xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-10 00:15:51 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Misc] add ignore mapper for quark quantization (#28275)

Browse Source 
Signed-off-by: Haoyang Li <lihaoyang0109@gmail.com>
This commit is contained in:
haoyangli-amd 2025-11-14 13:56:35 +08:00 committed by GitHub
parent 0aecd9138f
commit 0b25498990
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 9 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
vllm/model_executor/layers/quantization/quark/quark.py
View File

@ -2,7 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import fnmatch import fnmatch
from typing import Any, Optional, cast from typing import TYPE_CHECKING, Any, Optional, cast
import torch import torch
@ -34,6 +34,9 @@ from vllm.model_executor.layers.quantization.quark.utils import (
) )
from vllm.platforms import current_platform from vllm.platforms import current_platform
if TYPE_CHECKING:
from vllm.model_executor.models.utils import WeightsMapper
__all__ = ["QuarkLinearMethod"] __all__ = ["QuarkLinearMethod"]
logger = init_logger(__name__) logger = init_logger(__name__)
@ -54,6 +57,7 @@ class QuarkConfig(QuantizationConfig):
self.kv_cache_group = kv_cache_group self.kv_cache_group = kv_cache_group
self.kv_cache_config = kv_cache_config self.kv_cache_config = kv_cache_config
self.pack_method = pack_method self.pack_method = pack_method
self.ignore: list[str] = cast(list[str], self.quant_config.get("exclude", []))
def get_linear_method(self) -> "QuarkLinearMethod": def get_linear_method(self) -> "QuarkLinearMethod":
return QuarkLinearMethod(self) return QuarkLinearMethod(self)
@ -74,9 +78,8 @@ class QuarkConfig(QuantizationConfig):
from vllm.attention.layer import Attention # Avoid circular import from vllm.attention.layer import Attention # Avoid circular import
# Check if the layer is skipped for quantization. # Check if the layer is skipped for quantization.
exclude_layers = cast(list[str], self.quant_config.get("exclude"))
if should_ignore_layer( if should_ignore_layer(
prefix, ignore=exclude_layers, fused_mapping=self.packed_modules_mapping prefix, ignore=self.ignore, fused_mapping=self.packed_modules_mapping
): ):
return UnquantizedLinearMethod() return UnquantizedLinearMethod()
if isinstance(layer, LinearBase): if isinstance(layer, LinearBase):
@ -90,6 +93,9 @@ class QuarkConfig(QuantizationConfig):
return QuarkMoEMethod.get_moe_method(self, module=layer, layer_name=prefix) return QuarkMoEMethod.get_moe_method(self, module=layer, layer_name=prefix)
return None return None
def apply_vllm_mapper(self, hf_to_vllm_mapper: "WeightsMapper"):
self.ignore = hf_to_vllm_mapper.apply_list(self.ignore)
@classmethod @classmethod
def from_config(cls, config: dict[str, Any]) -> "QuarkConfig": def from_config(cls, config: dict[str, Any]) -> "QuarkConfig":
export_config = config.get("export") export_config = config.get("export")