[Bugfix] Fix layer skip logic with bitsandbytes (vllm-project#9887)

Signed-off-by: mgoin <[email protected]>
Signed-off-by: Loc Huynh <[email protected]>

Author: mgoin

vllm/model_executor/layers/quantization/bitsandbytes.py

@@ -119,7 +119,12 @@ def get_scaled_act_names(self) -> List[str]:
 
 
 def is_layer_skipped_bnb(prefix: str, llm_int8_skip_modules: List[str]):
-    return any(module_name in prefix for module_name in llm_int8_skip_modules)
+    # Split the prefix into its dot-separated components
+    components = prefix.split('.')
+
+    # Check if any of the skip modules exactly matches any component
+    return any(module_name in components
+               for module_name in llm_int8_skip_modules)
 
 
 class BitsAndBytesLinearMethod(LinearMethodBase):