_total_num_tokens, num_heads, head_size]
+            v_cache: [max_total_num_tokens, num_heads, head_size]
+            req_to_token: [max_num_reqs, max_context_len]
+            req_pool_indices: [num_seqs]
+            seq_lens: [num_seqs]
+            extend_prefix_lens: [num_seqs]
+            extend_seq_lens: [num_seqs]
+            scaling: float or None
+            enable_gqa: bool
+            causal: bool
+
+        Returns:
+            output: [num_tokens, num_heads, head_size]
+        """
+
+        assert seq_lens.shape[0] == extend_prefix_lens.shape[0]
+        assert seq_lens.shape[0] == extend_seq_lens.shape[0]
+
+        # [num_tokens, num_heads, head_size] -> [num_heads, num_tokens, head_size]
+        query = query.movedim(0, query.dim() - 2)
+
+        start_q, start_kv = 0, 0
+
+        for seq_idx in range(seq_lens.shape[0]):
+            # TODO: this loop process a sequence per iter, this is inefficient.
+            # Need optimize the performance later.
+            extend_seq_len_q = extend_seq_lens[seq_idx]
+            prefill_seq_len_q = extend_prefix_lens[seq_idx]
+
+            seq_len_kv = seq_lens[seq_idx]
+            end_q = start_q + extend_seq_len_q
+            end_kv = start_kv + seq_len_kv
+
+            per_req_query = query[:, start_q:end_q, :]
+            per_req_query_redundant = torch.empty(
+                (per_req_query.shape[0], seq_len_kv, per_req_query.shape[2]),
+                dtype=per_req_query.dtype,
+                device=per_req_query.device,
+            )
+
+            per_req_query_redundant[:, prefill_seq_len_q:, :] = per_req_query
+
+            # get key and value from cache. per_req_tokens contains the kv cache
+            # index for each token in the sequence.
+            req_pool_idx = req_pool_indices[seq_idx]
+            per_req_tokens = req_to_token[req_pool_idx, :seq_len_kv]
+            per_req_key = k_cache[per_req_tokens].movedim(0, query.dim() - 2)
+            per_req_value = v_cache[per_req_tokens].movedim(0, query.dim() - 2)
+
+            if not causal:
+                raise NotImplementedError("Non-causal mode is not yet implemented.")
+
+            per_req_out_redundant = (
+                self.flex_attention(
+                    per_req_query_redundant.unsqueeze(0),
+                    per_req_key.unsqueeze(0),
+                    per_req_value.unsqueeze(0),
+                    block_mask=self.extend_block_masks[seq_idx],
+                    scale=scaling,
+                    enable_gqa=enable_gqa,
+                )
+                .squeeze(0)
+                .movedim(query.dim() - 2, 0)
+            )
+            output[start_q:end_q, :, :] = per_req_out_redundant[
+                prefill_seq_len_q:, :, :
+            ]
+            start_q, start_kv = end_q, end_kv
+        return output
+
+    def _run_flex_forward_decode(
+        self,
+        query: torch.Tensor,
+        output: torch.Tensor,
+        k_cache: torch.Tensor,
+        v_cache: torch.Tensor,
+        req_to_token: torch.Tensor,
+        req_pool_indices: torch.Tensor,
+        seq_lens: torch.Tensor,
+        scaling=None,
+        enable_gqa=False,
+        causal=False,
+    ):
+        """Run the decode forward by using torch flex attention op.
+
+        Args:
+            query: [num_tokens, num_heads, head_size]
+            output: [num_tokens, num_heads, head_size]
+            k_cache: [max_total_num_tokens, num_heads, head_size]
+            v_cache: [max_total_num_tokens, num_heads, head_size]
+            req_to_token: [max_num_reqs, max_context_len]
+            req_pool_indices: [num_seqs]
+            seq_lens: [num_seqs]
+            scaling: float or None
+            enable_gqa: bool
+            causal: bool
+
+        Returns:
+            output: [num_tokens, num_heads, head_size]
+        """
+
+        # [num_tokens, num_heads, head_size] -> [num_heads, num_tokens, head_size]
+        query = query.movedim(0, query.dim() - 2)
+
+        start_q, start_kv = 0, 0
+        for seq_idx in range(seq_lens.shape[0]):
+            # TODO: this loop process a sequence per iter, this is inefficient.
+            # Need optimize the performance later.
+
+            seq_len_q = 1
+            seq_len_kv = seq_lens[seq_idx]
+            end_q = start_q + seq_len_q
+            end_kv = start_kv + seq_len_kv
+
+            per_req_query = query[:, start_q:end_q, :]
+
+            # get key and value from cache. per_req_tokens contains the kv cache
+            # index for each token in the sequence.
+            req_pool_idx = req_pool_indices[seq_idx]
+            per_req_tokens = req_to_token[req_pool_idx, :seq_len_kv]
+            per_req_key = k_cache[per_req_tokens].movedim(0, query.dim() - 2)
+            per_req_value = v_cache[per_req_tokens].movedim(0, query.dim() - 2)
+
+            per_req_out = (
+                self.flex_attention(
+                    per_req_query.unsqueeze(0),
+                    per_req_key.unsqueeze(0),
+                    per_req_value.unsqueeze(0),
+                    block_mask=self.decode_block_masks[seq_idx],
+                    scale=scaling,
+                    enable_gqa=enable_gqa,
+                )
+                .squeeze(0)
+                .movedim(query.dim() - 2, 0)
+            )
+
+            output[start_q:end_q, :, :] = per_req_out
+            start_q, start_kv = end_q, end_kv
+
+        return output
+
+    def forward_extend(
+        self,
+        q,
+        k,
+        v,
+        layer: RadixAttention,
+        forward_batch: ForwardBatch,
+        save_kv_cache=True,
+    ):
+        if layer.qk_head_dim != layer.v_head_dim:
+            o = q.new_empty((q.shape[0], layer.tp_q_head_num * layer.v_head_dim))
+        else:
+            o = torch.empty_like(q)
+
+        if save_kv_cache:
+            forward_batch.token_to_kv_pool.set_kv_buffer(
+                layer, forward_batch.out_cache_loc, k, v
+            )
+
+        use_gqa = layer.tp_q_head_num != layer.tp_k_head_num
+
+        q_ = q.view(-1, layer.tp_q_head_num, layer.qk_head_dim)
+        o_ = o.view(-1, layer.tp_q_head_num, layer.v_head_dim)
+
+        causal = True
+        if layer.is_cross_attention or layer.attn_type == AttentionType.ENCODER_ONLY:
+            raise NotImplementedError(
+                "TorchFlexAttnBackend does not support non-causal attention for now."
+            )
+
+        self._run_flex_forward_extend(
+            q_,
+            o_,
+            forward_batch.token_to_kv_pool.get_key_buffer(layer.layer_id),
+            forward_batch.token_to_kv_pool.get_value_buffer(layer.layer_id),
+            forward_batch.req_to_token_pool.req_to_token,
+            forward_batch.req_pool_indices,
+            forward_batch.seq_lens,
+            forward_batch.extend_prefix_lens,
+            forward_batch.extend_seq_lens,
+            scaling=layer.scaling,
+            enable_gqa=use_gqa,
+            causal=causal,
+        )
+        return o
+
+    def forward_decode(
+        self,
+        q,
+        k,
+        v,
+        layer: RadixAttention,
+        forward_batch: ForwardBatch,
+        save_kv_cache=True,
+    ):
+        # During torch.compile, there is a bug in rotary_emb that causes the
+        # output value to have a 3D tensor shape. This reshapes the output correctly.
+        q = q.reshape(-1, layer.tp_q_head_num * layer.qk_head_dim)
+
+        if layer.qk_head_dim != layer.v_head_dim:
+            o = q.new_empty((q.shape[0], layer.tp_q_head_num * layer.v_head_dim))
+        else:
+            o = torch.empty_like(q)
+
+        if save_kv_cache:
+            forward_batch.token_to_kv_pool.set_kv_buffer(
+                layer, forward_batch.out_cache_loc, k, v
+            )
+
+        use_gqa = layer.tp_q_head_num != layer.tp_k_head_num
+        q_ = q.view(-1, layer.tp_q_head_num, layer.qk_head_dim)
+        o_ = o.view(-1, layer.tp_q_head_num, layer.v_head_dim)
+
+        self._run_flex_forward_decode(
+            q_,
+            o_,
+            forward_batch.token_to_kv_pool.get_key_buffer(layer.layer_id),
+            forward_batch.token_to_kv_pool.get_value_buffer(layer.layer_id),
+            forward_batch.req_to_token_pool.req_to_token,
+            forward_batch.req_pool_indices,
+            forward_batch.seq_lens,
+            scaling=layer.scaling,
+            enable_gqa=use_gqa,
+            causal=False,
+        )
+
+        return o
+
+    def support_triton(self):
+        return False

python/sglang/srt/model_executor/model_runner.py

@@ -1786,6 +1786,12 @@ class ModelRunner:
             )
 
             return TorchNativeAttnBackend(self)
+        elif backend_str == "flex_attention":
+            from sglang.srt.layers.attention.torch_flex_backend import (
+                TorchFlexAttnBackend,
+            )
+
+            return TorchFlexAttnBackend(self)
         elif backend_str == "flashmla":
             from sglang.srt.layers.attention.flashmla_backend import FlashMLABackend
 

python/sglang/srt/server_args.py

@@ -93,6 +93,7 @@ ATTENTION_BACKEND_CHOICES = [
     # Common
     "triton",
     "torch_native",
+    "flex_attention",
     # NVIDIA specific
     "cutlass_mla",
     "fa3",
@@ -592,6 +593,15 @@ class ServerArgs:
             )
             self.disable_cuda_graph = True
 
+        if self.attention_backend == "flex_attention":
+            logger.warning(
+                "Cuda graph is disabled because of using torch Flex Attention backend"
+            )
+            self.disable_cuda_graph = True
+            assert (
+                self.speculative_algorithm is None
+            ), "Speculative decoding is currently not supported with Flex Attention backend"
+
         if is_npu() and self.attention_backend in ["ascend", "hybrid_linear_attn"]:
             logger.warning(
                 "At this moment Ascend attention backend only supports a page_size of 128, change page_size to 128."

test/srt/test_torch_flex_attention_backend.py

@@ -0,0 +1,49 @@
+"""
+Usage:
+python3 -m unittest test_torch_flex_attention_backend.TestTorchFlexAttnBackend.test_gsm8k
+"""
+
+import unittest
+from types import SimpleNamespace
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
+from sglang.test.test_utils import (
+    DEFAULT_MODEL_NAME_FOR_TEST,
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    CustomTestCase,
+    popen_launch_server,
+)
+
+
+class TestTorchFlexAttnBackend(CustomTestCase):
+    def test_gsm8k(self):
+        model = DEFAULT_MODEL_NAME_FOR_TEST
+        base_url = DEFAULT_URL_FOR_TEST
+        process = popen_launch_server(
+            model,
+            base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=["--attention-backend", "flex_attention"],
+        )
+
+        try:
+            args = SimpleNamespace(
+                num_shots=8,
+                data_path=None,
+                num_questions=100,
+                parallel=10,
+                max_new_tokens=512,
+                host="http://127.0.0.1",
+                port=int(base_url.split(":")[-1]),
+            )
+            metrics = run_eval_few_shot_gsm8k(args)
+            print(f"{metrics=}")
+            self.assertGreater(metrics["accuracy"], 0.62)
+        finally:
+            kill_process_tree(process.pid)
+
+
+if __name__ == "__main__":
+    unittest.main()