Support Gemma4 MoE NVFP4 (#25054)
This commit is contained in:
@@ -826,7 +826,14 @@ class Scheduler(
|
||||
self.model_config.hf_config, "text_config", self.model_config.hf_config
|
||||
)
|
||||
|
||||
if hasattr(config_to_check, "num_experts_per_tok"):
|
||||
# Different MoE architectures expose the per-token expert count under
|
||||
# different attribute names (e.g. Gemma4 uses ``top_k_experts``).
|
||||
moe_topk_attrs = (
|
||||
"num_experts_per_tok",
|
||||
"num_experts_per_token",
|
||||
"top_k_experts",
|
||||
)
|
||||
if any(hasattr(config_to_check, attr) for attr in moe_topk_attrs):
|
||||
initialize_moe_config(self.server_args)
|
||||
|
||||
# Initialize GEMM-related configuration for FP8 and FP4 backends.
|
||||
|
||||
Reference in New Issue
Block a user