Support Llama 4 for fused_marlin_moe (#20457)

Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin
2025-07-04 07:55:10 +00:00
committed by GitHub
parent 1caca5a589
commit 0e3fe896e2
6 changed files with 11 additions and 17 deletions
@@ -700,6 +700,7 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
global_scale1=layer.w13_weight_scale_2,
global_scale2=layer.w2_weight_scale_2,
quant_type_id=scalar_types.float4_e2m1f.id,
apply_router_weight_on_input=apply_router_weight_on_input,
global_num_experts=global_num_experts,
expert_map=expert_map)