fix kimi k2.5's moe gemm config init (#18064)
This commit is contained in:
@@ -485,7 +485,12 @@ class Scheduler(
|
|||||||
)[0]
|
)[0]
|
||||||
|
|
||||||
def init_moe_gemm_config(self):
|
def init_moe_gemm_config(self):
|
||||||
if hasattr(self.model_config.hf_config, "num_experts_per_tok"):
|
# For the MM models, check the text_config for MoE settings
|
||||||
|
config_to_check = getattr(
|
||||||
|
self.model_config.hf_config, "text_config", self.model_config.hf_config
|
||||||
|
)
|
||||||
|
|
||||||
|
if hasattr(config_to_check, "num_experts_per_tok"):
|
||||||
initialize_moe_config(self.server_args)
|
initialize_moe_config(self.server_args)
|
||||||
|
|
||||||
# Initialize GEMM-related configuration for FP8 and FP4 backends.
|
# Initialize GEMM-related configuration for FP8 and FP4 backends.
|
||||||
|
|||||||
Reference in New Issue
Block a user