kanshan/TensorRT-LLMs
1
0
Fork 0
mirror of https://github.com/NVIDIA/TensorRT-LLM.git synced 2026-01-14 06:27:45 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity

tests: add ds r1 tp4 test (#5197)

Browse Source 
Signed-off-by: xinhe-nv <200704525+xinhe-nv@users.noreply.github.com>
This commit is contained in:
xinhe-nv 2025-06-19 12:48:33 +08:00 committed by GitHub
parent dedce8ab0e
commit e5400eeae0
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 73 additions and 16 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

80
tests/integration/defs/accuracy/test_llm_api_pytorch.py
View File

@ -987,25 +987,80 @@ class TestDeepSeekV3Lite(LlmapiAccuracyTestHarness):
task.evaluate(llm)
@pytest.mark.timeout(7200)
@pytest.mark.skip_less_device_memory(80000)
class TestDeepSeekR1(LlmapiAccuracyTestHarness):
MODEL_NAME = "deepseek-ai/DeepSeek-R1"
MODEL_PATH = f"{llm_models_root()}/DeepSeek-R1/DeepSeek-R1"
@pytest.mark.skip_less_mpi_world_size(8)
@skip_pre_blackwell
@pytest.mark.parametrize(
"tp_size,pp_size,ep_size,mtp_nextn,fp8kv,attention_dp,cuda_graph,overlap_scheduler,max_batch_size,moe_backend",
[
# Use a larger batch_size to speed up the tests
(8, 1, 4, 3, False, False, True, True, 32, "CUTLASS"),
(8, 1, 4, 3, False, False, True, True, 32, "TRTLLM"),
(8, 1, 8, 0, True, True, True, True, 32, "CUTLASS"),
(8, 1, 1, 0, True, True, True, True, 32, "CUTLASS"),
pytest.param(8,
1,
4,
3,
False,
False,
True,
True,
32,
"CUTLASS",
marks=pytest.mark.skip_less_device(8)),
pytest.param(8,
1,
4,
3,
False,
False,
True,
True,
32,
"TRTLLM",
marks=pytest.mark.skip_less_device(8)),
pytest.param(8,
1,
8,
0,
True,
True,
True,
True,
32,
"CUTLASS",
marks=pytest.mark.skip_less_device(8)),
pytest.param(8,
1,
1,
0,
True,
True,
True,
True,
32,
"CUTLASS",
marks=pytest.mark.skip_less_device(8)),
pytest.param(4,
1,
1,
0,
True,
True,
True,
True,
16,
"CUTLASS",
marks=pytest.mark.skip_less_device(4)),
],
ids=["latency", "latency_trtllmgen", "throughput", "throughput_tp8"])
def test_nvfp4_8gpus(self, tp_size, pp_size, ep_size, mtp_nextn, fp8kv,
attention_dp, cuda_graph, overlap_scheduler,
max_batch_size, moe_backend):
ids=[
"latency", "latency_trtllmgen", "throughput", "throughput_tp8",
"throughput_tp4"
])
def test_nvfp4_multi_gpus(self, tp_size, pp_size, ep_size, mtp_nextn, fp8kv,
attention_dp, cuda_graph, overlap_scheduler,
max_batch_size, moe_backend):
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9)
pytorch_config = dict(disable_overlap_scheduler=not overlap_scheduler,
@ -1042,9 +1097,10 @@ class TestDeepSeekR1(LlmapiAccuracyTestHarness):
task.evaluate(llm)
task = GSM8K(self.MODEL_NAME)
task.evaluate(llm)
task = GPQADiamond(self.MODEL_NAME)
task.evaluate(llm,
extra_evaluator_kwargs=dict(apply_chat_template=True))
# Commented out because GPQA takes too long to run
# task = GPQADiamond(self.MODEL_NAME)
# task.evaluate(llm,
# extra_evaluator_kwargs=dict(apply_chat_template=True))
@pytest.mark.skip_less_mpi_world_size(8)
@skip_pre_hopper

9
tests/integration/test_lists/qa/examples_test_list.txt
View File

@ -468,10 +468,11 @@ accuracy/test_llm_api_pytorch.py::TestLlama3_1NemotronNano8Bv1::test_fp8_prequan
accuracy/test_llm_api_pytorch.py::TestNemotronUltra::test_fp8_prequantized[tp8ep4-cuda_graph=True]
accuracy/test_llm_api_pytorch.py::TestNemotronH::test_reasoning_fp8_prequantized
accuracy/test_llm_api_pytorch.py::TestQwen2_7BInstruct::test_auto_dtype
accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_nvfp4_8gpus[latency]
accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_nvfp4_8gpus[latency_trtllmgen]
accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_nvfp4_8gpus[throughput]
accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_nvfp4_8gpus[throughput_tp8]
accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_nvfp4_multi_gpus[latency]
accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_nvfp4_multi_gpus[latency_trtllmgen]
accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_nvfp4_multi_gpus[throughput]
accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_nvfp4_multi_gpus[throughput_tp8]
accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_nvfp4_multi_gpus[throughput_tp4]
accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_fp8_block_scales[latency]
accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_fp8_block_scales[latency]
accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[latency_moe_cutlass]