[Bugfix] Fix topk_ids indices_type for CUTLASS w8a8 FP8 MoE (#20166)
Signed-off-by: Ming Yang <yming@meta.com>
This commit is contained in:
+9
-5
@@ -929,9 +929,12 @@ class CompressedTensorsW8A8Fp8MoECutlassMethod(CompressedTensorsMoEMethod):
|
||||
num_expert_group=num_expert_group,
|
||||
custom_routing_function=custom_routing_function,
|
||||
scoring_func=scoring_func,
|
||||
e_score_correction_bias=e_score_correction_bias,
|
||||
indices_type=self.topk_indices_dtype,
|
||||
)
|
||||
e_score_correction_bias=e_score_correction_bias)
|
||||
|
||||
a1_scale = layer.w13_input_scale
|
||||
a2_scale = layer.w2_input_scale
|
||||
per_act_token = a1_scale.numel() != 1 if a1_scale is not None else (
|
||||
a2_scale.numel() != 1 if a2_scale is not None else False)
|
||||
|
||||
return self.fused_experts(
|
||||
x,
|
||||
@@ -939,13 +942,14 @@ class CompressedTensorsW8A8Fp8MoECutlassMethod(CompressedTensorsMoEMethod):
|
||||
layer.w2_weight,
|
||||
topk_weights,
|
||||
topk_ids,
|
||||
per_act_token=per_act_token,
|
||||
activation=activation,
|
||||
global_num_experts=global_num_experts,
|
||||
expert_map=None if self.disable_expert_map else expert_map,
|
||||
w1_scale=layer.w13_weight_scale,
|
||||
w2_scale=layer.w2_weight_scale,
|
||||
a1_scale=layer.w13_input_scale,
|
||||
a2_scale=layer.w2_input_scale,
|
||||
a1_scale=a1_scale,
|
||||
a2_scale=a2_scale,
|
||||
)
|
||||
|
||||
|
||||
|
||||
Reference in New Issue
Block a user