Skip to content

Commit 3fb19de

Browse files
farazkh80farazkh80
committed
l0 tests for gb10
Signed-off-by: list <[email protected]>
1 parent 7187afe commit 3fb19de

File tree

4 files changed

+94
-11
lines changed

4 files changed

+94
-11
lines changed

‎tests/integration/defs/accuracy/test_llm_api_pytorch.py‎

Lines changed: 28 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -4096,6 +4096,34 @@ def test_w4_1gpu(self, kv_cache_dtype, moe_backend, cuda_graph,
40964096
task.evaluate(llm,
40974097
extra_evaluator_kwargs=self.extra_evaluator_kwargs)
40984098

4099+
# on spark 120b accuracy takes 2.2 hours, so we do 20b for now
4100+
def test_w4_1gpu_20b_spark(self, mocker):
4101+
mocker.patch.object(GSM8K, "MAX_OUTPUT_LEN", 8192)
4102+
mocker.patch.dict(GSM8K.EVALUATE_KWARGS,
4103+
{"scores_filter": "exact_match,flexible-extract"})
4104+
4105+
pytorch_config = dict(
4106+
disable_overlap_scheduler=False,
4107+
cuda_graph_config=CudaGraphConfig())
4108+
4109+
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.5,
4110+
dtype="auto")
4111+
4112+
model_path = f"{llm_models_root()}/gpt_oss/gpt-oss-20b"
4113+
llm = LLM(model_path,
4114+
tensor_parallel_size=1,
4115+
pipeline_parallel_size=1,
4116+
moe_expert_parallel_size=1,
4117+
kv_cache_config=kv_cache_config,
4118+
**pytorch_config,
4119+
moe_config=MoeConfig(backend="CUTLASS"))
4120+
4121+
with llm:
4122+
model_name = "GPT-OSS/20B-MXFP4"
4123+
task = GSM8K(model_name)
4124+
task.evaluate(llm,
4125+
extra_evaluator_kwargs=self.extra_evaluator_kwargs)
4126+
40994127
def test_dummy_load_format(self):
41004128
llm = LLM(
41014129
self.MODEL_PATH,

‎tests/integration/defs/test_e2e.py‎

Lines changed: 33 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1902,11 +1902,43 @@ def test_ptp_quickstart(llm_root, llm_venv):
19021902
marks=skip_pre_blackwell),
19031903
pytest.param(
19041904
'GPT-OSS-120B', 'gpt_oss/gpt-oss-120b', marks=skip_pre_blackwell),
1905+
pytest.param(
1906+
'Qwen3-8b-fp8',
1907+
'Qwen3/nvidia-Qwen3-8B-FP8',
1908+
marks=skip_pre_blackwell),
1909+
pytest.param(
1910+
'Qwen3-8b-nvfp4',
1911+
'Qwen3/nvidia-Qwen3-8B-NVFP4',
1912+
marks=skip_pre_blackwell),
1913+
pytest.param(
1914+
'Qwen3-14b-fp8',
1915+
'Qwen3/nvidia-Qwen3-14B-FP8',
1916+
marks=skip_pre_blackwell),
1917+
pytest.param(
1918+
'Qwen3-14b-nvfp4',
1919+
'Qwen3/nvidia-Qwen3-14B-NVFP4',
1920+
marks=skip_pre_blackwell),
1921+
pytest.param(
1922+
'Qwen3-32b-nvfp4',
1923+
'Qwen3/nvidia-Qwen3-32B-NVFP4',
1924+
marks=skip_pre_blackwell),
1925+
pytest.param(
1926+
'Phi4-Reasoning-Plus-fp8',
1927+
'nvidia-Phi-4-reasoning-plus-FP8',
1928+
marks=skip_pre_blackwell),
1929+
pytest.param(
1930+
'Phi4-Reasoning-Plus-nvfp4',
1931+
'nvidia-Phi-4-reasoning-plus-NVFP4',
1932+
marks=skip_pre_blackwell),
1933+
pytest.param(
1934+
'Nemotron-Nano-v2-nvfp4',
1935+
'NVIDIA-Nemotron-Nano-9B-v2-NVFP4',
1936+
marks=skip_pre_blackwell),
19051937
])
19061938
def test_ptp_quickstart_advanced(llm_root, llm_venv, model_name, model_path):
19071939
print(f"Testing {model_name}.")
19081940
example_root = Path(os.path.join(llm_root, "examples", "llm-api"))
1909-
if model_name == "Nemotron-H-8B":
1941+
if model_name in ("Nemotron-H-8B", "Nemotron-Nano-v2-nvfp4"):
19101942
llm_venv.run_cmd([
19111943
str(example_root / "quickstart_advanced.py"),
19121944
"--disable_kv_cache_reuse",

‎tests/integration/test_lists/test-db/l0_gb10.yml‎

Lines changed: 28 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -16,10 +16,13 @@ l0_gb10:
1616
backend: pytorch
1717
tests:
1818
# ------------- PyTorch tests ---------------
19-
- unittest/_torch/attention/test_attention_mla.py
20-
- test_e2e.py::test_ptp_quickstart_bert[VANILLA-BertForSequenceClassification-bert/bert-base-uncased-yelp-polarity]
21-
- test_e2e.py::test_ptp_quickstart_bert[TRTLLM-BertForSequenceClassification-bert/bert-base-uncased-yelp-polarity]
22-
- accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_bf16[latency]
19+
# - unittest/_torch/attention/test_attention_mla.py
20+
# - test_e2e.py::test_ptp_quickstart_bert[VANILLA-BertForSequenceClassification-bert/bert-base-uncased-yelp-polarity]
21+
# - test_e2e.py::test_ptp_quickstart_bert[TRTLLM-BertForSequenceClassification-bert/bert-base-uncased-yelp-polarity]
22+
# - accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_bf16[latency]
23+
- accuracy/test_llm_api_pytorch.py::TestLlama3_1_8B::test_auto_dtype
24+
- accuracy/test_llm_api_pytorch.py::TestLlama3_1_8B::test_nvfp4
25+
- accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_1gpu_20b_spark
2326
- condition:
2427
ranges:
2528
system_gpu_count:
@@ -35,8 +38,25 @@ l0_gb10:
3538
backend: pytorch
3639
tests:
3740
# ------------- PyTorch tests ---------------
38-
# Below cases which are commented out due to they failed on gb10
39-
# - unittest/_torch/modeling -k "modeling_mllama"
41+
- unittest/_torch/modeling -k "modeling_mllama"
4042
- unittest/_torch/modeling -k "modeling_out_of_tree"
41-
# - unittest/_torch/modules/test_fused_moe.py::test_fused_moe_nvfp4[CUTLASS-dtype0]
42-
# - unittest/_torch/modules/test_fused_moe.py::test_fused_moe_nvfp4[CUTLASS-dtype1]
43+
- test_e2e.py::test_ptp_quickstart_advanced[Llama3.1-8B-BF16-llama-3.1-model/Meta-Llama-3.1-8B]
44+
- test_e2e.py::test_ptp_quickstart_advanced[Llama3.1-8B-FP8-llama-3.1-model/Llama-3.1-8B-Instruct-FP8]
45+
- test_e2e.py::test_ptp_quickstart_advanced[Llama3.1-8B-NVFP4-nvfp4-quantized/Meta-Llama-3.1-8B]
46+
- test_e2e.py::test_ptp_quickstart_advanced[GPT-OSS-20B-gpt_oss/gpt-oss-20b]
47+
- test_e2e.py::test_ptp_quickstart_advanced[GPT-OSS-120B-gpt_oss/gpt-oss-120b]
48+
- test_e2e.py::test_ptp_quickstart_advanced[Qwen3-8b-fp8-Qwen3/nvidia-Qwen3-8B-FP8]
49+
- test_e2e.py::test_ptp_quickstart_advanced[Qwen3-8b-nvfp4-Qwen3/nvidia-Qwen3-8B-NVFP4]
50+
- test_e2e.py::test_ptp_quickstart_advanced[Qwen3-14b-fp8-Qwen3/nvidia-Qwen3-14B-FP8]
51+
- test_e2e.py::test_ptp_quickstart_advanced[Qwen3-14b-nvfp4-Qwen3/nvidia-Qwen3-14B-NVFP4]
52+
- test_e2e.py::test_ptp_quickstart_multimodal_phi4mm[phi4-multimodal-instruct-fp8-multimodals/Phi-4-multimodal-instruct-FP8-image_audio]
53+
- test_e2e.py::test_ptp_quickstart_multimodal_phi4mm[phi4-multimodal-instruct-fp4-multimodals/Phi-4-multimodal-instruct-FP4-image_audio]
54+
- test_e2e.py::test_ptp_quickstart_advanced[Qwen3-32b-nvfp4-Qwen3/nvidia-Qwen3-32B-NVFP4]
55+
- test_e2e.py::test_ptp_quickstart_advanced[Phi4-Reasoning-Plus-fp8-nvidia-Phi-4-reasoning-plus-FP8]
56+
- test_e2e.py::test_ptp_quickstart_advanced[Phi4-Reasoning-Plus-nvfp4-nvidia-Phi-4-reasoning-plus-NVFP4]
57+
- test_e2e.py::test_ptp_quickstart_advanced[Nemotron-Nano-v2-nvfp4-NVIDIA-Nemotron-Nano-9B-v2-NVFP4]
58+
- test_e2e.py::test_ptp_quickstart_advanced[Nemotron-Super-49B-v1-FP8-nemotron-nas/Llama-3_3-Nemotron-Super-49B-v1-FP8]
59+
- test_e2e.py::test_ptp_quickstart_advanced[Qwen3-30B-A3B_fp8_hf-Qwen3/saved_models_Qwen3-30B-A3B_fp8_hf]
60+
- test_e2e.py::test_ptp_quickstart_advanced[Qwen3-30B-A3B_nvfp4_hf-Qwen3/saved_models_Qwen3-30B-A3B_nvfp4_hf]
61+
- test_e2e.py::test_ptp_quickstart_advanced[Llama3.1-70B-NVFP4-nvfp4-quantized/Meta-Llama-3.1-70B]
62+
- test_e2e.py::test_ptp_quickstart_advanced[Llama3.1-70B-FP8-llama-3.1-model/Llama-3.1-70B-Instruct-FP8]

‎tests/unittest/_torch/modeling/test_modeling_mllama.py‎

Lines changed: 5 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -10,6 +10,7 @@
1010
from transformers import MllamaConfig
1111
from transformers import \
1212
MllamaForConditionalGeneration as HFMllamaForConditionalGeneration
13+
from utils.util import getSMVersion
1314

1415
import tensorrt_llm
1516
from tensorrt_llm._torch.attention_backend.utils import get_attention_backend
@@ -392,9 +393,10 @@ def test_mllama_allclose_to_hf_text_only(self, scenario: Scenario) -> None:
392393
position_ids=position_ids,
393394
use_cache=True)
394395

396+
atol = 0.35 if getSMVersion() >= 121 else 0.3
395397
torch.testing.assert_close(logits,
396398
ref.logits[:, -1].float(),
397-
atol=0.3,
399+
atol=atol,
398400
rtol=0.3)
399401

400402
# gen
@@ -458,9 +460,10 @@ def run_forward(input_ids, position_ids, attn_metadata):
458460
past_key_values=ref.past_key_values,
459461
use_cache=True)
460462

463+
atol = 0.35 if getSMVersion() >= 121 else 0.3
461464
torch.testing.assert_close(logits,
462465
ref.logits[:, -1].float(),
463-
atol=0.3,
466+
atol=atol,
464467
rtol=0.3)
465468
if graph_runner is not None:
466469
graph_runner.clear()

0 commit comments

Comments
 (0)