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 11, 12:30 UTC
[Bugfix] fix _get_quant_method of FusedMoE for deepseekV3.2 on non-NV… (#30057)
Signed-off-by: tjp_zju <tanjianpingzju1990@gmail.com>
tjp_zju committed
Dec 14, 2025 at 18:20 UTC
6ecc1e411ba3e720ef85aa34bba338581bcb7f76
1 file changed
+5
vllm/model_executor/layers/quantization/moe_wna16.py
+5
@@ -17,6 +17,9 @@ from vllm.model_executor.layers.fused_moe.layer import (
17
FusedMoEMethodBase,
18
FusedMoeWeightScaleSupported,
19
)
20
+from vllm.model_executor.layers.fused_moe.unquantized_fused_moe_method import (
21
+ UnquantizedFusedMoEMethod,
22
+)
23
from vllm.model_executor.layers.linear import LinearBase, UnquantizedLinearMethod
24
from vllm.model_executor.layers.quantization import QuantizationMethods
25
from vllm.model_executor.layers.quantization.base_config import (
@@ -162,6 +165,8 @@ class MoeWNA16Config(QuantizationConfig):
165
self, layer: torch.nn.Module, prefix: str
166
) -> Optional["QuantizeMethodBase"]:
167
if is_layer_skipped_quant(prefix, self.modules_to_not_convert):
168
+ if isinstance(layer, FusedMoE):
169
+ return UnquantizedFusedMoEMethod(layer.moe_config)
170
return UnquantizedLinearMethod()
171
elif isinstance(layer, LinearBase):
172
# Avoid circular import