From 939f9dd182d88b0fd35cedef8b620c2f6d5f02f7 Mon Sep 17 00:00:00 2001 From: nv-guomingz <137257613+nv-guomingz@users.noreply.github.com> Date: Wed, 24 Jun 2026 10:59:24 +0000 Subject: [PATCH] [TRTLLM-13600][test] Clean up Qwen3 test cases Signed-off-by: nv-guomingz <137257613+nv-guomingz@users.noreply.github.com> --- .../defs/accuracy/test_llm_api_pytorch.py | 218 +++++++----------- .../test_lists/test-db/l0_b200.yml | 16 +- .../test_lists/test-db/l0_dgx_b200.yml | 6 +- .../test_lists/test-db/l0_dgx_b300.yml | 2 +- .../test_lists/test-db/l0_h100.yml | 8 +- 5 files changed, 101 insertions(+), 149 deletions(-) diff --git a/tests/integration/defs/accuracy/test_llm_api_pytorch.py b/tests/integration/defs/accuracy/test_llm_api_pytorch.py index 9666884d93e5..9316d0e0698f 100644 --- a/tests/integration/defs/accuracy/test_llm_api_pytorch.py +++ b/tests/integration/defs/accuracy/test_llm_api_pytorch.py @@ -4202,15 +4202,12 @@ class TestQwen3_8B(LlmapiAccuracyTestHarness): MODEL_NAME = "Qwen3/Qwen3-8B" @skip_pre_hopper - @pytest.mark.parametrize( - "tp_size,pp_size,ep_size,attention_dp,cuda_graph,overlap_scheduler", - [(1, 1, 1, False, True, True)], - ids=["latency"]) - def test_fp8_block_scales(self, tp_size, pp_size, ep_size, attention_dp, - cuda_graph, overlap_scheduler): - pytorch_config = dict( - disable_overlap_scheduler=not overlap_scheduler, - cuda_graph_config=CudaGraphConfig() if cuda_graph else None) + @pytest.mark.parametrize("tp_size,pp_size,ep_size,attention_dp", + [(1, 1, 1, False)], + ids=["latency"]) + def test_fp8_block_scales(self, tp_size, pp_size, ep_size, attention_dp): + pytorch_config = dict(cuda_graph_config=CudaGraphConfig( + max_batch_size=32, enable_padding=True)) with LLM(f"{llm_models_root()}/Qwen3/Qwen3-8B-FP8", tensor_parallel_size=tp_size, @@ -4218,7 +4215,7 @@ def test_fp8_block_scales(self, tp_size, pp_size, ep_size, attention_dp, moe_expert_parallel_size=ep_size, **pytorch_config, enable_attention_dp=attention_dp, - max_batch_size=64) as llm: + max_batch_size=32) as llm: task = CnnDailymail(self.MODEL_NAME) task.evaluate(llm) task = MMLU(self.MODEL_NAME) @@ -4230,7 +4227,8 @@ def test_fp8_block_scales_early_first_token_response(self): pytorch_config = dict( disable_overlap_scheduler=False, enable_early_first_token_response=True, - cuda_graph_config=CudaGraphConfig(), + cuda_graph_config=CudaGraphConfig(enable_padding=True, + max_batch_size=32), ) kv_cache_config = KvCacheConfig(enable_block_reuse=False) @@ -4238,7 +4236,7 @@ def test_fp8_block_scales_early_first_token_response(self): **pytorch_config, kv_cache_config=kv_cache_config, enable_chunked_prefill=False, - max_batch_size=64) as llm: + max_batch_size=32) as llm: task = CnnDailymail(self.MODEL_NAME) task.evaluate(llm) task = MMLU(self.MODEL_NAME) @@ -4271,13 +4269,15 @@ def test_bf16(self, tp_size, pp_size, ep_size, attention_dp, cuda_graph, overlap_scheduler, is_cached): pytorch_config = dict( disable_overlap_scheduler=not overlap_scheduler, - cuda_graph_config=CudaGraphConfig() if cuda_graph else None) + cuda_graph_config=CudaGraphConfig( + enable_padding=True, max_batch_size=32) if cuda_graph else None) with LLM(f"{llm_models_root()}/Qwen3/Qwen3-8B" if is_cached else "Qwen/Qwen3-8B", tensor_parallel_size=tp_size, pipeline_parallel_size=pp_size, moe_expert_parallel_size=ep_size, + max_batch_size=32, **pytorch_config, enable_attention_dp=attention_dp) as llm: task = CnnDailymail(self.MODEL_NAME) @@ -4306,7 +4306,7 @@ def test_eagle3(self, eagle3_one_model, enable_chunked_prefill, 350: 2 } if enable_draft_len_schedule else None cuda_graph_config = (CudaGraphConfig( - max_batch_size=500) if enable_draft_len_schedule + enable_padding=True, max_batch_size=32) if enable_draft_len_schedule or enable_max_concurrency else None) max_draft_len = 4 @@ -4334,6 +4334,7 @@ def test_eagle3(self, eagle3_one_model, enable_chunked_prefill, kv_cache_config=kv_cache_config, enable_chunked_prefill=enable_chunked_prefill, max_num_tokens=256 if enable_chunked_prefill else 8192, + max_batch_size=32, speculative_config=spec_config) with llm: @@ -4365,22 +4366,21 @@ def test_dflash(self): task.evaluate(llm) @skip_pre_blackwell - @pytest.mark.parametrize( - "tp_size,pp_size,ep_size,attention_dp,cuda_graph,overlap_scheduler", - [(1, 1, 1, False, True, True)], - ids=["latency"]) + @pytest.mark.parametrize("tp_size,pp_size,ep_size,attention_dp", + [(1, 1, 1, False)], + ids=["latency"]) @pytest.mark.parametrize("activation_dtype", ["fp8", "mxfp8"]) def test_w4a8_mxfp4(self, tp_size, pp_size, ep_size, attention_dp, - cuda_graph, overlap_scheduler, activation_dtype): - pytorch_config = dict( - disable_overlap_scheduler=not overlap_scheduler, - cuda_graph_config=CudaGraphConfig() if cuda_graph else None) + activation_dtype): + pytorch_config = dict(cuda_graph_config=CudaGraphConfig( + enable_padding=True, max_batch_size=32)) llm = LLM( f"{llm_models_root()}/mxfp4-qwen3/saved_models_Qwen3-8B_w4a8_mxfp4_{activation_dtype}_kv_none_hf", tensor_parallel_size=tp_size, pipeline_parallel_size=pp_size, moe_expert_parallel_size=ep_size, + max_batch_size=32, **pytorch_config, enable_attention_dp=attention_dp) with llm: @@ -4394,22 +4394,19 @@ class TestQwen3_30B_A3B(LlmapiAccuracyTestHarness): @skip_pre_hopper @skip_post_blackwell @parametrize_with_ids("torch_compile", [False, True]) - @pytest.mark.parametrize( - "tp_size,pp_size,ep_size,attention_dp,cuda_graph,overlap_scheduler", - [(1, 1, 1, False, False, True)], - ids=["latency"]) - def test_fp8_block_scales(self, tp_size, pp_size, ep_size, attention_dp, - cuda_graph, overlap_scheduler, torch_compile): + @pytest.mark.parametrize("tp_size,ep_size,attention_dp,cuda_graph", + [(1, 1, False, False)], + ids=["latency"]) + def test_fp8_block_scales(self, tp_size, ep_size, attention_dp, cuda_graph, + torch_compile): torch_compile_config = _get_default_torch_compile_config(torch_compile) pytorch_config = dict( - disable_overlap_scheduler=not overlap_scheduler, cuda_graph_config=CudaGraphConfig() if cuda_graph else None, torch_compile_config=torch_compile_config) with LLM(f"{llm_models_root()}/Qwen3/Qwen3-30B-A3B-FP8", tensor_parallel_size=tp_size, - pipeline_parallel_size=pp_size, moe_expert_parallel_size=ep_size, **pytorch_config, enable_attention_dp=attention_dp) as llm: @@ -4443,41 +4440,36 @@ def test_dummy_load_format(self): @skip_pre_ada @parametrize_with_ids("torch_compile", [False, True]) - @pytest.mark.parametrize( - "tp_size,pp_size,ep_size,attention_dp,cuda_graph,overlap_scheduler", - [(1, 1, 1, True, True, True)], - ids=["latency"]) - def test_fp8(self, tp_size, pp_size, ep_size, attention_dp, cuda_graph, - overlap_scheduler, torch_compile): + @pytest.mark.parametrize("tp_size,ep_size", [(1, 1)], ids=["latency"]) + def test_fp8(self, tp_size, ep_size, torch_compile): "RCCA: https://nvbugspro.nvidia.com/bug/5284463" "Need to check Ada support" torch_compile_config = _get_default_torch_compile_config(torch_compile) - pytorch_config = dict( - disable_overlap_scheduler=not overlap_scheduler, - cuda_graph_config=CudaGraphConfig() if cuda_graph else None, - torch_compile_config=torch_compile_config) + pytorch_config = dict(cuda_graph_config=CudaGraphConfig( + max_batch_size=32, enable_padding=True), + torch_compile_config=torch_compile_config) with LLM(f"{llm_models_root()}/Qwen3/saved_models_Qwen3-30B-A3B_fp8_hf", tensor_parallel_size=tp_size, - pipeline_parallel_size=pp_size, moe_expert_parallel_size=ep_size, + max_batch_size=32, **pytorch_config, - enable_attention_dp=attention_dp) as llm: + enable_attention_dp=True) as llm: task = GSM8K(self.MODEL_NAME) task.evaluate(llm) @skip_pre_blackwell @parametrize_with_ids("torch_compile", [False, True]) @pytest.mark.parametrize( - "tp_size,pp_size,ep_size,attention_dp,cuda_graph,overlap_scheduler,moe_backend", + "tp_size,pp_size,ep_size,attention_dp,moe_backend", [ - (1, 1, 1, False, True, True, "CUTLASS"), - (1, 1, 1, False, True, True, "TRTLLM"), - (4, 1, 4, True, True, True, "CUTLASS"), - (4, 1, 4, True, True, True, "TRTLLM"), - (4, 1, 4, False, True, True, "CUTLASS"), - (4, 1, 4, False, True, True, "TRTLLM"), + (1, 1, 1, False, "CUTLASS"), + (1, 1, 1, False, "TRTLLM"), + (4, 1, 4, True, "CUTLASS"), + (4, 1, 4, True, "TRTLLM"), + (4, 1, 4, False, "CUTLASS"), + (4, 1, 4, False, "TRTLLM"), ], ids=[ "latency_moe_cutlass", @@ -4494,8 +4486,6 @@ def test_nvfp4( pp_size, ep_size, attention_dp, - cuda_graph, - overlap_scheduler, moe_backend, torch_compile, ): @@ -4506,8 +4496,8 @@ def test_nvfp4( torch_compile_config = _get_default_torch_compile_config(torch_compile) pytorch_config = dict( - disable_overlap_scheduler=not overlap_scheduler, - cuda_graph_config=CudaGraphConfig() if cuda_graph else None, + cuda_graph_config=CudaGraphConfig(max_batch_size=32, + enable_padding=True), moe_config=MoeConfig(backend=moe_backend), torch_compile_config=torch_compile_config, ) @@ -4528,17 +4518,15 @@ def test_nvfp4( task.evaluate(llm) @pytest.mark.parametrize("moe_backend", ["CUTLASS", "TRTLLM"]) - @pytest.mark.parametrize( - "tp_size,pp_size,ep_size,attention_dp,cuda_graph,overlap_scheduler", [ - (1, 1, 1, False, True, True), - (2, 1, 2, False, True, True), - (4, 1, 4, False, True, True), - ], - ids=["latency", "ep2", "ep4"]) + @pytest.mark.parametrize("tp_size,pp_size,ep_size", [ + (1, 1, 1), + (2, 1, 2), + (4, 1, 4), + ], + ids=["latency", "ep2", "ep4"]) @pytest.mark.parametrize("activation_dtype", ["static_fp8", "mxfp8"], ids=["fp8", "mxfp8"]) def test_w4a8_mxfp4(self, moe_backend, tp_size, pp_size, ep_size, - attention_dp, cuda_graph, overlap_scheduler, activation_dtype): if moe_backend in ["CUTLASS", "TRTLLM"] and get_sm_version() < 100: pytest.skip( @@ -4549,17 +4537,17 @@ def test_w4a8_mxfp4(self, moe_backend, tp_size, pp_size, ep_size, pytest.skip( "Mxfp8 is only supported for TRTLLM or CUTLASS moe backend.") - pytorch_config = dict( - disable_overlap_scheduler=not overlap_scheduler, - cuda_graph_config=CudaGraphConfig() if cuda_graph else None) + pytorch_config = dict(cuda_graph_config=CudaGraphConfig( + enable_padding=True, max_batch_size=32)) llm = LLM( f"{llm_models_root()}/mxfp4-qwen3/saved_models_Qwen3-30B-A3B_w4a8_mxfp4_{activation_dtype}_kv_none_hf_moeonly", tensor_parallel_size=tp_size, pipeline_parallel_size=pp_size, moe_expert_parallel_size=ep_size, + max_batch_size=32, **pytorch_config, - enable_attention_dp=attention_dp, + enable_attention_dp=False, moe_config=MoeConfig(backend=moe_backend)) with llm: task = MMLU(self.MODEL_NAME) @@ -4569,23 +4557,22 @@ def test_w4a8_mxfp4(self, moe_backend, tp_size, pp_size, ep_size, pytest.param("TRITON", marks=skip_no_hopper), pytest.param("TRTLLM", marks=skip_pre_blackwell) ]) - @pytest.mark.parametrize( - "tp_size,pp_size,ep_size,attention_dp,cuda_graph,overlap_scheduler", [ - (1, 1, 1, False, True, True), - ], - ids=["latency"]) + @pytest.mark.parametrize("tp_size,pp_size,ep_size,attention_dp", [ + (1, 1, 1, False), + ], + ids=["latency"]) def test_w4a16_mxfp4(self, tp_size, pp_size, ep_size, attention_dp, - cuda_graph, overlap_scheduler, moe_backend): - pytorch_config = dict( - disable_overlap_scheduler=not overlap_scheduler, - cuda_graph_config=CudaGraphConfig() if cuda_graph else None, - moe_config=MoeConfig(backend=moe_backend)) + moe_backend): + pytorch_config = dict(cuda_graph_config=CudaGraphConfig( + enable_padding=True, max_batch_size=32), + moe_config=MoeConfig(backend=moe_backend)) llm = LLM( f"{llm_models_root()}/mxfp4-qwen3/saved_models_Qwen3-30B-A3B_w4a16_mxfp4_kv_none_hf_moeonly", tensor_parallel_size=tp_size, pipeline_parallel_size=pp_size, moe_expert_parallel_size=ep_size, + max_batch_size=32, enable_attention_dp=attention_dp, **pytorch_config) with llm: @@ -4593,32 +4580,6 @@ def test_w4a16_mxfp4(self, tp_size, pp_size, ep_size, attention_dp, task.evaluate(llm) -class TestQwen3_32B(LlmapiAccuracyTestHarness): - MODEL_NAME = "Qwen3/Qwen3-32B" - - @skip_pre_hopper - @pytest.mark.parametrize( - "tp_size,pp_size,ep_size,attention_dp,cuda_graph,overlap_scheduler", - [(1, 1, 1, False, False, True)], - ids=["latency"]) - def test_fp8_block_scales(self, tp_size, pp_size, ep_size, attention_dp, - cuda_graph, overlap_scheduler): - pytorch_config = dict( - disable_overlap_scheduler=not overlap_scheduler, - cuda_graph_config=CudaGraphConfig() if cuda_graph else None) - - with LLM(f"{llm_models_root()}/Qwen3/Qwen3-32B-FP8", - tensor_parallel_size=tp_size, - pipeline_parallel_size=pp_size, - moe_expert_parallel_size=ep_size, - **pytorch_config, - enable_attention_dp=attention_dp) as llm: - task = CnnDailymail(self.MODEL_NAME) - task.evaluate(llm) - task = MMLU(self.MODEL_NAME) - task.evaluate(llm) - - class TestQwen3_235B_A22B(LlmapiAccuracyTestHarness): MODEL_NAME = "Qwen3/Qwen3-235B-A22B" @@ -4653,22 +4614,20 @@ def test_fp8(self, tp_size, pp_size, ep_size, attention_dp, cuda_graph, @skip_pre_hopper @pytest.mark.skip_less_device(8) - @pytest.mark.parametrize( - "tp_size,pp_size,ep_size,attention_dp,cuda_graph,overlap_scheduler,moe_backend", - [(8, 1, 8, True, True, True, "DEEPGEMM"), - (8, 1, 8, False, True, True, "DEEPGEMM"), - (8, 1, 8, True, True, True, "TRTLLM"), - (8, 1, 8, False, True, True, "TRTLLM")], - ids=[ - "latency_deepgemm", "throughput_latency_deepgemm", "latency_trtllm", - "throughput_latency_trtllm" - ]) + @pytest.mark.parametrize("tp_size,pp_size,ep_size,attention_dp,moe_backend", + [(8, 1, 8, True, "DEEPGEMM"), + (8, 1, 8, False, "DEEPGEMM"), + (8, 1, 8, True, "TRTLLM"), + (8, 1, 8, False, "TRTLLM")], + ids=[ + "latency_deepgemm", + "throughput_latency_deepgemm", + "latency_trtllm", "throughput_latency_trtllm" + ]) def test_fp8_block_scales(self, tp_size, pp_size, ep_size, attention_dp, - cuda_graph, overlap_scheduler, moe_backend): - pytorch_config = dict( - disable_overlap_scheduler=not overlap_scheduler, - cuda_graph_config=CudaGraphConfig() if cuda_graph else None, - moe_config=MoeConfig(backend=moe_backend)) + moe_backend): + pytorch_config = dict(cuda_graph_config=CudaGraphConfig(), + moe_config=MoeConfig(backend=moe_backend)) kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.6) with LLM(f"{llm_models_root()}/Qwen3/Qwen3-235B-A22B-FP8", @@ -4686,13 +4645,13 @@ def test_fp8_block_scales(self, tp_size, pp_size, ep_size, attention_dp, @skip_pre_blackwell @pytest.mark.skip_less_mpi_world_size(8) @pytest.mark.parametrize( - "tp_size,pp_size,ep_size,attention_dp,cuda_graph,overlap_scheduler,moe_backend,eagle3", + "tp_size,pp_size,ep_size,attention_dp,moe_backend,eagle3", [ - (8, 1, 8, True, True, True, "CUTLASS", False), - (8, 1, 8, False, True, True, "TRTLLM", False), - (8, 1, 8, True, True, True, "TRTLLM", False), - (8, 1, 8, False, True, True, "TRTLLM", True), - (8, 1, 8, True, True, True, "TRTLLM", True), + (8, 1, 8, True, "CUTLASS", False), + (8, 1, 8, False, "TRTLLM", False), + (8, 1, 8, True, "TRTLLM", False), + (8, 1, 8, False, "TRTLLM", True), + (8, 1, 8, True, "TRTLLM", True), ], ids=[ "latency_moe_cutlass", @@ -4702,17 +4661,15 @@ def test_fp8_block_scales(self, tp_size, pp_size, ep_size, attention_dp, "latency_moe_trtllm_eagle3_attention_dp", ], ) - def test_nvfp4(self, tp_size, pp_size, ep_size, attention_dp, cuda_graph, - overlap_scheduler, moe_backend, eagle3): + def test_nvfp4(self, tp_size, pp_size, ep_size, attention_dp, moe_backend, + eagle3): sm_version = get_sm_version() if moe_backend == "TRTLLM" and sm_version in (120, 121): pytest.skip(f"{moe_backend} backend does not support SM 120 or 121") - pytorch_config = dict( - disable_overlap_scheduler=not overlap_scheduler, - cuda_graph_config=CudaGraphConfig() if cuda_graph else None, - moe_config=MoeConfig(backend=moe_backend)) + pytorch_config = dict(cuda_graph_config=CudaGraphConfig(), + moe_config=MoeConfig(backend=moe_backend)) kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.4, enable_block_reuse=not eagle3) @@ -5698,11 +5655,6 @@ def test_w4_4gpus_online_eplb(self, kv_cache_dtype, mocker): extra_evaluator_kwargs=self.extra_evaluator_kwargs) -class TestQwQ_32B(LlmapiAccuracyTestHarness): - MODEL_NAME = "Qwen/QwQ-32B" - MODEL_PATH = f"{llm_models_root()}/QwQ-32B" - - @skip_pre_hopper @pytest.mark.skip_less_device_memory(80000) class TestQwen3NextInstruct(LlmapiAccuracyTestHarness): diff --git a/tests/integration/test_lists/test-db/l0_b200.yml b/tests/integration/test_lists/test-db/l0_b200.yml index 6c9a9331a277..6daf0cfc66f0 100644 --- a/tests/integration/test_lists/test-db/l0_b200.yml +++ b/tests/integration/test_lists/test-db/l0_b200.yml @@ -51,17 +51,9 @@ l0_b200: - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_1gpu[v2_kv_cache-True-True-trtllm-fp8] - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_dummy_load_format - accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_fp8_block_scales[latency] # Cover nvbugs 5461712 and 5505402 - - accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_w4a8_mxfp4[fp8-latency] - - accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_w4a8_mxfp4[mxfp8-latency] - - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[latency_moe_cutlass-torch_compile=False] - - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[latency_moe_cutlass-torch_compile=True] - - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[latency_moe_trtllm-torch_compile=False] - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[latency_moe_trtllm-torch_compile=True] - - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_w4a8_mxfp4[fp8-latency-CUTLASS] - - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_w4a16_mxfp4[latency-TRITON] - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_w4a8_mxfp4[fp8-latency-TRTLLM] - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_w4a8_mxfp4[mxfp8-latency-TRTLLM] - - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_w4a8_mxfp4[mxfp8-latency-CUTLASS] - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_w4a16_mxfp4[latency-TRTLLM] - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B_Instruct_2507::test_skip_softmax_attention[target_sparsity_0.9-fp8kv=True] - accuracy/test_llm_api_pytorch.py::TestQwen3_5_35B_A3B::test_fp8[enable_block_reuse=True] @@ -319,6 +311,14 @@ l0_b200: # canary in post_merge for CUDA IPC / virtual memory / profiling paths. - unittest/_torch/misc - accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_nvfp4[tp1_block_reuse-cutlass] + - accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_w4a8_mxfp4[fp8-latency] + - accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_w4a8_mxfp4[mxfp8-latency] + - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[latency_moe_cutlass-torch_compile=False] + - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[latency_moe_trtllm-torch_compile=False] + - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[latency_moe_cutlass-torch_compile=True] + - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_w4a8_mxfp4[fp8-latency-CUTLASS] + - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_w4a16_mxfp4[latency-TRITON] + - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_w4a8_mxfp4[mxfp8-latency-CUTLASS] - accuracy/test_llm_api_pytorch.py::TestQwen3_5_35B_A3B::test_fp8[enable_block_reuse=False] - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=0-fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False] - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4[moe_backend=CUTLASS-mtp_nextn=0-fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=True] diff --git a/tests/integration/test_lists/test-db/l0_dgx_b200.yml b/tests/integration/test_lists/test-db/l0_dgx_b200.yml index c1ea19000626..6506a453bde4 100644 --- a/tests/integration/test_lists/test-db/l0_dgx_b200.yml +++ b/tests/integration/test_lists/test-db/l0_dgx_b200.yml @@ -51,9 +51,6 @@ l0_dgx_b200: - accuracy/test_llm_api_pytorch.py::TestNemotronV3Ultra::test_nvfp4_4gpus_static_eplb[moe_backend=CUTLASS] - accuracy/test_llm_api_pytorch.py::TestNemotronV3Ultra::test_nvfp4_4gpus_static_eplb[moe_backend=TRTLLM] - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[tep4_latency_moe_trtllm-torch_compile=True] - - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[dep4_latency_moe_trtllm-torch_compile=False] - - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[dep4_latency_moe_cutlass-torch_compile=False] - - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[dep4_latency_moe_cutlass-torch_compile=True] - disaggregated/test_disaggregated.py::test_disaggregated_deepseek_v3_lite_fp8_ucx[DeepSeek-V3-Lite-fp8] - disaggregated/test_disaggregated.py::test_disaggregated_deepseek_v3_lite_fp8_nixl[DeepSeek-V3-Lite-fp8] - disaggregated/test_disaggregated.py::test_disaggregated_gpt_oss_120b_harmony[gpt_oss/gpt-oss-120b] @@ -279,6 +276,9 @@ l0_dgx_b200: - accuracy/test_llm_api_pytorch.py::TestNemotronV3Super::test_auto_dtype_4gpus[4-1-True-True-True] - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[tep4_latency_moe_trtllm-torch_compile=False] - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[tep4_latency_moe_cutlass-torch_compile=False] + - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[dep4_latency_moe_trtllm-torch_compile=False] + - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[dep4_latency_moe_cutlass-torch_compile=False] + - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[dep4_latency_moe_cutlass-torch_compile=True] - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[tep4_latency_moe_cutlass-torch_compile=True] - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[dep4_latency_moe_trtllm-torch_compile=True] - accuracy/test_llm_api_pytorch.py::TestQwen3_5_397B_A17B::test_nvfp4[tep4_cutedsl] diff --git a/tests/integration/test_lists/test-db/l0_dgx_b300.yml b/tests/integration/test_lists/test-db/l0_dgx_b300.yml index 48ea3217d2ff..6e3c5f49f860 100644 --- a/tests/integration/test_lists/test-db/l0_dgx_b300.yml +++ b/tests/integration/test_lists/test-db/l0_dgx_b300.yml @@ -50,6 +50,7 @@ l0_dgx_b300: - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4_4gpus[moe_backend=CUTLASS-mtp_nextn=0-tp4-fp8kv=False-attention_dp=True-cuda_graph=False-overlap_scheduler=False-low_precision_combine=True-torch_compile=False] - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4_4gpus[moe_backend=CUTLASS-mtp_nextn=0-tp4-fp8kv=True-attention_dp=True-cuda_graph=True-overlap_scheduler=True-low_precision_combine=True-torch_compile=False] - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[tep4_latency_moe_trtllm-torch_compile=True] + - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[tep4_latency_moe_trtllm-torch_compile=False] - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[tep4_latency_moe_cutlass-torch_compile=True] - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[dep4_latency_moe_trtllm-torch_compile=False] - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[dep4_latency_moe_cutlass-torch_compile=True] @@ -117,7 +118,6 @@ l0_dgx_b300: - unittest/_torch/modules/moe/test_moe_module.py::test_configurable_moe_multi_gpu_eplb -k "MEGAMOE_CUTEDSL" - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16_4gpus[ep4-mtp_nextn=2-attention_dp=True-cuda_graph=True-overlap_scheduler=True-torch_compile=False] - accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16_4gpus[tp2pp2-mtp_nextn=2-attention_dp=True-cuda_graph=True-overlap_scheduler=True-torch_compile=False] - - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[tep4_latency_moe_trtllm-torch_compile=False] - disaggregated/test_disaggregated.py::test_disaggregated_deepseek_v3_lite_fp8_nixl[DeepSeek-V3-Lite-fp8] - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_4gpus[v1_kv_cache-ep4-trtllm-auto] - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_4gpus[v2_kv_cache-ep4-trtllm-auto] diff --git a/tests/integration/test_lists/test-db/l0_h100.yml b/tests/integration/test_lists/test-db/l0_h100.yml index fc5b2283e600..307524d53302 100644 --- a/tests/integration/test_lists/test-db/l0_h100.yml +++ b/tests/integration/test_lists/test-db/l0_h100.yml @@ -140,13 +140,9 @@ l0_h100: - accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_fp8_block_scales[latency] - accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_fp8_block_scales_early_first_token_response - accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_dummy_load_format - - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_fp8[latency-torch_compile=False] - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_fp8[latency-torch_compile=True] - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_dummy_load_format - - accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_eagle3[eagle3_one_model=True-enable_chunked_prefill=True-enable_max_concurrency=False-enable_draft_len_schedule=False] - accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_eagle3[eagle3_one_model=True-enable_chunked_prefill=False-enable_max_concurrency=False-enable_draft_len_schedule=False] - - accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_eagle3[eagle3_one_model=True-enable_chunked_prefill=False-enable_max_concurrency=False-enable_draft_len_schedule=True] - - accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_eagle3[eagle3_one_model=True-enable_chunked_prefill=False-enable_max_concurrency=True-enable_draft_len_schedule=False] - accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_dflash - accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_dflash - accuracy/test_llm_api_pytorch.py::TestQwen3_5_4B::test_bf16 @@ -338,6 +334,9 @@ l0_h100: orchestrator: mpi tests: # ------------- PyTorch tests --------------- + - accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_eagle3[eagle3_one_model=True-enable_chunked_prefill=True-enable_max_concurrency=False-enable_draft_len_schedule=False] + - accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_eagle3[eagle3_one_model=True-enable_chunked_prefill=False-enable_max_concurrency=False-enable_draft_len_schedule=True] + - accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_eagle3[eagle3_one_model=True-enable_chunked_prefill=False-enable_max_concurrency=True-enable_draft_len_schedule=False] - accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_bfloat16[attn_backend=FLASHINFER-torch_compile=False] - accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_bfloat16[attn_backend=FLASHINFER-torch_compile=True] - accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_bfloat16[attn_backend=TRTLLM-torch_compile=False] @@ -397,6 +396,7 @@ l0_h100: - accuracy/test_llm_api_pytorch.py::TestGemma3_1BInstruct::test_fp8_prequantized[torch_compile=True] - accuracy/test_llm_api_pytorch.py::TestGemma3_27BInstruct::test_fp8_prequantized - accuracy/test_llm_api_pytorch.py::TestGemma3_27BInstruct::test_auto_dtype TIMEOUT (90) + - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_fp8[latency-torch_compile=False] - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_fp8_block_scales[latency-torch_compile=False] - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_fp8_block_scales[latency-torch_compile=True] - accuracy/test_llm_api_pytorch.py::TestLlama3_1_8BInstruct::test_guided_decoding[llguidance]