[Kernel] optimize moe_align_block_size for cuda graph and large num_experts (e.g. DeepSeek-V3) (#12222)

Signed-off-by: Jinzhen Lin <linjinzhen@hotmail.com>
Co-authored-by: Michael Goin <mgoin@redhat.com>
Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>
This commit is contained in:
Jinzhen Lin
2025-01-20 16:42:16 -08:00
committed by GitHub
co-authored by Michael Goin Tyler Michael Smith
parent 06a760d6e8
commit 750f4cabfa
2 changed files with 58 additions and 37 deletions
+1 -1
View File
@@ -607,7 +607,7 @@ class ModelConfig:
self.max_seq_len_to_capture = min(self.max_seq_len_to_capture,
self.max_model_len)
MODEL_NOT_SUPPORT_CUDA_GRAPH = ['deepseek_v3', 'mllama']
MODEL_NOT_SUPPORT_CUDA_GRAPH = ['mllama']
if (self.hf_config.model_type in MODEL_NOT_SUPPORT_CUDA_GRAPH
and not self.enforce_eager):
logger.warning(