[ROCm][Quant] Minimax-M3: Enable fp8_per_channel for bf16 weights on mi300x (#45854)
Signed-off-by: Hongxia Yang <hongxia.yang@amd.com> Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com> Co-authored-by: tjtanaa <tunjian.tan@embeddedllm.com>
This commit is contained in:
@@ -787,6 +787,8 @@ class Fp8MoEMethod(FusedMoEMethodBase):
|
||||
a2_scale=a2_scale,
|
||||
block_shape=self.weight_block_size,
|
||||
swiglu_limit=getattr(layer, "swiglu_limit", None),
|
||||
gemm1_alpha=getattr(layer, "swiglu_alpha", None),
|
||||
gemm1_beta=getattr(layer, "swiglu_beta", None),
|
||||
)
|
||||
|
||||
# Inject biases into the quant config if the model has them
|
||||
|
||||
@@ -483,6 +483,8 @@ class _Fp8OnlineMoEBase(OnlineMoEMethodBase):
|
||||
per_act_token_quant=self.per_act_token_quant,
|
||||
per_out_ch_quant=self.per_out_ch_quant,
|
||||
swiglu_limit=getattr(layer, "swiglu_limit", None),
|
||||
gemm1_alpha=getattr(layer, "swiglu_alpha", None),
|
||||
gemm1_beta=getattr(layer, "swiglu_beta", None),
|
||||
)
|
||||
|
||||
|
||||
|
||||
Reference in New Issue
Block a user