[V1][CUDA] Full cudagraph support for FlashInfer (#21367)

vllm/v1/attention/backends/flash_attn.py

@@ -25,7 +25,8 @@ if is_flash_attn_varlen_func_available():
 from vllm.config import VllmConfig, get_layers_from_vllm_config
 from vllm.logger import init_logger
 from vllm.utils import cdiv
-from vllm.v1.attention.backends.utils import (AttentionMetadataBuilder,
+from vllm.v1.attention.backends.utils import (AttentionCGSupport,
+                                              AttentionMetadataBuilder,
                                               CommonAttentionMetadata,
                                               get_kv_cache_layout)
 from vllm.v1.kv_cache_interface import AttentionSpec
@@ -153,7 +154,9 @@ def _get_sliding_window_configs(
 
 class FlashAttentionMetadataBuilder(
         AttentionMetadataBuilder[FlashAttentionMetadata]):
-    full_cudagraph_supported: ClassVar[bool] = get_flash_attn_version() == 3
+    attn_cudagraph_support: ClassVar[AttentionCGSupport] = \
+        AttentionCGSupport.NEVER if get_flash_attn_version() == 2 \
+        else AttentionCGSupport.ALWAYS
 
     def __init__(self, kv_cache_spec: AttentionSpec, layer_names: list[str],
                  vllm_config: VllmConfig, device: torch.device):

vllm/v1/attention/backends/mla/flashmla.py

@@ -18,6 +18,7 @@ from vllm.v1.attention.backends.mla.common import (MLACommonBackend,
                                                    MLACommonImpl,
                                                    MLACommonMetadata,
                                                    MLACommonMetadataBuilder)
+from vllm.v1.attention.backends.utils import AttentionCGSupport
 from vllm.v1.kv_cache_interface import AttentionSpec
 
 logger = init_logger(__name__)
@@ -54,7 +55,8 @@ class FlashMLAMetadata(MLACommonMetadata[FlashMLADecodeMetadata]):
 
 
 class FlashMLAMetadataBuilder(MLACommonMetadataBuilder[FlashMLAMetadata]):
-    full_cudagraph_supported: ClassVar[bool] = True  # Decode-only
+    attn_cudagraph_support: ClassVar[AttentionCGSupport] = \
+        AttentionCGSupport.PURE_DECODE_ONLY
 
     def __init__(self, kv_cache_spec: AttentionSpec, layer_names: list[str],
                  vllm_config: VllmConfig, device: torch.device):

vllm/v1/attention/backends/mla/rocm_aiter_mla.py

@@ -17,6 +17,7 @@ from vllm.v1.attention.backends.mla.common import (MLACommonBackend,
                                                    MLACommonImpl,
                                                    MLACommonMetadata,
                                                    MLACommonMetadataBuilder)
+from vllm.v1.attention.backends.utils import AttentionCGSupport
 from vllm.v1.kv_cache_interface import AttentionSpec
 
 # yapf: enable
@@ -64,7 +65,8 @@ class AiterMLAMetadata(MLACommonMetadata[AiterMLADecodeMetadata]):
 
 
 class AiterMLAMetadataBuilder(MLACommonMetadataBuilder[AiterMLAMetadata]):
-    full_cudagraph_supported: ClassVar[bool] = True  # decode only
+    attn_cudagraph_support: ClassVar[AttentionCGSupport] = \
+        AttentionCGSupport.PURE_DECODE_ONLY
 
     def __init__(self, kv_cache_spec: AttentionSpec, layer_names: list[str],
                  vllm_config: VllmConfig, device: torch.device):

vllm/v1/attention/backends/triton_attn.py

@@ -18,7 +18,8 @@ from vllm.config import VllmConfig
 from vllm.logger import init_logger
 from vllm.platforms import current_platform
 from vllm.v1.attention.backends.flash_attn import FlashAttentionMetadata
-from vllm.v1.attention.backends.utils import (AttentionMetadataBuilder,
+from vllm.v1.attention.backends.utils import (AttentionCGSupport,
+                                              AttentionMetadataBuilder,
                                               CommonAttentionMetadata)
 from vllm.v1.kv_cache_interface import AttentionSpec
 
@@ -57,7 +58,8 @@ class TritonAttentionMetadata:
 
 class TritonAttentionMetadataBuilder(
         AttentionMetadataBuilder[TritonAttentionMetadata]):
-    full_cudagraph_supported: ClassVar[bool] = True
+    attn_cudagraph_support: ClassVar[AttentionCGSupport] = \
+        AttentionCGSupport.ALWAYS
 
     def __init__(self, kv_cache_spec: AttentionSpec, layer_names: list[str],
                  vllm_config: VllmConfig, device: torch.device):

vllm/v1/attention/backends/utils.py

 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import abc
+import enum
 import functools
 from abc import abstractmethod
 from dataclasses import dataclass, make_dataclass
@@ -65,9 +66,24 @@ class CommonAttentionMetadata:
 M = TypeVar("M")
 
 
+class AttentionCGSupport(enum.Enum):
+    """ Constants for the cudagraph support of the attention backend
+    Here we do not consider the cascade attention, as currently
+    it is never cudagraph supported."""
+
+    NEVER = 0
+    """NO cudagraph support"""
+    PURE_DECODE_ONLY = 1
+    """Cudagraph supported for pure decode, need to run without
+    cudagraph for mixed prefill-decode batches"""
+    ALWAYS = 2
+    """Cudagraph always supported"""
+
+
 class AttentionMetadataBuilder(abc.ABC, Generic[M]):
     # Does this backend/builder support CUDA Graphs for attention.
-    full_cudagraph_supported: ClassVar[bool] = False
+    attn_cudagraph_support: ClassVar[AttentionCGSupport] = \
+        AttentionCGSupport.NEVER
 
     @abstractmethod
     def __init__(self, kv_cache_spec: AttentionSpec, layer_names: list[str],

vllm/v1/worker/gpu_model_runner.py

@@ -47,7 +47,7 @@ from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, DeviceMemoryProfiler,
                         is_pin_memory_available, round_up, supports_dynamo)
 from vllm.v1.attention.backends.mamba_selectors import get_mamba_attn_backend
 from vllm.v1.attention.backends.utils import (
-    AttentionMetadataBuilder, CommonAttentionMetadata,
+    AttentionCGSupport, AttentionMetadataBuilder, CommonAttentionMetadata,
     make_kv_sharing_fast_prefill_attention_metadata,
     make_local_attention_virtual_batches)
 from vllm.v1.core.encoder_cache_manager import compute_encoder_budget
@@ -2619,12 +2619,22 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
             self.device,
         )
 
-        if (self.full_cuda_graph
-                and not attn_metadata_builder_i.full_cudagraph_supported):
-            raise ValueError(
-                f"Full CUDAGraph not supported for "
-                f"{attn_backend_i.__name__}. Turn off CompilationConfig."
-                f"full_cuda_graph or use a different attention backend.")
+        if self.full_cuda_graph:
+            if attn_metadata_builder_i.attn_cudagraph_support == \
+                AttentionCGSupport.NEVER:
+                raise ValueError(f"Full CUDAGraph not supported for "
+                                 f"{attn_backend_i.__name__}. Turn off "
+                                 f"CompilationConfig.full_cuda_graph or use a "
+                                 f" different attention backend.")
+            if attn_metadata_builder_i.attn_cudagraph_support == \
+                AttentionCGSupport.PURE_DECODE_ONLY:
+                # Limit the max cudagraph size to the max number of
+                # sequences for pure decode only cudagraph backend,
+                # whose max_query_len is 1.
+                self.cudagraph_batch_sizes = [
+                    size for size in self.cudagraph_batch_sizes
+                    if size <= self.scheduler_config.max_num_seqs
+                ]
         return attn_backend_i, attn_metadata_builder_i
 
     def initialize_attn_backend(self, kv_cache_config: KVCacheConfig) -> None:

vllm/v1/worker/gpu_worker.py

@@ -321,11 +321,16 @@ class Worker(WorkerBase):
         if get_pp_group().is_last_rank:
             max_num_reqs = min(self.scheduler_config.max_num_seqs,
                                self.scheduler_config.max_num_batched_tokens)
+            # activate building attn_metadata for this dummy run to avoid
+            # potential illegal memory access for full cudagraph relay.
+            attn_cudagraph = self.compilation_config.full_cuda_graph and\
+                not self.model_config.enforce_eager
 
             # We skip EPLB here since we don't want to record dummy metrics
             hidden_states, last_hidden_states = \
                 self.model_runner._dummy_run(
                     num_tokens=max_num_reqs,
+                    capture_attn_cudagraph=attn_cudagraph,
                     skip_eplb=True,
                 )
             if self.model_runner.is_pooling_model: