[TRTLLM-8638][fix] fix test issues (NVIDIA#8557)

Signed-off-by: Xin He (SW-GPU) <200704525+xinhe-nv@users.noreply.github.com>

Author: xinhe-nv

tests/integration/defs/.test_durations

@@ -570,7 +570,7 @@
     "examples/test_exaone.py::test_llm_exaone_2gpu[exaone_3.0_7.8b_instruct-float16-nb:1]": 7155.35844087804435,
     "examples/test_gemma.py::test_hf_gemma_fp8_base_bf16_multi_lora[gemma-2-27b-it]": 317.7816583644599,
     "examples/test_gemma.py::test_hf_gemma_fp8_base_bf16_multi_lora[gemma-2-9b-it]": 317.7816583644599,
-    "examples/test_gemma.py::test_llm_gemma_1gpu_summary[gemma-2-27b-it-other-bfloat16-8]": 1620.1766637390829,
+    "examples/test_gemma.py::test_llm_gemma_1gpu_summary[gemma-2-27b-it-other-bfloat16-8]": 7020.1766637390829,
     "examples/test_gemma.py::test_llm_gemma_1gpu_summary_vswa[gemma-3-1b-it-other-bfloat16-8]": 195.3050664511975,
     "examples/test_gemma.py::test_llm_hf_gemma_quantization_1gpu[gemma-2-27b-it-fp8-bfloat16-8]": 317.7816583644599,
     "examples/test_gpt.py::test_llm_gpt2_medium_1gpu[non_streaming-use_cpp_session-enable_gemm_plugin]": 114.20040711760521,

tests/integration/defs/accuracy/test_llm_api_pytorch.py

@@ -2219,6 +2219,8 @@ def test_nvfp4_multi_gpus_corner_case(self):
 
     @pytest.mark.skip_less_mpi_world_size(8)
     @skip_pre_hopper
+    @pytest.mark.skipif(get_sm_version() >= 100,
+                        reason="https://nvbugs/5547584 WNF")
     @pytest.mark.skip_less_device_memory(140000)
     @pytest.mark.parametrize(
         "tp_size,pp_size,ep_size,mtp_nextn,fp8kv,attention_dp,cuda_graph,overlap_scheduler,max_batch_size,moe_backend",
@@ -2272,6 +2274,7 @@ def test_fp8_blockscale(self, tp_size, pp_size, ep_size, mtp_nextn, fp8kv,
 
     @pytest.mark.skip_less_mpi_world_size(8)
     @skip_pre_hopper
+    @pytest.mark.skip_less_device_memory(140000)
     @pytest.mark.parametrize(
         "tp_size,pp_size,ep_size,mtp_nextn,fp8kv,attention_dp,cuda_graph,overlap_scheduler,max_batch_size",
         [(8, 1, 4, 3, False, False, True, True, 1),
@@ -3573,6 +3576,7 @@ def test_w4_chunked_prefill(self, kv_cache_dtype, moe_backend, mocker):
                           extra_evaluator_kwargs=self.extra_evaluator_kwargs)
 
 
+@skip_pre_hopper
 class TestEXAONE4(LlmapiAccuracyTestHarness):
     MODEL_NAME = "LGAI-EXAONE/EXAONE-4.0-32B"
     kv_cache_config = KvCacheConfig(enable_block_reuse=False,

tests/integration/defs/examples/test_gemma.py

@@ -210,7 +210,7 @@ def test_llm_gemma_1gpu_summary_vswa(batch_size, data_type, gemma_model_root,
                        max_attention_window)
 
 
-@pytest.mark.timeout(5400)
+@pytest.mark.timeout(7200)
 @pytest.mark.parametrize("batch_size", [8])
 @pytest.mark.parametrize("data_type", ['float16', 'bfloat16'])
 @pytest.mark.parametrize("test_case", [

tests/integration/defs/test_e2e.py

@@ -3534,23 +3534,31 @@ def test_ptp_quickstart_advanced_llama_multi_nodes(llm_root, llm_venv,
                  marks=skip_pre_hopper),
 ])
 def test_multi_nodes_eval(llm_venv, model_path, tp_size, pp_size, ep_size,
-                          eval_task):
+                          eval_task, mmlu_dataset_root):
     if "Llama-4" in model_path and tp_size == 16:
         pytest.skip("Llama-4 with tp16 is not supported")
 
     mmlu_threshold = 81.5
+    model_dir = f"{llm_models_root()}/{model_path}"
     run_cmd = [
         "trtllm-llmapi-launch",
         "trtllm-eval",
-        f"--model={llm_models_root()}/{model_path}",
+        f"--model={model_dir}",
         f"--ep_size={ep_size}",
         f"--tp_size={tp_size}",
         f"--pp_size={pp_size}",
         f"--kv_cache_free_gpu_memory_fraction={_MEM_FRACTION_80}",
         "--max_batch_size=32",
-        eval_task,
+        "--backend=pytorch",
     ]
 
+    if "Kimi" in model_path:
+        run_cmd.append("--trust_remote_code")
+    else:
+        run_cmd.append(f"--tokenizer={model_dir}")
+
+    run_cmd.extend([eval_task, f"--dataset_path={mmlu_dataset_root}"])
+
     llm_venv._new_env["TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL"] = "1"
     output = check_output(" ".join(run_cmd), shell=True, env=llm_venv._new_env)