[TRTLLM-8113][test] Add pytorch workflow e2e tests with pp enabled (NVIDIA#8357)

Signed-off-by: Stanley Sun <[email protected]>
Signed-off-by: Mike Iovine <[email protected]>
Signed-off-by: FredricZ-2007 <[email protected]>

Author: StanleySun639

tests/integration/defs/test_e2e.py

@@ -2332,6 +2332,40 @@ def test_ptp_quickstart_advanced_multi_gpus(llm_root, llm_venv, model_name,
                              gpu_count)
 
 
+@pytest.mark.skip_less_device_memory(80000)
+@pytest.mark.parametrize("cuda_graph", [False, True])
+@pytest.mark.parametrize("tp_size, pp_size", [
+    pytest.param(2, 2, marks=pytest.mark.skip_less_device(4)),
+    pytest.param(2, 4, marks=pytest.mark.skip_less_device(8)),
+])
+@pytest.mark.parametrize("model_name,model_path", [
+    pytest.param('Llama3.3-70B-FP8',
+                 'llama-3.3-models/Llama-3.3-70B-Instruct-FP8',
+                 marks=skip_pre_hopper),
+])
+def test_ptp_quickstart_advanced_pp_enabled(llm_root, llm_venv, model_name,
+                                            model_path, cuda_graph, tp_size,
+                                            pp_size):
+    print(f"Testing {model_name} on 8 GPUs.")
+    example_root = Path(os.path.join(llm_root, "examples", "llm-api"))
+    cmd = [
+        str(example_root / "quickstart_advanced.py"),
+        "--enable_chunked_prefill",
+        "--model_dir",
+        f"{llm_models_root()}/{model_path}",
+        f"--tp_size={tp_size}",
+        f"--pp_size={pp_size}",
+        "--moe_ep_size=1",
+        "--kv_cache_fraction=0.5",
+    ]
+    if cuda_graph:
+        cmd.extend([
+            "--use_cuda_graph",
+            "--cuda_graph_padding_enabled",
+        ])
+    llm_venv.run_cmd(cmd)
+
+
 @skip_pre_hopper
 @pytest.mark.skip_less_device(8)
 @pytest.mark.parametrize("cuda_graph", [False, True])

tests/integration/test_lists/qa/llm_function_core.txt

@@ -417,6 +417,7 @@ accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_fp8_beam_search[
 accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_fp8_beam_search[enable_cuda_graph=True-enable_padding=True-disable_overlap_scheduler=False]
 accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_fp8_beam_search[enable_cuda_graph=True-enable_padding=True-disable_overlap_scheduler=True]
 accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_fp8_tp4
+accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_fp8_tp2pp2
 accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_nvfp4_tp4
 accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_fp8_eagle3_tp8[eagle3_one_model=True]
 accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_fp8_eagle3_tp8[eagle3_one_model=False]
@@ -692,6 +693,10 @@ test_e2e.py::test_ptp_quickstart_multimodal_multiturn[phi4-multimodal-instruct-f
 test_e2e.py::test_ptp_quickstart_multimodal_multiturn[phi4-multimodal-instruct-fp4-multimodals/Phi-4-multimodal-instruct-FP4]
 test_e2e.py::test_ptp_quickstart_bert[VANILLA-BertForSequenceClassification-bert/bert-base-uncased-yelp-polarity]
 test_e2e.py::test_ptp_quickstart_bert[TRTLLM-BertForSequenceClassification-bert/bert-base-uncased-yelp-polarity]
+test_e2e.py::test_ptp_quickstart_advanced_pp_enabled[Llama3.3-70B-FP8-llama-3.3-models/Llama-3.3-70B-Instruct-FP8-2-2-False]
+test_e2e.py::test_ptp_quickstart_advanced_pp_enabled[Llama3.3-70B-FP8-llama-3.3-models/Llama-3.3-70B-Instruct-FP8-2-2-True]
+test_e2e.py::test_ptp_quickstart_advanced_pp_enabled[Llama3.3-70B-FP8-llama-3.3-models/Llama-3.3-70B-Instruct-FP8-2-4-False]
+test_e2e.py::test_ptp_quickstart_advanced_pp_enabled[Llama3.3-70B-FP8-llama-3.3-models/Llama-3.3-70B-Instruct-FP8-2-4-True]
 test_e2e.py::test_ptp_star_attention_example[Llama3.1-8B-BF16-llama-3.1-model/Meta-Llama-3.1-8B]
 test_e2e.py::test_trtllm_bench_pytorch_backend_sanity[meta-llama/Llama-3.1-8B-llama-3.1-8b-hf-nvfp4-False-False]
 test_e2e.py::test_ptp_scaffolding[DeepSeek-R1-Distill-Qwen-7B-DeepSeek-R1/DeepSeek-R1-Distill-Qwen-7B]

tests/integration/test_lists/qa/llm_function_core_sanity.txt

@@ -129,6 +129,7 @@ accuracy/test_llm_api_pytorch.py::TestLlama3_2_1B::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestLlama3_2_1B::test_fp8_prequantized
 accuracy/test_llm_api_pytorch.py::TestLlama3_2_3B::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestLlama3_2_3B::test_fp8_prequantized
+accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_fp8_tp2pp2
 accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_fp8_tp4
 accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_nvfp4_tp4
 accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_fp8_eagle3_tp8[eagle3_one_model=True]