[None][chore] Add failed cases into waives.txt (NVIDIA#7841)

Signed-off-by: xinhe-nv <[email protected]>
Signed-off-by: Xin He (SW-GPU) <[email protected]>

Author: xinhe-nv

tests/integration/defs/accuracy/test_llm_api_pytorch.py

@@ -2010,42 +2010,7 @@ def test_nvfp4_multi_gpus_chunked_prefill(self, tp_size, pp_size, ep_size,
             assert llm.args.moe_config.backend == moe_backend
             assert llm.args.quant_config.quant_algo == QuantAlgo.NVFP4
 
-    def test_nvfp4_multi_gpus_corner_case(self):
-        """
-        This test is used to test the corner case of the NVFP4 model.
-        When using the same value for max_seq_len and max_num_tokens, there will be no
-        enough kv block for the dummy requests in CUDA graph warmup when creating
-        the py_executor before estimating kv cache. Then CUDA graph capture will be
-        triggered when estimating kv cache. This may cause some errors.
-        More info in https://nvbugs/5485325.
-        """
-        kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.80,
-                                        dtype="fp8",
-                                        enable_block_reuse=False)
-        pytorch_config = dict(disable_overlap_scheduler=False,
-                              cuda_graph_config=CudaGraphConfig(
-                                  enable_padding=True, max_batch_size=1024),
-                              moe_config=MoeConfig(backend="TRTLLM"))
-
-        mtp_config = MTPDecodingConfig(num_nextn_predict_layers=1)
-        with LLM(f"{llm_models_root()}/DeepSeek-R1/DeepSeek-R1-FP4",
-                 tensor_parallel_size=8,
-                 pipeline_parallel_size=1,
-                 moe_expert_parallel_size=8,
-                 kv_cache_config=kv_cache_config,
-                 **pytorch_config,
-                 enable_attention_dp=False,
-                 speculative_config=mtp_config,
-                 max_seq_len=5120,
-                 max_num_tokens=5120) as llm:
-
-            assert llm.args.quant_config.quant_algo == QuantAlgo.NVFP4
-
-            task = MMLU(self.MODEL_NAME)
-            task.evaluate(llm)
-            task = GSM8K(self.MODEL_NAME)
-            task.evaluate(llm)
-
+    @skip_pre_blackwell
     def test_nvfp4_multi_gpus_corner_case(self):
         """
         This test is used to test the corner case of the NVFP4 model.

tests/integration/defs/test_e2e.py

@@ -1735,8 +1735,10 @@ def test_openai_multinodes_chat_tp8pp2(llm_root, llm_venv):
 
 
 @pytest.mark.skip_less_device_memory(80000)
-@pytest.mark.parametrize(
-    "model_name", ["llama-3.1-model/Meta-Llama-3.1-8B", "gpt_oss/gpt-oss-20b"])
+@pytest.mark.parametrize("model_name", [
+    "llama-3.1-model/Meta-Llama-3.1-8B",
+    pytest.param("gpt_oss/gpt-oss-20b", marks=skip_pre_hopper)
+])
 def test_trtllm_benchmark_serving(llm_venv, model_name):
     test_root = unittest_path() / "llmapi" / "apps"
     llm_venv.run_cmd([

tests/integration/test_lists/waives.txt

@@ -349,3 +349,7 @@ accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_ctx_pp_gen
 examples/test_llama.py::test_llm_llama_1gpu_fp8_kv_cache[llama-v2-7b-hf-bfloat16] SKIP (https://nvbugs/5527940)
 accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16_4gpus[tp4-mtp_nextn=2-attention_dp=True-cuda_graph=True-overlap_scheduler=True-torch_compile=True] SKIP (https://nvbugs/5528070)
 accuracy/test_llm_api_autodeploy.py::TestLlama3_1_8B::test_auto_dtype SKIP (https://nvbugs/5527956)
+test_e2e.py::test_ptp_quickstart_multimodal[mistral-small-3.1-24b-instruct-Mistral-Small-3.1-24B-Instruct-2503-mixture_text_image-True] SKIP (https://nvbugs/5509024)
+accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_fp8_blockscale[throughput] SKIP (https://nvbugs/5481198)
+accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_fp8_blockscale_chunked_prefill[latency] SKIP (https://nvbugs/5481198)
+accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_fp8_blockscale_chunked_prefill[throughput] SKIP (https://nvbugs/5481198)

tests/unittest/llmapi/apps/_test_trtllm_serve_benchmark.py

@@ -3,7 +3,7 @@
 import sys
 
 import pytest
-from utils.util import skip_gpu_memory_less_than_80gb
+from utils.util import skip_gpu_memory_less_than_80gb, skip_pre_hopper
 
 from .openai_server import RemoteOpenAIServer
 
@@ -45,9 +45,11 @@ def dataset_path(dataset_name: str):
 
 
 @skip_gpu_memory_less_than_80gb
-@pytest.mark.parametrize(
-    "model_name", ["llama-3.1-model/Meta-Llama-3.1-8B", "gpt_oss/gpt-oss-20b"],
-    indirect=True)
+@pytest.mark.parametrize("model_name", [
+    "llama-3.1-model/Meta-Llama-3.1-8B",
+    pytest.param("gpt_oss/gpt-oss-20b", marks=skip_pre_hopper)
+],
+                         indirect=True)
 def test_trtllm_serve_benchmark(server: RemoteOpenAIServer, benchmark_root: str,
                                 model_path: str):
     model_name = model_path.split("/")[-1]