diff --git a/tests/integration/test_lists/test-db/l0_dgx_h200.yml b/tests/integration/test_lists/test-db/l0_dgx_h200.yml index c31066f3d4ca..ef6a7a057aed 100644 --- a/tests/integration/test_lists/test-db/l0_dgx_h200.yml +++ b/tests/integration/test_lists/test-db/l0_dgx_h200.yml @@ -31,8 +31,6 @@ l0_dgx_h200: - accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_ctx_pp_gen_tp_asymmetric[MMLU-gen_tp=2-ctx_pp=4] - disaggregated/test_disaggregated.py::test_disaggregated_ctxtp2pp2_gentp2pp2[TinyLlama-1.1B-Chat-v1.0] - disaggregated/test_disaggregated.py::test_disaggregated_ctxpp4_genpp4[TinyLlama-1.1B-Chat-v1.0] - - unittest/_torch/multi_gpu_modeling/test_llama4.py::test_llama4[pp1-ep1-disable_adp-enable_graph-tp8-trtllm-scout] - - unittest/_torch/multi_gpu_modeling/test_llama4.py::test_llama4[pp1-ep4-enable_adp-enable_graph-tp8-trtllm-scout] - unittest/llmapi/test_llm_pytorch.py::test_nemotron_nas_lora - condition: ranges: diff --git a/tests/integration/test_lists/waives.txt b/tests/integration/test_lists/waives.txt index 41d01293dc64..10d4513930ce 100644 --- a/tests/integration/test_lists/waives.txt +++ b/tests/integration/test_lists/waives.txt @@ -285,7 +285,6 @@ triton_server/test_triton.py::test_python_bls_unit_tests[python-bls-unit-tests] triton_server/test_triton.py::test_mistral_ib[mistral-ib] SKIP (https://nvbugs/5477399) triton_server/test_triton.py::test_eagle[eagle] SKIP (https://nvbugs/5477378) examples/test_mixtral.py::test_llm_mixtral_moe_plugin_lora_4gpus[Mixtral-8x7B-v0.1-chinese-mixtral-lora] SKIP (https://nvbugs/5477421) -unittest/_torch/multi_gpu_modeling/test_llama4.py::test_llama4[pp1-ep4-enable_adp-enable_graph-tp8-trtllm-scout] SKIP (https://nvbugs/5477730) test_e2e.py::test_openai_chat_example[trt] SKIP (https://nvbugs/5477444) examples/test_multimodal.py::test_llm_multimodal_general[fuyu-8b-pp:1-tp:1-float16-bs:8-cpp_e2e:False-nb:1] SKIP (https://nvbugs/5448462) examples/test_multimodal.py::test_llm_multimodal_general[fuyu-8b-pp:1-tp:1-float16-bs:8-cpp_e2e:True-nb:1] SKIP (https://nvbugs/5448462) @@ -339,7 +338,6 @@ accuracy/test_cli_flow.py::TestMixtral8x7B::test_fp8_tp2pp2_manage_weights SKIP accuracy/test_cli_flow.py::TestLlama3_1_8B::test_tp4[enable_gemm_allreduce_plugin] SKIP (https://nvbugs/5532023) accuracy/test_llm_api.py::TestLlama3_1_8BInstruct::test_tp2cp2 SKIP (https://nvbugs/5532023) accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4_4gpus[moe_backend=CUTLASS-mtp_nextn=2-tp2pp2-fp8kv=True-attention_dp=True-cuda_graph=True-overlap_scheduler=True-torch_compile=False] SKIP (https://nvbugs/5503479) -unittest/_torch/multi_gpu_modeling/test_llama4.py::test_llama4[pp1-ep1-disable_adp-enable_graph-tp8-trtllm-scout] SKIP (https://nvbugs/5541545) full:L20/accuracy/test_llm_api_pytorch.py::TestLlama3_2_1B::test_fp8_prequantized SKIP (https://nvbugs/5542862) full:L20/accuracy/test_llm_api_pytorch.py::TestMinistral8BInstruct::test_fp8 SKIP (https://nvbugs/5542862) full:L40S/accuracy/test_llm_api_pytorch.py::TestLlama3_2_1B::test_fp8_prequantized SKIP (https://nvbugs/5542862) diff --git a/tests/unittest/_torch/multi_gpu_modeling/test_llama4.py b/tests/unittest/_torch/multi_gpu_modeling/test_llama4.py deleted file mode 100644 index 6149201d5825..000000000000 --- a/tests/unittest/_torch/multi_gpu_modeling/test_llama4.py +++ /dev/null @@ -1,91 +0,0 @@ -from difflib import SequenceMatcher - -import pytest -import torch -from utils.llm_data import llm_models_root - -from tensorrt_llm import LLM, SamplingParams -from tensorrt_llm.llmapi import CudaGraphConfig, KvCacheConfig - - -@pytest.mark.parametrize( - "model_name", - ["Llama-4-Maverick-17B-128E-Instruct", "Llama-4-Scout-17B-16E-Instruct"], - ids=['maverick', 'scout']) -@pytest.mark.parametrize("backend", ["TRTLLM", "FLASHINFER"], - ids=["trtllm", "flashinfer"]) -@pytest.mark.parametrize("tp_size", [1, 8], ids=["tp1", "tp8"]) -@pytest.mark.parametrize("use_cuda_graph", [True, False], - ids=["enable_graph", "disable_graph"]) -@pytest.mark.parametrize("enable_attention_dp", [True, False], - ids=["enable_adp", "disable_adp"]) -@pytest.mark.parametrize("ep_size", [4, 1], ids=["ep4", "ep1"]) -@pytest.mark.parametrize("pp_size", [1, 8], ids=["pp1", "pp8"]) -def test_llama4(model_name, backend, tp_size, use_cuda_graph, - enable_attention_dp, ep_size, pp_size): - if pp_size > 1 and (ep_size > 1 or tp_size > 1): - return - - if pp_size == 1 and tp_size == 1: - return - - if enable_attention_dp and not (tp_size == 8 and ep_size == 4 - and pp_size == 1): - pytest.skip("Skip this attention DP test case to avoid too many tests") - - prompts = [ - { - "prompt": "The president of the United States is" - }, - { - # NOTE: Long context accuracy testing (RULER) is not available in CI yet. - # This test cannot be removed until long context is covered. - "prompt": - "This is a very long prompt to exercise long context. Count up to 10000 from 1, 2, 3," - + ", ".join(str(i) for i in range(4, 9000)) - }, - { - "prompt": "<|image|>This image is of color", - "multi_modal_data": { - "image": [torch.ones(3, 1024, 1024)] - } - }, - ] - - expected_outputs = [ - " the head of state and head of government of the", ", 9000, 9001, ", - " white. What is the color of the background of" - ] - - pytorch_config = dict(attn_backend=backend) - model_dir = str(llm_models_root() / "llama4-models" / model_name) - - kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.25, ) - llm = LLM( - model=model_dir, - tensor_parallel_size=tp_size, - moe_expert_parallel_size=ep_size, - moe_tensor_parallel_size=tp_size // ep_size, - cuda_graph_config=CudaGraphConfig() if use_cuda_graph else None, - **pytorch_config, - pipeline_parallel_size=pp_size, - enable_attention_dp=enable_attention_dp, - kv_cache_config=kv_cache_config, - enable_chunked_prefill=True, - ) - with llm: - outputs = llm.generate( - prompts, - sampling_params=SamplingParams(max_tokens=10), - ) - - assert len(outputs) == len(expected_outputs), "Output length mismatch" - - def similar(a, b, threshold=0.9): - return SequenceMatcher(None, a, b).ratio() >= threshold - - for output, expected in zip(outputs, expected_outputs): - output_text = output.outputs[0].text - assert similar( - output_text, - expected), f"Expected '{expected}' but get '{output_text}'"