[BugFix] Fix VllmConfig() construction on all platforms (#20695)

Signed-off-by: Nick Hill <nhill@redhat.com>
This commit is contained in:
Nick Hill
2025-07-10 07:00:20 +00:00
committed by GitHub
parent 59389c927b
commit ffbcc9e757
5 changed files with 19 additions and 16 deletions
+5 -3
View File
@@ -166,17 +166,19 @@ class CudaPlatformBase(Platform):
logger.info(
"Forcing kv cache block size to 64 for FlashMLA backend.")
compilation_config = vllm_config.compilation_config
if (envs.VLLM_ALL2ALL_BACKEND == "deepep_high_throughput"
and parallel_config.data_parallel_size > 1
and vllm_config.compilation_config.use_cudagraph):
and compilation_config.use_cudagraph):
logger.info(
"Data Parallel: Forcing enforce eager to be True 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.")
vllm_config.compilation_config.use_cudagraph = False
vllm_config.model_config.enforce_eager = True
compilation_config.use_cudagraph = False
if model_config is not None:
model_config.enforce_eager = True
# TODO (varun): Turning this ON gives incorrect results for the
# Deepseek-V2-lite model.
vllm_config.compilation_config.use_inductor = False