Update llama torch model + llm api pytorch test

Signed-off-by: jintaop <jintaop@nvidia.com>

Author: JintaoPengCS

tensorrt_llm/_torch/models/modeling_llama.py

@@ -635,7 +635,7 @@ def __init__(
         )
         self.is_nvfp4 = self.is_quanted and model_config.quant_config.quant_mode.has_nvfp4(
         )
-
+        # Self Attention
         self.self_attn = LlamaAttention(
             model_config,
             layer_idx=layer_idx,

tests/integration/defs/accuracy/test_llm_api_pytorch.py

@@ -95,6 +95,19 @@ def test_nvfp4(self):
             task.evaluate(llm)
             task = MMLU(self.MODEL_NAME)
             task.evaluate(llm)
+            
+    def test_nvfp4_with_norm_quant(self, monkeypatch):
+        model_path = f"{llm_models_root()}/nvfp4-quantized/Meta-Llama-3.1-8B"
+        with LLM(model_path) as llm:
+            sm_version = get_sm_version()
+            if sm_version not in (100, 103):
+                pytest.skip(f"test_nvfp4_with_norm_quant supports SM 100 and 103 only")
+            monkeypatch.setenv("TRTLLM_DISABLE_NVFP4_LAYERNORM_FUSION", "0")
+            assert llm.args.quant_config.quant_algo == QuantAlgo.NVFP4
+            task = CnnDailymail(self.MODEL_NAME)
+            task.evaluate(llm)
+            task = MMLU(self.MODEL_NAME)
+            task.evaluate(llm)
 
     @skip_pre_blackwell
     @pytest.mark.parametrize("stream_interval", [4, 64],