[CLI env var] Add VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH in env variables (vllm-project#25274)

Signed-off-by: qqma <[email protected]>
Signed-off-by: DarkLight1337 <[email protected]>
Co-authored-by: qqma <[email protected]>
Co-authored-by: Cyrus Leung <[email protected]>

Author: 3 people

tests/compile/piecewise/test_full_cudagraph.py

@@ -46,7 +46,10 @@ class BackendConfig:
     # FA3 on Hopper
     "FA3":
     BackendConfig(name="FA3",
-                  env_vars={"VLLM_FLASH_ATTN_VERSION": "3"},
+                  env_vars={
+                      "VLLM_FLASH_ATTN_VERSION": "3",
+                      "VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH": "16",
+                  },
                   comp_config={
                       "cudagraph_mode": "FULL",
                   },
@@ -66,6 +69,7 @@ class BackendConfig:
     BackendConfig(name="FlashAttentionMLA",
                   env_vars={
                       "VLLM_ATTENTION_BACKEND": "FLASH_ATTN_MLA",
+                      "VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH": "16",
                   },
                   comp_config={
                       "cudagraph_mode": "FULL_DECODE_ONLY",
@@ -89,7 +93,10 @@ class BackendConfig:
     # FA2
     "FA2":
     BackendConfig(name="FA2",
-                  env_vars={"VLLM_FLASH_ATTN_VERSION": "2"},
+                  env_vars={
+                      "VLLM_FLASH_ATTN_VERSION": "2",
+                      "VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH": "16",
+                  },
                   comp_config={
                       "cudagraph_mode": "FULL",
                   }),

tests/v1/cudagraph/test_cudagraph_mode.py

@@ -47,7 +47,10 @@ class BackendConfig:
     # FA3 on Hopper
     "FA3":
     BackendConfig(name="FA3",
-                  env_vars={"VLLM_FLASH_ATTN_VERSION": "3"},
+                  env_vars={
+                      "VLLM_FLASH_ATTN_VERSION": "3",
+                      "VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH": "16",
+                  },
                   comp_config={
                       "cudagraph_mode": "FULL",
                   },
@@ -67,6 +70,7 @@ class BackendConfig:
     BackendConfig(name="FlashAttentionMLA",
                   env_vars={
                       "VLLM_ATTENTION_BACKEND": "FLASH_ATTN_MLA",
+                      "VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH": "16",
                   },
                   comp_config={
                       "cudagraph_mode": "FULL_DECODE_ONLY",
@@ -75,7 +79,10 @@ class BackendConfig:
     # FA2
     "FA2":
     BackendConfig(name="FA2",
-                  env_vars={"VLLM_FLASH_ATTN_VERSION": "2"},
+                  env_vars={
+                      "VLLM_FLASH_ATTN_VERSION": "2",
+                      "VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH": "16",
+                  },
                   comp_config={
                       "cudagraph_mode": "FULL_AND_PIECEWISE",
                   }),

vllm/envs.py

@@ -119,6 +119,7 @@
     VLLM_SERVER_DEV_MODE: bool = False
     VLLM_V1_OUTPUT_PROC_CHUNK_SIZE: int = 128
     VLLM_MLA_DISABLE: bool = False
+    VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH: int = 16
     VLLM_RAY_PER_WORKER_GPUS: float = 1.0
     VLLM_RAY_BUNDLE_INDICES: str = ""
     VLLM_CUDART_SO_PATH: Optional[str] = None
@@ -946,6 +947,12 @@ def get_vllm_port() -> Optional[int]:
     "VLLM_MLA_DISABLE":
     lambda: bool(int(os.getenv("VLLM_MLA_DISABLE", "0"))),
 
+    # If set, vLLM will pick up the provided Flash Attention MLA
+    # max number splits for cuda graph decode
+    "VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH":
+    lambda: int(os.getenv("VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH",
+                          "16")),
+
     # Number of GPUs per worker in Ray, if it is set to be a fraction,
     # it allows ray to schedule multiple actors on a single GPU,
     # so that users can colocate other actors on the same GPUs as vLLM.
@@ -1379,6 +1386,7 @@ def compute_hash() -> str:
     environment_variables_to_hash = [
         "VLLM_PP_LAYER_PARTITION",
         "VLLM_MLA_DISABLE",
+        "VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH",
         "VLLM_USE_TRITON_FLASH_ATTN",
         "VLLM_USE_TRITON_AWQ",
         "VLLM_DP_RANK",

vllm/v1/attention/backends/flash_attn.py

@@ -8,6 +8,7 @@
 import torch
 
 from vllm import _custom_ops as ops
+from vllm import envs
 from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
                                               AttentionMetadata, AttentionType,
                                               is_quantized_kv_cache)
@@ -33,9 +34,6 @@
 
 logger = init_logger(__name__)
 
-# NOTE(woosuk): This is an arbitrary number. Tune it if needed.
-_DEFAULT_MAX_NUM_SPLITS_FOR_CUDA_GRAPH = 16
-
 
 class FlashAttentionBackend(AttentionBackend):
 
@@ -215,7 +213,8 @@ def __init__(self, kv_cache_spec: AttentionSpec, layer_names: list[str],
             # When using cuda graph, we need to set the upper bound of the
             # number of splits so that large enough intermediate buffers are
             # pre-allocated during capture.
-            self.max_num_splits = _DEFAULT_MAX_NUM_SPLITS_FOR_CUDA_GRAPH
+            self.max_num_splits = (
+                envs.VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH)
 
         # Sliding window size to be used with the AOT scheduler will be
         # populated on first build() call.

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

@@ -6,6 +6,7 @@
 
 import torch
 
+from vllm import envs
 from vllm.attention.backends.abstract import (AttentionLayer, AttentionType,
                                               is_quantized_kv_cache)
 from vllm.attention.utils.fa_utils import (flash_attn_supports_mla,
@@ -24,10 +25,6 @@
 
 logger = init_logger(__name__)
 
-# NOTE(matt): This is an arbitrary number, copied from
-# woosuk's implementation in standard FlashAttention backend
-_DEFAULT_MAX_NUM_SPLITS_FOR_CUDA_GRAPH = 16
-
 
 class FlashAttnMLABackend(MLACommonBackend):
 
@@ -97,7 +94,8 @@ def __init__(self, kv_cache_spec: AttentionSpec, layer_names: list[str],
             # When using cuda graph, we need to set the upper bound of the
             # number of splits so that large enough intermediate buffers are
             # pre-allocated during capture.
-            self.max_num_splits = _DEFAULT_MAX_NUM_SPLITS_FOR_CUDA_GRAPH
+            self.max_num_splits = (
+                envs.VLLM_FLASH_ATTN_MAX_NUM_SPLITS_FOR_CUDA_GRAPH)
 
         # TODO(lucas): Until we add support for the DCP custom masking we need
         #   to restrict decodes to q_len == 1 when DCP is enabled.