[Core] Enable CUDA graphs for DP + All2All kernels (#18724)

Signed-off-by: Varun Sundar Rabindranath <varun@neuralmagic.com>
Co-authored-by: Varun Sundar Rabindranath <varun@neuralmagic.com>
This commit is contained in:
Varun Sundar Rabindranath
2025-05-28 22:55:30 +00:00
committed by GitHub
co-authored by Varun Sundar Rabindranath
parent 6dbe5b5c93
commit 7951d78738
4 changed files with 100 additions and 37 deletions
-11
View File
@@ -106,7 +106,6 @@ class CudaPlatformBase(Platform):
def check_and_update_config(cls, vllm_config: "VllmConfig") -> None:
parallel_config = vllm_config.parallel_config
scheduler_config = vllm_config.scheduler_config
compilation_config = vllm_config.compilation_config
model_config = vllm_config.model_config
if parallel_config.worker_cls == "auto":
@@ -154,16 +153,6 @@ class CudaPlatformBase(Platform):
logger.info(
"Forcing kv cache block size to 64 for FlashMLA backend.")
if (parallel_config.data_parallel_size > 1
and compilation_config.use_cudagraph):
logger.info(
"Data Parallel: Forcing enforce eager to be True since DP is "
"currently not supported with CUDA Graphs.")
vllm_config.model_config.enforce_eager = True
compilation_config.use_cudagraph = False
# FIXME: inductor breaks cudagraph (from @bnell)
compilation_config.use_inductor = False
@classmethod
def get_current_memory_usage(cls,
device: Optional[torch.types.Device] = None