xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-11 04:54:59 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bug] Fix torch Compilation Cache Hit Error (#25093)

Browse Source 
Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Wentao Ye 2025-09-18 15:38:37 -04:00 committed by GitHub
parent 75fb112d80
commit d2a30a2d93
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 10 additions and 19 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
vllm/config/compilation.py
View File

@ -563,18 +563,6 @@ class CompilationConfig:
self.cudagraph_mode = CUDAGraphMode.FULL self.cudagraph_mode = CUDAGraphMode.FULL
self.splitting_ops = [] self.splitting_ops = []
if envs.VLLM_ALL2ALL_BACKEND == "deepep_high_throughput":
# exclude MoE dispatch/combine from capture by ensuring
# piecewise splitting includes them, so communication remains
# outside CUDA graphs while compute can still be graphed.
moe_ops = [
"vllm.moe_forward",
"vllm.moe_forward_shared",
]
for op in moe_ops:
if op not in self.splitting_ops:
self.splitting_ops.append(op)
def splitting_ops_contain_attention(self) -> bool: def splitting_ops_contain_attention(self) -> bool:
return self.splitting_ops is not None and all( return self.splitting_ops is not None and all(
op in self.splitting_ops for op in self._attention_ops) op in self.splitting_ops for op in self._attention_ops)

17
vllm/platforms/cuda.py
View File

@ -191,14 +191,17 @@ class CudaPlatformBase(Platform):
compilation_config = vllm_config.compilation_config compilation_config = vllm_config.compilation_config
if (envs.VLLM_ALL2ALL_BACKEND == "deepep_high_throughput" if (envs.VLLM_ALL2ALL_BACKEND == "deepep_high_throughput"
and parallel_config.data_parallel_size > 1 and parallel_config.data_parallel_size > 1
and compilation_config.cudagraph_mode and compilation_config.cudagraph_mode != CUDAGraphMode.NONE):
not in [CUDAGraphMode.NONE, CUDAGraphMode.PIECEWISE]): # TODO: Piecewise Cuda graph might be enabled
# if torch compile cache key issue fixed
# See https://github.com/vllm-project/vllm/pull/25093
logger.info( logger.info(
"Data Parallel with DeepEP high-throughput: using PIECEWISE " "Data Parallel: disabling cudagraphs since DP "
"CUDA graphs and excluding MoE ops from capture. Set " "with DeepEP high-throughput kernels are not CUDA Graph "
"VLLM_ALL2ALL_BACKEND=deepep_low_latency if you need MoE " "compatible. The DeepEP low-latency kernels are CUDA Graph "
"graphs captured as well.") "compatible. Set the all_to_all backend to deepep_low_latency "
compilation_config.cudagraph_mode = CUDAGraphMode.PIECEWISE "to use those kernels instead.")
compilation_config.cudagraph_mode = CUDAGraphMode.NONE
@classmethod @classmethod
def get_current_memory_usage(cls, def get_current_memory_usage(cls,