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
@@ -889,8 +889,6 @@ class Fp8MoEMethod(FusedMoEMethodBase):
elif self.use_marlin:
assert activation == "silu", (
f"{activation} not supported for Marlin MoE.")
assert not apply_router_weight_on_input, (
"Apply router weight on input not supported for Marlin MoE.")
return torch.ops.vllm.fused_marlin_moe(
x,
layer.w13_weight,
@@ -901,6 +899,7 @@ class Fp8MoEMethod(FusedMoEMethodBase):
topk_weights,
topk_ids,
quant_type_id=scalar_types.float8_e4m3fn.id,
apply_router_weight_on_input=apply_router_weight_on_input,
global_num_experts=global_num_experts,
expert_map=expert_map)
else: