Add ModelOpt Qwen3 nvfp4 support (#20101)

Signed-off-by: Zhiyu Cheng <zhiyuc@nvidia.com>

vllm/model_executor/models/qwen3_moe.py

@@ -48,7 +48,8 @@ from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead, VocabParallelEmbedding)
-from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+from vllm.model_executor.model_loader.weight_utils import (
+    default_weight_loader, maybe_remap_kv_scale_name)
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.sequence import IntermediateTensors
 
@@ -471,12 +472,21 @@ class Qwen3MoeModel(nn.Module):
                 # Skip layers on other devices.
                 if is_pp_missing_parameter(name, self):
                     continue
+                if name.endswith("scale"):
+                    # Remapping the name of FP8 kv-scale.
+                    name = maybe_remap_kv_scale_name(name, params_dict)
+                    if name is None:
+                        continue
                 if name not in params_dict:
                     continue
 
                 param = params_dict[name]
-                weight_loader = param.weight_loader
-                weight_loader(param, loaded_weight, shard_id)
+                weight_loader = getattr(param, "weight_loader",
+                                        default_weight_loader)
+                if weight_loader == default_weight_loader:
+                    weight_loader(param, loaded_weight)
+                else:
+                    weight_loader(param, loaded_weight, shard_id)
                 break
             else:
                 is_expert_weight = False