Applying scales rename to fp8 config

vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py

@@ -425,6 +425,10 @@ def get_cache_scale(self, name: str) -> Optional[str]:
             return name.replace(".k_proj.output_scale", ".attn.k_scale")
         if name.endswith(".output_scale") and ".v_proj" in name:
             return name.replace(".v_proj.output_scale", ".attn.v_scale")
+        if name.endswith(".output_scale") and ".q_proj" in name:
+            return name.replace(".q_proj.output_scale", ".attn.q_scale")
+        if name.endswith("self_attn.prob_output_scale"):
+            return name.replace(".prob_output_scale", ".attn.prob_scale")
         # If no matches, return None
         return None
 

vllm/model_executor/layers/quantization/fp8.py

@@ -113,6 +113,26 @@ def get_quant_method(self, layer: torch.nn.Module,
             return Fp8KVCacheMethod(self)
         return None
 
+    def get_cache_scale(self, name: str) -> Optional[str]:
+        """
+        Check whether the param name matches the format for k/v cache scales
+        in compressed-tensors. If this is the case, return its equivalent
+        param name expected by vLLM
+
+        :param name: param name
+        :return: matching param name for KV cache scale in vLLM
+        """
+        if name.endswith(".output_scale") and ".k_proj" in name:
+            return name.replace(".k_proj.output_scale", ".attn.k_scale")
+        if name.endswith(".output_scale") and ".v_proj" in name:
+            return name.replace(".v_proj.output_scale", ".attn.v_scale")
+        if name.endswith(".output_scale") and ".q_proj" in name:
+            return name.replace(".q_proj.output_scale", ".attn.q_scale")
+        if name.endswith("self_attn.prob_output_scale"):
+            return name.replace(".prob_output_scale", ".attn.prob_scale")
+        # If no matches, return None
+        return None
+
 
 class Fp8LinearMethod(LinearMethodBase):
     """Linear method for FP8.