[Bugfix] Fix activation quantization dispatch for WNA4Int/WNA8Int (#48785)

Signed-off-by: HDCharles <charlesdavidhernandez@gmail.com>
This commit is contained in:
HDCharles
2026-07-16 17:13:02 -04:00
committed by GitHub
parent fb5ec0dc9e
commit ab3c1aedf3
3 changed files with 13 additions and 6 deletions
@@ -96,7 +96,7 @@ class CompressedTensorsWNA4Int(CompressedTensorsScheme):
return 75
def _build_input_quant_config(self) -> dict | None:
"""Build the config dict that HummingInputSchema.from_config expects."""
"""Build the config dict that BaseInputSchema.from_config expects."""
if self.input_quant is None:
return None
iq = self.input_quant
@@ -112,7 +112,7 @@ class CompressedTensorsWNA4Int(CompressedTensorsScheme):
"dynamic": iq.dynamic,
"group_size": iq.group_size or 0,
"quant_method": "compressed-tensors",
"format": self.quant_format,
"format": "int-quantized",
}
def create_weights(
@@ -96,7 +96,7 @@ class CompressedTensorsWNA8Int(CompressedTensorsScheme):
return 75
def _build_input_quant_config(self) -> dict | None:
"""Build the config dict that HummingInputSchema.from_config expects."""
"""Build the config dict that BaseInputSchema.from_config expects."""
if self.input_quant is None:
return None
iq = self.input_quant
@@ -112,7 +112,7 @@ class CompressedTensorsWNA8Int(CompressedTensorsScheme):
"dynamic": iq.dynamic,
"group_size": iq.group_size or 0,
"quant_method": "compressed-tensors",
"format": self.quant_format,
"format": "int-quantized",
}
def create_weights(
@@ -441,6 +441,7 @@ def prepare_humming_layer(
input_quant_config: dict | None = None,
):
from vllm.utils.humming import (
BaseInputSchema,
BaseWeightSchema,
HummingInputSchema,
HummingMethod,
@@ -448,7 +449,7 @@ def prepare_humming_layer(
weight_schema = BaseWeightSchema.from_config(quant_config)
if input_quant_config is not None:
input_schema = HummingInputSchema.from_config(input_quant_config)
input_schema = BaseInputSchema.from_config(input_quant_config)
else:
input_schema = HummingInputSchema()
@@ -463,13 +464,19 @@ def prepare_humming_layer(
shape_k_stacks = [input_size_per_partition]
shape_n_stacks = layer.output_partition_sizes
# Step 1: convert weight to humming standard format
# Step 1: convert weight and input schemas to humming standard format
weight_schema, tensors = weight_schema.convert_humming(
tensors=dict(layer.named_parameters()),
shape_n_stacks=shape_n_stacks,
shape_k_stacks=shape_k_stacks,
param_dtype=layer.params_dtype,
)
input_schema, _ = input_schema.convert_humming(
tensors={},
shape_n_stacks=shape_n_stacks,
shape_k_stacks=shape_k_stacks,
param_dtype=layer.params_dtype,
)
layer.weight_schema = weight_schema