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 qwen3 moe fp8 accuracy issue (#23031)
Signed-off-by: Jinzhen Lin <jinzhen.ljz@antgroup.com>
Jinzhen Lin committed
Aug 17, 2025 at 08:41 UTC
a258ad8bcc0014c04d11a9bc8c6591b379c31b68
1 file changed
+4
vllm/model_executor/layers/quantization/fp8.py
+4
@@ -125,6 +125,10 @@ class Fp8Config(QuantizationConfig):
125
ignored_layers = cls.get_from_keys_or(config, ["ignored_layers"], None)
126
weight_block_size = cls.get_from_keys_or(config, ["weight_block_size"],
127
None)
128
+ if not ignored_layers:
129
+ ignored_layers = cls.get_from_keys_or(config,
130
+ ["modules_to_not_convert"],
131
+ None)
132
return cls(is_checkpoint_fp8_serialized=is_checkpoint_fp8_serialized,
133
activation_scheme=activation_scheme,
134
ignored_layers=ignored_layers,