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

Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Wentao Ye
2025-09-18 12:38:37 -07:00
committed by GitHub
parent 75fb112d80
commit d2a30a2d93
2 changed files with 10 additions and 19 deletions
+10 -7
View File
@@ -191,14 +191,17 @@ class CudaPlatformBase(Platform):
compilation_config = vllm_config.compilation_config
if (envs.VLLM_ALL2ALL_BACKEND == "deepep_high_throughput"
and parallel_config.data_parallel_size > 1
and compilation_config.cudagraph_mode
not in [CUDAGraphMode.NONE, CUDAGraphMode.PIECEWISE]):
and compilation_config.cudagraph_mode != CUDAGraphMode.NONE):
# 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(
"Data Parallel with DeepEP high-throughput: using PIECEWISE "
"CUDA graphs and excluding MoE ops from capture. Set "
"VLLM_ALL2ALL_BACKEND=deepep_low_latency if you need MoE "
"graphs captured as well.")
compilation_config.cudagraph_mode = CUDAGraphMode.PIECEWISE
"Data Parallel: disabling cudagraphs since DP "
"with DeepEP high-throughput kernels are not CUDA Graph "
"compatible. The DeepEP low-latency kernels are CUDA Graph "
"compatible. Set the all_to_all backend to deepep_low_latency "
"to use those kernels instead.")
compilation_config.cudagraph_mode = CUDAGraphMode.NONE
@classmethod
def get_current_memory_usage(cls,