[Bugfix] fix qwen3 moe fp8 accuracy issue (#23031)

Signed-off-by: Jinzhen Lin <jinzhen.ljz@antgroup.com>
2025-08-17 08:41:23 +08:00
parent bf7f470b22
commit a258ad8bcc
1 changed files with 4 additions and 0 deletions
--- a/vllm/model_executor/layers/quantization/fp8.py
+++ b/vllm/model_executor/layers/quantization/fp8.py
@ -125,6 +125,10 @@ class Fp8Config(QuantizationConfig):
        ignored_layers = cls.get_from_keys_or(config, ["ignored_layers"], None)
        weight_block_size = cls.get_from_keys_or(config, ["weight_block_size"],
                                                 None)
+        if not ignored_layers:
+            ignored_layers = cls.get_from_keys_or(config,
+                                                  ["modules_to_not_convert"],
+                                                  None)
        return cls(is_checkpoint_fp8_serialized=is_checkpoint_fp8_serialized,
                   activation_scheme=activation_scheme,
                   ignored_layers=ignored_layers,