[V1] EP/TP MoE + DP Attention (#13931)
This commit is contained in:
@@ -111,6 +111,7 @@ 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
|
||||
|
||||
if parallel_config.worker_cls == "auto":
|
||||
if scheduler_config.is_multi_step:
|
||||
@@ -150,6 +151,14 @@ class CudaPlatformBase(Platform):
|
||||
"FlashMLA: Forcing kv cache block size to 64 since this"
|
||||
" is currently the only block size supported by the kernel.")
|
||||
|
||||
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
|
||||
|
||||
@classmethod
|
||||
def get_current_memory_usage(cls,
|
||||
device: Optional[torch.types.Device] = None
|
||||
|
||||
Reference in New Issue
Block a user