[Bugfix] Fix MiniMax-M3 compressed-tensors FP8 MoE SwiGLU params (#46845)
Signed-off-by: Tan Pin Siang <tanpinsiang@gmail.com>
This commit is contained in:
+2
@@ -352,6 +352,8 @@ class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
|
||||
per_act_token_quant=is_per_token,
|
||||
per_out_ch_quant=is_per_token,
|
||||
block_shape=self.weight_block_size,
|
||||
gemm1_alpha=getattr(layer, "swiglu_alpha", None),
|
||||
gemm1_beta=getattr(layer, "swiglu_beta", None),
|
||||
swiglu_limit=getattr(layer, "swiglu_limit", None),
|
||||
layer=layer,
|
||||
)
|
||||
|
||||
Reference in New Issue
Block a user