[MoE Refactor] Migrate ModelOptMxFp8FusedMoE to oracle (#42768)
Signed-off-by: Bill Nell <bnell@redhat.com> Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
This commit is contained in:
@@ -371,19 +371,18 @@ class _Fp8OnlineMoEBase(OnlineMoEMethodBase):
|
||||
a1_scale = layer.w13_input_scale
|
||||
a2_scale = layer.w2_input_scale
|
||||
|
||||
quant_config = make_fp8_moe_quant_config(
|
||||
return make_fp8_moe_quant_config(
|
||||
fp8_backend=self.fp8_backend,
|
||||
w1_scale=w1_scale,
|
||||
w2_scale=w2_scale,
|
||||
a1_scale=a1_scale,
|
||||
a2_scale=a2_scale,
|
||||
w1_bias=getattr(layer, "w13_bias", None),
|
||||
w2_bias=getattr(layer, "w2_bias", None),
|
||||
block_shape=self.weight_block_size,
|
||||
swiglu_limit=getattr(layer, "swiglu_limit", None),
|
||||
)
|
||||
|
||||
self._maybe_inject_biases(quant_config, layer)
|
||||
return quant_config
|
||||
|
||||
|
||||
class Fp8PerTensorOnlineMoEMethod(_Fp8OnlineMoEBase):
|
||||
"""Online tensorwise FP8 MoE quantization.
|
||||
|
||||
Reference in New Issue
Block a user