Fix stage_activation: use Triton staging kernel from vLLM patch instead of fp4_utils

src/nvfp4_megamoe_kernel/nvfp4_mega_moe.py

@@ -161,13 +161,13 @@ def nvfp4_mega_moe_l2(
 
 def stage_activation(x_bf16):
     """Quantize BF16 activation to FP4 (E2M1) with UE4M3 block16 scales.
-
-    This replaces the Triton staging kernel from patches/staging_kernel.py.
+    
+    Uses the Triton staging kernel from the vLLM deepseek_v4 patch.
     """
-    from vllm.model_executor.layers.quantization.utils.fp4_utils import (
-        per_token_cast_to_fp4,
+    from vllm.model_executor.models.staging_kernel import (
+        _stage_deepseek_v4_mega_moe_inputs,
     )
-    return per_token_cast_to_fp4(x_bf16)
+    return _stage_deepseek_v4_mega_moe_inputs(x_bf16)
 
 
 def nvfp4_mega_moe_full(