register custom op for flash attn and use from torch.ops

.buildkite/test-pipeline.yaml

@@ -162,6 +162,13 @@ steps:
     - pytest -v -s models/test_oot_registration.py # it needs a clean process
     - pytest -v -s models -m \"not vlm\" --ignore=models/test_oot_registration.py
 
+- label: torch compile integration test
+  source_file_dependencies:
+  - vllm/
+  commands:
+    - pytest -v -s ./compile/test_full_graph.py
+
+
 - label: Vision Language Models Test # 42min
   mirror_hardwares: [amd]
   source_file_dependencies:

tests/compile/test_full_graph.py

@@ -0,0 +1,20 @@
+import os
+
+import pytest
+
+
+@pytest.mark.parametrize("model", ["meta-llama/Meta-Llama-3-8B"])
+def test_full_graph(model):
+    # make sure these models can be captured in full graph mode
+    os.environ["VLLM_TEST_DYNAMO_GRAPH_CAPTURE"] = "1"
+
+    from vllm import LLM, SamplingParams
+    prompts = [
+        "Hello, my name is",
+        "The president of the United States is",
+        "The capital of France is",
+        "The future of AI is",
+    ]
+    sampling_params = SamplingParams(temperature=0)
+    llm = LLM(model="meta-llama/Meta-Llama-3-8B")
+    llm.generate(prompts, sampling_params)

tests/kernels/test_flash_attn.py

@@ -2,7 +2,8 @@
 
 import pytest
 import torch
-from vllm_flash_attn import flash_attn_varlen_func, flash_attn_with_kvcache
+
+import vllm.attention.backends.flash_attn  # noqa: F401
 
 NUM_HEADS = [(16, 16), (32, 8), (64, 8)]
 HEAD_SIZES = [128, 256]
@@ -105,17 +106,31 @@ def test_flash_attn_with_paged_kv(
                                  (num_seqs, max_num_blocks_per_seq),
                                  dtype=torch.int32)
 
-    output = flash_attn_with_kvcache(
-        q=query.unsqueeze(1),
-        k_cache=key_cache,
-        v_cache=value_cache,
+    output = torch.ops.vllm.flash_attn_with_kvcache(
+        decode_query=query.unsqueeze(1),
+        key_cache=key_cache,
+        value_cache=value_cache,
         softmax_scale=scale,
         causal=True,
         block_table=block_tables,
         cache_seqlens=kv_lens_tensor,
         softcap=soft_cap if soft_cap is not None else 0,
     ).squeeze(1)
 
+    torch.library.opcheck(torch.ops.vllm.flash_attn_with_kvcache,
+                          args=tuple(),
+                          kwargs=dict(
+                              decode_query=query.unsqueeze(1),
+                              key_cache=key_cache,
+                              value_cache=value_cache,
+                              softmax_scale=scale,
+                              causal=True,
+                              block_table=block_tables,
+                              cache_seqlens=kv_lens_tensor,
+                              softcap=soft_cap if soft_cap is not None else 0,
+                          ),
+                          test_utils=("test_faketensor", ))
+
     ref_output = ref_paged_attn(
         query=query,
         key_cache=key_cache,
@@ -185,7 +200,7 @@ def test_varlen_with_paged_kv(
                                  (num_seqs, max_num_blocks_per_seq),
                                  dtype=torch.int32)
 
-    output = flash_attn_varlen_func(
+    output = torch.ops.vllm.flash_attn_varlen_func(
         q=query,
         k=key_cache,
         v=value_cache,
@@ -200,6 +215,24 @@ def test_varlen_with_paged_kv(
         softcap=soft_cap if soft_cap is not None else 0,
     )
 
+    torch.library.opcheck(torch.ops.vllm.flash_attn_varlen_func,
+                          args=tuple(),
+                          kwargs=dict(
+                              q=query,
+                              k=key_cache,
+                              v=value_cache,
+                              cu_seqlens_q=cu_query_lens,
+                              cu_seqlens_k=cu_kv_lens,
+                              max_seqlen_q=max_query_len,
+                              max_seqlen_k=max_kv_len,
+                              softmax_scale=scale,
+                              causal=True,
+                              window_size=window_size,
+                              block_table=block_tables,
+                              softcap=soft_cap if soft_cap is not None else 0,
+                          ),
+                          test_utils=("test_faketensor", ))
+
     ref_output = ref_paged_attn(
         query=query,
         key_cache=key_cache,

vllm/attention/backends/flash_attn.py

@@ -3,7 +3,6 @@
 from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Type
 
 import torch
-from vllm_flash_attn import flash_attn_varlen_func, flash_attn_with_kvcache
 
 from vllm import _custom_ops as ops
 from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
@@ -18,6 +17,108 @@
 if TYPE_CHECKING:
     from vllm.worker.model_runner import ModelInputForGPUBuilder
 
+from vllm_flash_attn import flash_attn_varlen_func as _flash_attn_varlen_func
+from vllm_flash_attn import flash_attn_with_kvcache as _flash_attn_with_kvcache
+
+
+@torch.library.custom_op("vllm::flash_attn_varlen_func", mutates_args=[])
+def flash_attn_varlen_func(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    cu_seqlens_q: torch.Tensor,
+    cu_seqlens_k: torch.Tensor,
+    max_seqlen_q: int,
+    max_seqlen_k: int,
+    softmax_scale: Optional[float] = None,
+    causal: bool = False,
+    window_size: Optional[List[int]] = None,
+    softcap: float = 0.0,
+    alibi_slopes: Optional[torch.Tensor] = None,
+    block_table: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    # custom op does not support tuple input
+    real_window_size: Tuple[int, int]
+    if window_size is None:
+        real_window_size = (-1, -1)
+    else:
+        assert len(window_size) == 2
+        real_window_size = (window_size[0], window_size[1])
+    return _flash_attn_varlen_func(
+        q=q,
+        k=k,
+        v=v,
+        cu_seqlens_q=cu_seqlens_q,
+        cu_seqlens_k=cu_seqlens_k,
+        max_seqlen_q=max_seqlen_q,
+        max_seqlen_k=max_seqlen_k,
+        softmax_scale=softmax_scale,
+        causal=causal,
+        window_size=real_window_size,
+        softcap=softcap,
+        alibi_slopes=alibi_slopes,
+        block_table=block_table,
+    )
+
+
+@flash_attn_varlen_func.register_fake  # type: ignore
+def _(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    cu_seqlens_q: torch.Tensor,
+    cu_seqlens_k: torch.Tensor,
+    max_seqlen_q: int,
+    max_seqlen_k: int,
+    softmax_scale: Optional[float] = None,
+    causal: bool = False,
+    window_size: Optional[List[int]] = None,
+    softcap: float = 0.0,
+    alibi_slopes: Optional[torch.Tensor] = None,
+    block_table: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    return torch.empty_like(q)
+
+
+@torch.library.custom_op("vllm::flash_attn_with_kvcache", mutates_args=[])
+def flash_attn_with_kvcache(
+    decode_query: torch.Tensor,
+    key_cache: torch.Tensor,
+    value_cache: torch.Tensor,
+    cache_seqlens: Optional[torch.Tensor] = None,
+    block_table: Optional[torch.Tensor] = None,
+    softmax_scale: Optional[float] = None,
+    causal: bool = False,
+    alibi_slopes: Optional[torch.Tensor] = None,
+    softcap: float = 0.0,
+) -> torch.Tensor:
+    return _flash_attn_with_kvcache(
+        decode_query,
+        key_cache,
+        value_cache,
+        cache_seqlens=cache_seqlens,
+        block_table=block_table,
+        softmax_scale=softmax_scale,
+        causal=causal,
+        alibi_slopes=alibi_slopes,
+        softcap=softcap,
+    )
+
+
+@flash_attn_with_kvcache.register_fake  # type: ignore
+def _(
+    decode_query: torch.Tensor,
+    key_cache: torch.Tensor,
+    value_cache: torch.Tensor,
+    cache_seqlens: Optional[torch.Tensor] = None,
+    block_table: Optional[torch.Tensor] = None,
+    softmax_scale: Optional[float] = None,
+    causal: bool = False,
+    alibi_slopes: Optional[torch.Tensor] = None,
+    softcap: float = 0.0,
+) -> torch.Tensor:
+    return torch.empty_like(decode_query)
+
 
 class FlashAttentionBackend(AttentionBackend):
 
@@ -517,7 +618,7 @@ def forward(
                 # normal attention
                 # When block_tables are not filled, it means q and k are the
                 # prompt, and they have the same length.
-                out = flash_attn_varlen_func(
+                out = torch.ops.vllm.flash_attn_varlen_func(
                     q=query,
                     k=key,
                     v=value,
@@ -537,34 +638,36 @@ def forward(
                 # prefix-enabled attention
                 assert prefill_meta.seq_lens is not None
                 max_seq_len = max(prefill_meta.seq_lens)
-                output[:num_prefill_tokens] = flash_attn_varlen_func(
-                    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,
+                output[:
+                       num_prefill_tokens] = torch.ops.vllm.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=self.scale,
+                           causal=True,
+                           alibi_slopes=self.alibi_slopes,
+                           block_table=prefill_meta.block_tables,
+                           softcap=self.logits_soft_cap,
+                       )
+
+        if decode_meta := attn_metadata.decode_metadata:
+            # Decoding run.
+            output[
+                num_prefill_tokens:] = torch.ops.vllm.flash_attn_with_kvcache(
+                    decode_query.unsqueeze(1),
+                    key_cache,
+                    value_cache,
+                    block_table=decode_meta.block_tables,
+                    cache_seqlens=decode_meta.seq_lens_tensor,
                     softmax_scale=self.scale,
                     causal=True,
                     alibi_slopes=self.alibi_slopes,
-                    block_table=prefill_meta.block_tables,
                     softcap=self.logits_soft_cap,
-                )
-
-        if decode_meta := attn_metadata.decode_metadata:
-            # Decoding run.
-            output[num_prefill_tokens:] = flash_attn_with_kvcache(
-                decode_query.unsqueeze(1),
-                key_cache,
-                value_cache,
-                block_table=decode_meta.block_tables,
-                cache_seqlens=decode_meta.seq_lens_tensor,
-                softmax_scale=self.scale,
-                causal=True,
-                alibi_slopes=self.alibi_slopes,
-                softcap=self.logits_soft_cap,
-            ).squeeze(1)
+                ).squeeze(1)
 
         # Reshape the output tensor.
         return output.view(num_tokens, hidden_size)

vllm/attention/backends/flashinfer.py

@@ -4,9 +4,9 @@
 try:
     from flashinfer import BatchDecodeWithPagedKVCacheWrapper
     from flashinfer.prefill import BatchPrefillWithPagedKVCacheWrapper
-    from vllm_flash_attn import flash_attn_varlen_func
+
+    import vllm.attention.backends.flash_attn  # noqa
 except ImportError:
-    flash_attn_varlen_func = None
     BatchDecodeWithPagedKVCacheWrapper = None
     BatchPrefillWithPagedKVCacheWrapper = None
 
@@ -520,7 +520,7 @@ def forward(
             # This happens when vllm runs the profiling to
             # determine the number of blocks.
             if kv_cache is None:
-                output = flash_attn_varlen_func(
+                output = torch.ops.vllm.flash_attn_varlen_func(
                     q=query,
                     k=key,
                     v=value,