fix(quantization): Fix AWQ dequantize on Intel XPU and refactor AutoAWQ config (#42727)
Signed-off-by: Alex <alex.tech.lab@outlook.com> Signed-off-by: AlexHuang <jihuihuang@tencent.com> Co-authored-by: Claude <noreply@anthropic.com> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
co-authored by
Claude
Kunshang Ji
mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent
351c72d6e5
commit
8d4f54966c
@@ -8,9 +8,8 @@ import torch
|
||||
from torch.nn.parameter import Parameter
|
||||
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.quantization.auto_awq import AutoAWQConfig
|
||||
from vllm.model_executor.layers.quantization.auto_gptq import AutoGPTQConfig
|
||||
from vllm.model_executor.layers.quantization.awq import AWQConfig
|
||||
from vllm.model_executor.layers.quantization.awq_marlin import AWQMarlinConfig
|
||||
from vllm.model_executor.layers.quantization.utils.marlin_utils import (
|
||||
check_marlin_supported,
|
||||
)
|
||||
@@ -125,12 +124,12 @@ class INCWNA16LinearScheme(INCLinearScheme):
|
||||
)
|
||||
|
||||
if use_marlin:
|
||||
from vllm.model_executor.layers.quantization.awq_marlin import (
|
||||
AWQMarlinLinearMethod,
|
||||
from vllm.model_executor.layers.quantization.auto_awq import (
|
||||
AutoAWQMarlinLinearMethod,
|
||||
)
|
||||
|
||||
return AWQMarlinLinearMethod(
|
||||
AWQMarlinConfig(
|
||||
return AutoAWQMarlinLinearMethod(
|
||||
AutoAWQConfig(
|
||||
weight_bits=self.layer_config.bits,
|
||||
group_size=self.layer_config.group_size,
|
||||
zero_point=not self.layer_config.sym,
|
||||
@@ -140,13 +139,16 @@ class INCWNA16LinearScheme(INCLinearScheme):
|
||||
)
|
||||
)
|
||||
|
||||
from vllm.model_executor.layers.quantization.awq import AWQLinearMethod
|
||||
from vllm.model_executor.layers.quantization.auto_awq import (
|
||||
AutoAWQLinearMethod,
|
||||
)
|
||||
|
||||
return AWQLinearMethod(
|
||||
AWQConfig(
|
||||
return AutoAWQLinearMethod(
|
||||
AutoAWQConfig(
|
||||
weight_bits=self.layer_config.bits,
|
||||
group_size=self.layer_config.group_size,
|
||||
zero_point=not self.layer_config.sym,
|
||||
lm_head_quantized=False,
|
||||
)
|
||||
)
|
||||
|
||||
|
||||
@@ -4,8 +4,8 @@
|
||||
from typing import TYPE_CHECKING
|
||||
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.quantization.auto_awq import AutoAWQConfig
|
||||
from vllm.model_executor.layers.quantization.auto_gptq import AutoGPTQConfig
|
||||
from vllm.model_executor.layers.quantization.awq_marlin import AWQMarlinConfig
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.scalar_type import scalar_types
|
||||
|
||||
@@ -154,7 +154,7 @@ def _resolve_gptq_moe(layer: "torch.nn.Module", layer_config: "INCLayerConfig"):
|
||||
|
||||
|
||||
def _resolve_awq_moe(layer: "torch.nn.Module", layer_config: "INCLayerConfig"):
|
||||
from vllm.model_executor.layers.quantization.awq_marlin import AWQMarlinMoEMethod
|
||||
from vllm.model_executor.layers.quantization.auto_awq import AutoAWQMoEMethod
|
||||
from vllm.model_executor.layers.quantization.moe_wna16 import (
|
||||
MoeWNA16Config,
|
||||
MoeWNA16Method,
|
||||
@@ -177,8 +177,8 @@ def _resolve_awq_moe(layer: "torch.nn.Module", layer_config: "INCLayerConfig"):
|
||||
) and check_moe_marlin_supports_layer(layer, layer_config.group_size)
|
||||
|
||||
if use_marlin:
|
||||
return AWQMarlinMoEMethod(
|
||||
AWQMarlinConfig(
|
||||
return AutoAWQMoEMethod(
|
||||
AutoAWQConfig(
|
||||
weight_bits=layer_config.bits,
|
||||
group_size=layer_config.group_size,
|
||||
zero_point=not layer_config.sym,
|
||||
|
||||
Reference in New Issue
Block a user