[bugfix][quantization] fix quark qwen3 kv_cache quantization (#30308)

Signed-off-by: Haoyang Li <lihaoyang0109@gmail.com>

vllm/model_executor/models/qwen3_moe.py

@@ -403,6 +403,7 @@ class Qwen3MoeModel(nn.Module):
         self.padding_idx = config.pad_token_id
         self.vocab_size = config.vocab_size
         self.config = config
+        self.quant_config = quant_config
         self.embed_tokens = VocabParallelEmbedding(
             config.vocab_size,
             config.hidden_size,
@@ -505,6 +506,19 @@ class Qwen3MoeModel(nn.Module):
         loaded_params: set[str] = set()
         expert_params_mapping = self.get_expert_mapping()
         for name, loaded_weight in weights:
+            if self.quant_config is not None and (
+                scale_name := self.quant_config.get_cache_scale(name)
+            ):
+                # Loading kv cache quantization scales
+                param = params_dict[scale_name]
+                weight_loader = getattr(param, "weight_loader", default_weight_loader)
+                assert loaded_weight.numel() == 1, (
+                    f"KV scale numel {loaded_weight.numel()} != 1"
+                )
+                loaded_weight = loaded_weight.squeeze()
+                weight_loader(param, loaded_weight)
+                loaded_params.add(scale_name)
+                continue
             for param_name, weight_name, shard_id in stacked_params_mapping:
                 # Skip non-stacked layers and experts (experts handled below).
                 if weight_name not in name: