[Weight processing] Copy over new_data attributes in replace_parameter (#49601)

Signed-off-by: Felix Marty <Felix.Marty@amd.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
fxmarty-amd
2026-08-06 13:03:37 -04:00
committed by GitHub
co-authored by Claude mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent b38e111d3e
commit 81be2e09ae
10 changed files with 334 additions and 51 deletions
@@ -366,11 +366,6 @@ class GptOssMxfp4MoEMethod(FusedMoEMethodBase):
self.w13_precision_config = w13_scale
self.w2_precision_config = w2_scale
# AITER backend requires weights to be marked as shuffled.
if self.mxfp4_backend == Mxfp4MoeBackend.AITER_MXFP4_BF16:
layer.w13_weight.is_shuffled = True
layer.w2_weight.is_shuffled = True
if w13_bias is not None and w2_bias is not None:
replace_parameter(layer, "w13_bias", w13_bias)
replace_parameter(layer, "w2_bias", w2_bias)
@@ -760,14 +755,6 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
self.w13_precision_config = w13_scale
self.w2_precision_config = w2_scale
# AITER backend requires weights to be marked as shuffled.
if (
self.mxfp4_backend == Mxfp4MoeBackend.AITER_MXFP4_BF16
and not uses_triton_weight_format
):
layer.w13_weight.is_shuffled = True
layer.w2_weight.is_shuffled = True
if w13_bias is not None and w2_bias is not None:
replace_parameter(layer, "w13_bias", w13_bias)
replace_parameter(layer, "w2_bias", w2_bias)
@@ -816,6 +803,10 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
w13_scale_raw.view(-1, w13_scale_raw.shape[-1]), num_experts, guinterleave
)
w2_scale = e8m0_shuffle(w2_scale_raw.view(-1, w2_scale_raw.shape[-1]))
w13.is_shuffled = True
w2.is_shuffled = True
return w13, w2, w13_scale, w2_scale
def process_weights_after_loading(self, layer):