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 14, 05:30 UTC
[Bugfix] Fix AWQ marlin layer skipping (#27416)
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Isotr0py committed
Oct 24, 2025 at 02:30 UTC
81d5bb765a33792dc307f160ff65e73398e3b415
1 file changed
+9
-2
vllm/model_executor/layers/quantization/awq_marlin.py
+9
-2
@@ -178,7 +178,10 @@ class AWQMarlinConfig(QuantizationConfig):
178
isinstance(layer, ParallelLMHead) and self.lm_head_quantized
179
):
180
if is_layer_skipped(
181
- prefix, self.modules_to_not_convert, self.packed_modules_mapping
181
+ prefix,
182
+ self.modules_to_not_convert,
183
+ self.packed_modules_mapping,
184
+ skip_with_substr=True,
185
):
186
return UnquantizedLinearMethod()
187
# Check if the layer is supported by AWQMarlin.
@@ -194,7 +197,11 @@ class AWQMarlinConfig(QuantizationConfig):
197
elif isinstance(layer, FusedMoE):
198
from vllm.model_executor.layers.quantization.moe_wna16 import MoeWNA16Config
199
197
- if is_layer_skipped(prefix, getattr(self, "modules_to_not_convert", [])):
200
+ if is_layer_skipped(
201
+ prefix,
202
+ getattr(self, "modules_to_not_convert", []),
203
+ skip_with_substr=True,
204
+ ):
205
return UnquantizedFusedMoEMethod(layer.moe_config)
206
if not check_moe_marlin_supports_layer(layer, self.group_size):
207
logger.warning_once(