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] GPTQ quantization compatibility for Qwen3 Next MOE models (AutoGPTQ and AutoRound-GPTQ) (#25268)
Signed-off-by: JartX <sagformas@epdcenter.es>
JartX committed
Sep 20, 2025 at 05:18 UTC
36429096171ff8785645c40c662d859dddedd829
1 file changed
+5
-3
vllm/model_executor/models/qwen3_next.py
+5
-3
@@ -148,9 +148,11 @@ class Qwen3NextSparseMoeBlock(nn.Module):
148
149
def _maybe_ignore_quant_config(self, quant_config: QuantizationConfig):
150
# GPTQ configs do not have a list of ignored modules, however AutoGPTQ
151
- # seems to avoid gate quantization.
152
- # See: https://huggingface.co/Qwen/Qwen3-30B-A3B-GPTQ-Int4
153
- if isinstance(quant_config, (GPTQConfig, GPTQMarlinConfig)):
151
+ # seems to avoid gate quantization while AutoRound does.
152
+ if isinstance(
153
+ quant_config,
154
+ (GPTQConfig,
155
+ GPTQMarlinConfig)) and not quant_config.autoround_version:
156
return None
157
return quant_config
158