[Bugfix][Quantization] Give LM heads standard linear metadata (#58444)

Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Misha Goin
2026-09-24 16:39:49 -07:00
committed by GitHub
parent 9ac3b7c69f
commit 4f72bd0152
8 changed files with 93 additions and 70 deletions
@@ -119,8 +119,6 @@ class CompressedTensorsWNA16(CompressedTensorsScheme):
layer.output_size_per_partition = output_size_per_partition
layer.output_partition_sizes = output_partition_sizes
layer.params_dtype = params_dtype
if not hasattr(layer, "has_bias"):
layer.has_bias = False
mp_linear_kernel_config = MPLinearLayerConfig(
full_weight_shape=(input_size, output_size),
@@ -147,8 +147,6 @@ class CompressedTensorsWNA4Int(CompressedTensorsScheme):
layer.output_size_per_partition = output_size_per_partition
layer.output_partition_sizes = output_partition_sizes
layer.params_dtype = params_dtype
if not hasattr(layer, "has_bias"):
layer.has_bias = False
mp_config = MPLinearLayerConfig(
full_weight_shape=(input_size, output_size),
@@ -147,8 +147,6 @@ class CompressedTensorsWNA8Int(CompressedTensorsScheme):
layer.output_size_per_partition = output_size_per_partition
layer.output_partition_sizes = output_partition_sizes
layer.params_dtype = params_dtype
if not hasattr(layer, "has_bias"):
layer.has_bias = False
mp_config = MPLinearLayerConfig(
full_weight_shape=(input_size, output_size),
@@ -97,12 +97,8 @@ class CompressedTensorsWNA8O8Int(CompressedTensorsScheme):
output_size_per_partition = sum(output_partition_sizes)
layer.input_size_per_partition = input_size_per_partition
layer.output_size_per_partition = output_size_per_partition
# Set for kernels' weight prep; also covers ParallelLMHead, which does
# not set these in __init__.
layer.output_partition_sizes = output_partition_sizes
layer.params_dtype = params_dtype
if not hasattr(layer, "has_bias"):
layer.has_bias = False
mp_config = MPLinearLayerConfig(
full_weight_shape=(input_size, output_size),
@@ -2508,12 +2508,6 @@ class ModelOptLinearMethod(LinearMethodBase):
layer.logical_widths = output_partition_sizes
layer.input_size_per_partition = input_size_per_partition
layer.output_size_per_partition = sum(output_partition_sizes)
# Humming reads both off the layer in
# prepare_humming_linear_layer_config. LinearBase sets them itself;
# ParallelLMHead does not, so supply them here.
layer.output_partition_sizes = output_partition_sizes
if not hasattr(layer, "has_bias"):
layer.has_bias = getattr(layer, "bias", None) is not None
shapes = Shapes(output_partition_sizes, input_size_per_partition, params_dtype)
self.wkey.create_weights(layer, WEIGHT, self.ctx, shapes, weight_loader)