[feat] Support nvfp4 quantized model of Qwen3-Next (#17627)

python/sglang/srt/models/qwen3_next.py

@@ -625,13 +625,19 @@ class Qwen3HybridAttentionDecoderLayer(nn.Module):
             dtype=torch.get_default_dtype(),  # see impl of get_rope
         )
 
+        # qkv_proj is not quantized for fp4
         self.qkv_proj = QKVParallelLinear(
             config.hidden_size,
             self.head_dim,
             self.total_num_heads * (1 + self.attn_output_gate),
             self.total_num_kv_heads,
             bias=False,
-            quant_config=quant_config,
+            quant_config=(
+                quant_config
+                if quant_config is not None
+                and quant_config.get_name() != "modelopt_fp4"
+                else None
+            ),
             tp_rank=self.attn_tp_rank,
             tp_size=self.attn_tp_size,
             prefix=add_prefix("qkv_proj", prefix),
@@ -1123,6 +1129,11 @@ class Qwen3NextForCausalLM(nn.Module):
                     # if is_pp_missing_parameter(name, self):
                     #     continue
 
+                    if name.endswith("_scale") and name not in params_dict:
+                        assert (
+                            abs(loaded_weight.item() - 1.0) < 1e-6
+                        ), f"Expected 1.0, got {loaded_weight.item()} in skipped {name}"
+                        continue
                     param = params_dict[name]
                     weight_loader = getattr(
                         param, "weight_loader", default_weight_loader