Skip to content

Commit b0ac7c9

Browse files
xinhe-nvxinhe-nv
authored
test: skip failed cases on B200 (#3710)
* add skip condition to tests Signed-off-by: xinhe-nv <[email protected]> * fix error Signed-off-by: xinhe-nv <[email protected]> --------- Signed-off-by: xinhe-nv <[email protected]>
1 parent f08e599 commit b0ac7c9

File tree

7 files changed

+57
-7
lines changed

7 files changed

+57
-7
lines changed

‎tests/integration/defs/accuracy/test_cli_flow.py‎

Lines changed: 39 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -19,7 +19,8 @@
1919
from tensorrt_llm.quantization import QuantAlgo
2020

2121
from ..conftest import (llm_models_root, parametrize_with_ids, skip_no_nvls,
22-
skip_pre_ada, skip_pre_blackwell, skip_pre_hopper)
22+
skip_post_blackwell, skip_pre_ada, skip_pre_blackwell,
23+
skip_pre_hopper)
2324
from .accuracy_core import (MMLU, CliFlowAccuracyTestHarness, CnnDailymail,
2425
Humaneval, PassKeyRetrieval64k,
2526
PassKeyRetrieval128k, SlimPajama6B, ZeroScrolls)
@@ -57,6 +58,7 @@ def test_weight_only(self, precision: str):
5758
def test_int8_kv_cache(self):
5859
self.run(kv_cache_quant_algo=QuantAlgo.INT8)
5960

61+
@skip_post_blackwell
6062
@parametrize_with_ids("per_token,per_channel", [(False, False),
6163
(True, True)])
6264
def test_smooth_quant(self, per_token: bool, per_channel: bool):
@@ -142,6 +144,7 @@ class TestStarcoder2_15B(CliFlowAccuracyTestHarness):
142144
MODEL_PATH = f"{llm_models_root()}/starcoder2-model"
143145
EXAMPLE_FOLDER = "gpt"
144146

147+
@skip_post_blackwell
145148
def test_smooth_quant_ootb(self):
146149
self.run(tasks=[Humaneval(self.MODEL_NAME)],
147150
quant_algo=QuantAlgo.W8A8_SQ_PER_CHANNEL)
@@ -194,9 +197,11 @@ class TestPhi2(CliFlowAccuracyTestHarness):
194197
MODEL_PATH = f"{llm_models_root()}/phi-2"
195198
EXAMPLE_FOLDER = "phi"
196199

200+
@skip_post_blackwell
197201
def test_auto_dtype(self):
198202
self.run(dtype='auto')
199203

204+
@skip_post_blackwell
200205
@pytest.mark.skip_less_device(2)
201206
def test_tp2(self):
202207
self.run(tp_size=2)
@@ -316,6 +321,7 @@ def test_medusa(self, cuda_graph, mocker):
316321
extra_build_args=["--speculative_decoding_mode=medusa"],
317322
extra_summarize_args=extra_summarize_args)
318323

324+
@skip_post_blackwell
319325
@parametrize_with_ids("cuda_graph,chunked_context,typical_acceptance",
320326
[(False, False, False), (True, False, False),
321327
(True, True, False), (True, False, True)])
@@ -360,6 +366,7 @@ def test_beam_search(self):
360366
extra_build_args=["--max_beam_width=5"],
361367
extra_summarize_args=["--num_beams=5"])
362368

369+
@skip_post_blackwell
363370
def test_int4_gptq(self):
364371
self.run(
365372
quant_algo=QuantAlgo.W4A16_GPTQ,
@@ -386,6 +393,7 @@ class TestLlama2_7B(CliFlowAccuracyTestHarness):
386393
def test_auto_dtype(self):
387394
self.run(dtype='auto')
388395

396+
@skip_post_blackwell
389397
def test_smooth_quant(self):
390398
self.run(quant_algo=QuantAlgo.W8A8_SQ_PER_CHANNEL_PER_TOKEN_PLUGIN)
391399

@@ -433,21 +441,25 @@ def test_fp8_low_latency_gemm_plugin(self):
433441
extra_build_args=["--low_latency_gemm_plugin=fp8"])
434442

435443
@pytest.mark.skip_less_device(2)
444+
@skip_post_blackwell
436445
def test_smooth_quant_ootb_tp2(self):
437446
self.run(quant_algo=QuantAlgo.W8A8_SQ_PER_CHANNEL, tp_size=2)
438447

439448
@pytest.mark.skip_less_device(2)
449+
@skip_post_blackwell
440450
def test_int4_awq_tp2(self):
441451
self.run(quant_algo=QuantAlgo.W4A16_AWQ, tp_size=2)
442452

443453
@pytest.mark.skip_less_device(2)
454+
@skip_post_blackwell
444455
def test_int4_awq_prequantized_tp2(self, mocker):
445456
mocker.patch.object(
446457
self.__class__, "MODEL_PATH",
447458
f"{llm_models_root()}/llama-models-v2/Llama-2-7B-AWQ")
448459
self.run(quant_algo=QuantAlgo.W4A16_AWQ, tp_size=2)
449460

450461
@pytest.mark.skip_less_device(2)
462+
@skip_post_blackwell
451463
def test_int4_gptq_prequantized_tp2(self, mocker):
452464
mocker.patch.object(
453465
self.__class__, "MODEL_PATH",
@@ -469,16 +481,19 @@ def test_auto_dtype(self):
469481
def test_float32(self):
470482
self.run(dtype='float32')
471483

484+
@skip_post_blackwell
472485
@pytest.mark.parametrize("precision", ["int8", "int4"])
473486
def test_weight_only(self, precision: str):
474487
quant_algo = QuantAlgo.W8A16 if precision == "int8" else QuantAlgo.W4A16
475488
self.run(quant_algo=quant_algo)
476489

490+
@skip_post_blackwell
477491
@pytest.mark.parametrize("precision", ["int8", "int4"])
478492
def test_weight_only_int8_kv_cache(self, precision: str):
479493
quant_algo = QuantAlgo.W8A16 if precision == "int8" else QuantAlgo.W4A16
480494
self.run(quant_algo=quant_algo, kv_cache_quant_algo=QuantAlgo.INT8)
481495

496+
@skip_post_blackwell
482497
@pytest.mark.parametrize("precision", ["int8", "int4"])
483498
def test_weight_only_manage_weights(self, precision: str):
484499
quant_algo = QuantAlgo.W8A16 if precision == "int8" else QuantAlgo.W4A16
@@ -567,6 +582,7 @@ class TestLlama3_1_8B(CliFlowAccuracyTestHarness):
567582
def test_auto_dtype(self):
568583
self.run(dtype='auto')
569584

585+
@skip_post_blackwell
570586
def test_smooth_quant(self):
571587
self.run(quant_algo=QuantAlgo.W8A8_SQ_PER_CHANNEL_PER_TOKEN_PLUGIN)
572588

@@ -575,12 +591,14 @@ def test_fp8(self):
575591
self.run(quant_algo=QuantAlgo.FP8, kv_cache_quant_algo=QuantAlgo.FP8)
576592

577593
@skip_pre_ada
594+
@skip_post_blackwell
578595
def test_fp8_rowwise(self):
579596
self.run(tasks=[CnnDailymail(self.MODEL_NAME),
580597
MMLU(self.MODEL_NAME)],
581598
quant_algo=QuantAlgo.FP8_PER_CHANNEL_PER_TOKEN)
582599

583600
@skip_pre_ada
601+
@skip_post_blackwell
584602
def test_fp8_rowwise_meta_recipe(self):
585603
self.run(quant_algo=QuantAlgo.FP8_PER_CHANNEL_PER_TOKEN,
586604
extra_acc_spec="meta_recipe",
@@ -601,6 +619,7 @@ def test_tp4(self, gemm_allreduce: bool):
601619
extra_build_args=extra_build_args)
602620

603621
@skip_pre_ada
622+
@skip_post_blackwell
604623
@pytest.mark.skip_less_device(4)
605624
@pytest.mark.parametrize(
606625
"gemm_allreduce", [False, pytest.param(True, marks=skip_no_nvls)],
@@ -646,6 +665,7 @@ def test_fp8_prequantized(self, mocker):
646665
self.run(quant_algo=QuantAlgo.FP8, kv_cache_quant_algo=QuantAlgo.FP8)
647666

648667
@skip_pre_ada
668+
@skip_post_blackwell
649669
def test_medusa_fp8_prequantized(self, mocker):
650670
# nvidia/Llama-3.1-8B-Medusa-FP8
651671
mocker.patch.object(self.__class__, "MODEL_PATH",
@@ -670,23 +690,29 @@ class TestLlama3_2_1B(CliFlowAccuracyTestHarness):
670690
def test_auto_dtype(self):
671691
self.run(dtype='auto')
672692

693+
@skip_post_blackwell
673694
def test_smooth_quant(self):
674695
self.run(quant_algo=QuantAlgo.W8A8_SQ_PER_CHANNEL_PER_TOKEN_PLUGIN)
675696

697+
@skip_post_blackwell
676698
def test_smooth_quant_ootb(self):
677699
self.run(quant_algo=QuantAlgo.W8A8_SQ_PER_CHANNEL)
678700

701+
@skip_post_blackwell
679702
def test_smooth_quant_ootb_manage_weights(self):
680703
self.run(quant_algo=QuantAlgo.W8A8_SQ_PER_CHANNEL,
681704
extra_build_args=["--fast_build"])
682705

706+
@skip_post_blackwell
683707
def test_int4_awq(self):
684708
self.run(quant_algo=QuantAlgo.W4A16_AWQ)
685709

710+
@skip_post_blackwell
686711
def test_int4_awq_int8_kv_cache(self):
687712
self.run(quant_algo=QuantAlgo.W4A16_AWQ,
688713
kv_cache_quant_algo=QuantAlgo.INT8)
689714

715+
@skip_post_blackwell
690716
def test_int4_awq_manage_weights(self):
691717
self.run(quant_algo=QuantAlgo.W4A16_AWQ,
692718
extra_build_args=["--fast_build"])
@@ -733,10 +759,12 @@ def test_fp8_pp2(self):
733759
pp_size=2)
734760

735761
@skip_pre_ada
762+
@skip_post_blackwell
736763
def test_fp8_rowwise(self):
737764
self.run(quant_algo=QuantAlgo.FP8_PER_CHANNEL_PER_TOKEN)
738765

739766
@skip_pre_ada
767+
@skip_post_blackwell
740768
def test_fp8_rowwise_meta_recipe(self):
741769
self.run(quant_algo=QuantAlgo.FP8_PER_CHANNEL_PER_TOKEN,
742770
extra_acc_spec="meta_recipe",
@@ -830,6 +858,7 @@ def test_weight_only(self, precision: str):
830858
quant_algo = QuantAlgo.W8A16 if precision == "int8" else QuantAlgo.W4A16
831859
self.run(quant_algo=quant_algo, extra_convert_args=["--ckpt-type=hf"])
832860

861+
@skip_post_blackwell
833862
def test_smooth_quant(self):
834863
self.run(quant_algo=QuantAlgo.W8A8_SQ_PER_CHANNEL_PER_TOKEN_PLUGIN,
835864
extra_convert_args=[
@@ -841,6 +870,7 @@ def test_smooth_quant(self):
841870
def test_fp8(self):
842871
self.run(quant_algo=QuantAlgo.FP8, kv_cache_quant_algo=QuantAlgo.FP8)
843872

873+
@skip_post_blackwell
844874
def test_int4_awq(self):
845875
self.run(quant_algo=QuantAlgo.W4A16_AWQ)
846876

@@ -859,6 +889,7 @@ def test_weight_only(self, precision: str):
859889
quant_algo = QuantAlgo.W8A16 if precision == "int8" else QuantAlgo.W4A16
860890
self.run(quant_algo=quant_algo, extra_convert_args=["--ckpt-type=hf"])
861891

892+
@skip_post_blackwell
862893
@pytest.mark.skip_less_device_memory(50000)
863894
def test_smooth_quant(self):
864895
self.run(quant_algo=QuantAlgo.W8A8_SQ_PER_CHANNEL_PER_TOKEN_PLUGIN,
@@ -871,6 +902,7 @@ def test_smooth_quant(self):
871902
def test_fp8(self):
872903
self.run(quant_algo=QuantAlgo.FP8, kv_cache_quant_algo=QuantAlgo.FP8)
873904

905+
@skip_post_blackwell
874906
def test_int4_awq(self):
875907
self.run(quant_algo=QuantAlgo.W4A16_AWQ)
876908

@@ -887,6 +919,7 @@ def test_auto_dtype(self):
887919
dtype='auto',
888920
extra_convert_args=["--ckpt-type=hf"])
889921

922+
@skip_post_blackwell
890923
@pytest.mark.parametrize("precision", ["int8", "int4"])
891924
def test_weight_only(self, precision: str):
892925
quant_algo = QuantAlgo.W8A16 if precision == "int8" else QuantAlgo.W4A16
@@ -910,6 +943,7 @@ def test_auto_dtype(self):
910943
def test_weight_only(self):
911944
self.run(quant_algo=QuantAlgo.W8A16)
912945

946+
@skip_post_blackwell
913947
def test_int4_gptq_prequantized(self, mocker):
914948
mocker.patch.object(self.__class__, "MODEL_PATH",
915949
f"{llm_models_root()}/Qwen-7B-Chat-Int4")
@@ -938,6 +972,7 @@ class TestQwen2_0_5BInstruct(CliFlowAccuracyTestHarness):
938972
def test_auto_dtype(self):
939973
self.run(dtype='auto')
940974

975+
@skip_post_blackwell
941976
def test_weight_only(self):
942977
self.run(quant_algo=QuantAlgo.W8A16)
943978

@@ -956,9 +991,11 @@ class TestQwen2_7BInstruct(CliFlowAccuracyTestHarness):
956991
def test_auto_dtype(self):
957992
self.run(dtype='auto')
958993

994+
@skip_post_blackwell
959995
def test_weight_only(self):
960996
self.run(quant_algo=QuantAlgo.W8A16)
961997

998+
@skip_post_blackwell
962999
def test_int4_awq_prequantized(self, mocker):
9631000
mocker.patch.object(self.__class__, "MODEL_PATH",
9641001
f"{llm_models_root()}/Qwen2-7B-Instruct-AWQ")
@@ -990,6 +1027,7 @@ class TestQwen2_5_1_5BInstruct(CliFlowAccuracyTestHarness):
9901027
def test_auto_dtype(self):
9911028
self.run(dtype='auto')
9921029

1030+
@skip_post_blackwell
9931031
def test_weight_only(self):
9941032
self.run(quant_algo=QuantAlgo.W8A16)
9951033

‎tests/integration/defs/accuracy/test_llm_api.py‎

Lines changed: 4 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -18,7 +18,7 @@
1818
from tensorrt_llm.models.modeling_utils import QuantConfig
1919
from tensorrt_llm.quantization import QuantAlgo
2020

21-
from ..conftest import llm_models_root, skip_pre_ada
21+
from ..conftest import llm_models_root, skip_post_blackwell, skip_pre_ada
2222
from .accuracy_core import MMLU, CnnDailymail, LlmapiAccuracyTestHarness
2323

2424

@@ -27,6 +27,7 @@ class TestLlama3_1_8B(LlmapiAccuracyTestHarness):
2727
MODEL_PATH = f"{llm_models_root()}/llama-3.1-model/Meta-Llama-3.1-8B"
2828

2929
@skip_pre_ada
30+
@skip_post_blackwell
3031
def test_fp8_rowwise(self):
3132
quant_config = QuantConfig(QuantAlgo.FP8_PER_CHANNEL_PER_TOKEN)
3233

@@ -65,6 +66,7 @@ def test_auto_dtype(self):
6566
task.evaluate(llm,
6667
extra_evaluator_kwargs=self.EXTRA_EVALUATOR_KWARGS)
6768

69+
@skip_post_blackwell
6870
def test_weight_only(self):
6971
quant_config = QuantConfig(QuantAlgo.W8A16)
7072
with LLM(self.MODEL_PATH, quant_config=quant_config) as llm:
@@ -133,6 +135,7 @@ def test_auto_dtype(self):
133135
task = MMLU(self.MODEL_NAME)
134136
task.evaluate(llm)
135137

138+
@skip_post_blackwell
136139
def test_weight_only(self):
137140
quant_config = QuantConfig(QuantAlgo.W8A16)
138141
with LLM(self.MODEL_PATH, quant_config=quant_config) as llm:

‎tests/integration/defs/examples/test_eagle.py‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -22,6 +22,7 @@
2222
from defs.trt_test_alternative import check_call
2323

2424

25+
@skip_post_blackwell
2526
@pytest.mark.parametrize("use_dynamic_tree", [False, True],
2627
ids=['eagle1', 'eagle2'])
2728
@pytest.mark.parametrize("batch_size", [1, 8], ids=['bs1', 'bs8'])

‎tests/integration/defs/examples/test_exaone.py‎

Lines changed: 3 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -17,6 +17,7 @@
1717
import pytest
1818
from defs.common import (convert_weights, generate_summary_cmd, venv_check_call,
1919
venv_mpi_check_call)
20+
from defs.conftest import skip_post_blackwell
2021
from defs.trt_test_alternative import check_call
2122

2223

@@ -26,7 +27,8 @@
2627
@pytest.mark.parametrize("llm_exaone_model_root",
2728
['exaone_3.0_7.8b_instruct', 'exaone_deep_2.4b'],
2829
indirect=True)
29-
@pytest.mark.parametrize("use_weight_only", [True, False],
30+
@pytest.mark.parametrize("use_weight_only",
31+
[pytest.param(True, marks=skip_post_blackwell), False],
3032
ids=["enable_weight_only", "disable_weight_only"])
3133
def test_llm_exaone_1gpu(data_type, exaone_example_root, llm_exaone_model_root,
3234
llama_example_root, llm_datasets_root, llm_rouge_root,

‎tests/integration/defs/examples/test_mixtral.py‎

Lines changed: 3 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -23,7 +23,8 @@
2323
venv_mpi_check_call)
2424
from defs.conftest import (evaltool_mmlu_post_process,
2525
evaltool_wikilingua_post_process, llm_models_root,
26-
skip_pre_ada, skip_pre_blackwell)
26+
skip_post_blackwell, skip_pre_ada,
27+
skip_pre_blackwell)
2728
from defs.trt_test_alternative import check_call
2829
from evaltool.constants import (EVALTOOL_INFERENCE_SERVER_STARTUP_SCRIPT,
2930
EVALTOOL_INFERENCE_SERVER_STOP_SCRIPT,
@@ -888,6 +889,7 @@ def test_llm_mixtral_1gpu_fp4_llmapi(
888889
venv_check_call(llm_venv, mmlu_cmd)
889890

890891

892+
@skip_post_blackwell
891893
@pytest.mark.parametrize(
892894
"model_name", ['mixtral-8x7b-v0.1-AWQ', 'Mixtral-8x7B-Instruct-v0.1'])
893895
def test_llm_mixtral_int4_awq_1gpu_summary(llama_example_root,

‎tests/integration/defs/examples/test_multimodal.py‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -17,7 +17,7 @@
1717

1818
import pytest
1919
from defs.common import convert_weights, venv_check_call, venv_mpi_check_call
20-
from defs.conftest import get_device_memory, skip_pre_ada
20+
from defs.conftest import get_device_memory, skip_post_blackwell, skip_pre_ada
2121
from defs.trt_test_alternative import check_call
2222

2323

@@ -616,7 +616,7 @@ def _test_llm_multimodal_general(llm_venv,
616616
'neva-22b',
617617
'kosmos-2',
618618
'video-neva',
619-
'Phi-3-vision-128k-instruct',
619+
pytest.param('Phi-3-vision-128k-instruct', marks=skip_post_blackwell),
620620
'Phi-3.5-vision-instruct',
621621
'Phi-4-multimodal-instruct',
622622
'Llama-3.2-11B-Vision',

‎tests/integration/defs/examples/test_phi.py‎

Lines changed: 5 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -24,7 +24,8 @@
2424
evaltool_mmlu_post_process,
2525
evaltool_mtbench_post_process,
2626
evaltool_wikilingua_post_process, get_device_memory,
27-
skip_fp8_pre_ada, skip_pre_ada)
27+
get_sm_version, skip_fp8_pre_ada,
28+
skip_post_blackwell, skip_pre_ada)
2829
from defs.trt_test_alternative import check_call
2930
from evaltool.constants import (EVALTOOL_INFERENCE_SERVER_STARTUP_SCRIPT,
3031
EVALTOOL_INFERENCE_SERVER_STOP_SCRIPT,
@@ -421,6 +422,8 @@ def test_llm_phi_lora_1gpu(data_type, lora_data_type, phi_example_root,
421422
model_name = 'phi-3-lora'
422423
if data_type == 'fp8':
423424
skip_fp8_pre_ada(use_fp8=True)
425+
if get_sm_version() >= 100:
426+
pytest.skip("FP8 is not supported on post-Blackwell architectures")
424427
model_dir = quantize_data(
425428
llm_venv,
426429
phi_example_root,
@@ -570,6 +573,7 @@ def test_llm_phi_quantization_1gpu(data_type, llm_phi_model_root, llm_venv,
570573

571574

572575
@skip_pre_ada
576+
@skip_post_blackwell
573577
@pytest.mark.parametrize("llm_phi_model_root", [
574578
"phi-2", "Phi-3-mini-128k-instruct", "Phi-3-small-128k-instruct",
575579
"Phi-3.5-mini-instruct", "Phi-3.5-MoE-instruct", "Phi-4-mini-instruct"

0 commit comments

Comments
 (0)