A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 9, 17:06 UTC
[Bugfix] Fix missing is_layer_skipped check for FusedMoE in AWQConfig (#32935)
Signed-off-by: jon <joninco@bullpoint.org>
joninco committed
Jan 23, 2026 at 17:19 UTC
ebd0a17e0e0b5dd326dc6533cf4d8f49806e69e1
1 file changed
+3
-2
vllm/model_executor/layers/quantization/awq.py
+3
-2
@@ -106,7 +106,7 @@ class AWQConfig(QuantizationConfig):
106
return AWQLinearMethod(self)
107
elif isinstance(layer, FusedMoE):
108
# Lazy import to avoid circular import.
109
- from .awq_marlin import AWQMarlinConfig, AWQMarlinMoEMethod
109
+ from .awq_marlin import AWQMarlinConfig
110
from .moe_wna16 import MoeWNA16Config
111
from .utils.marlin_utils import check_moe_marlin_supports_layer
112
@@ -121,6 +121,7 @@ class AWQConfig(QuantizationConfig):
121
"group_size": self.group_size,
122
"zero_point": self.zero_point,
123
"lm_head": False,
124
+ "modules_to_not_convert": self.modules_to_not_convert,
125
}
126
return MoeWNA16Config.from_config(config).get_quant_method(
127
layer, prefix
@@ -136,7 +137,7 @@ class AWQConfig(QuantizationConfig):
137
awq_marlin_config = AWQMarlinConfig.from_config(
138
marlin_compatible_config_dict
139
)
139
- return AWQMarlinMoEMethod(awq_marlin_config, layer.moe_config)
140
+ return awq_marlin_config.get_quant_method(layer, prefix)
141
return None
142
143
def apply_vllm_mapper(self, hf_to_vllm_mapper: "WeightsMapper"):