[Fix][Spec Decode] Fix llama4 draft loading with different quantization (#27136)

Signed-off-by: linzebing <linzebing1995@gmail.com>

vllm/model_executor/models/llama4_eagle.py

@@ -60,6 +60,10 @@ class LlamaModel(nn.Module):
             prefix=maybe_prefix(prefix, "embed_tokens"),
         )
 
+        # Temporarily modify vllm_config.quant_config for draft model layers
+        original_quant_config = vllm_config.quant_config
+        vllm_config.quant_config = quant_config
+        try:
             self.layers = nn.ModuleList(
                 [
                     Llama4DecoderLayer(
@@ -70,6 +74,9 @@ class LlamaModel(nn.Module):
                     for i in range(self.config.num_hidden_layers)
                 ]
             )
+        finally:
+            # Restore original quant_config
+            vllm_config.quant_config = original_quant_config
         self.fc = torch.nn.Linear(
             self.config.hidden_size * 2, self.config.hidden_size, bias=False
         )