Skip to content

[https://nvbugs/5774869][infra] Use 2 GPUs to test skip softmax attention on H100. #10420

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
bobboli merged 4 commits into from
Jan 14, 2026
Merged

[https://nvbugs/5774869][infra] Use 2 GPUs to test skip softmax attention on H100. #10420

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
15ef830
Use 2 GPUs to test skip softmax attention on H100.
bobboli Jan 5, 2026
5a662a6
Update nvbugs for waive on Blackwell.
bobboli Jan 5, 2026
b07c566
Update ref accuracy.
bobboli Jan 6, 2026
b970446
Turn off KV block reuse.
bobboli Jan 10, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 3 additions & 3 deletions tests/integration/defs/accuracy/references/longbench_v1.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,8 +1,8 @@
Qwen3/Qwen3-30B-A3B-Instruct-2507:
# Skip Softmax Attention ref accuracy
- extra_acc_spec: "target_sparsity=0.0"
accuracy: 47.22
accuracy: 47.357
- extra_acc_spec: "target_sparsity=0.5"
accuracy: 47.22
accuracy: 47.102
- extra_acc_spec: "target_sparsity=0.9"
accuracy: 45.90
accuracy: 46.169
41 changes: 38 additions & 3 deletions tests/integration/defs/accuracy/test_llm_api_pytorch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3884,7 +3884,6 @@ class TestQwen3_30B_A3B_Instruct_2507(LlmapiAccuracyTestHarness):
MODEL_PATH = f"{llm_models_root()}/{MODEL_NAME}"

@skip_pre_hopper
# @pytest.mark.skip_less_device_memory(140000) # Only test for H200, B200
@pytest.mark.parametrize(
"target_sparsity,thr_prefill,thr_decode",
[
Expand All @@ -3903,10 +3902,11 @@ def test_skip_softmax_attention(self, target_sparsity: float,
"prefill": thr_prefill,
"decode": thr_decode,
})
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.85)
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.75,
enable_block_reuse=False)

if get_sm_version() >= 100:
pytest.skip("Bug to be fixed on Blackwell")
pytest.skip("https://nvbugs/5783509: Bug to be fixed on Blackwell")

with LLM(self.MODEL_PATH,
attn_backend="TRTLLM",
Expand All @@ -3918,6 +3918,41 @@ def test_skip_softmax_attention(self, target_sparsity: float,
task.evaluate(llm,
extra_acc_spec=f"target_sparsity={target_sparsity}")

@pytest.mark.parametrize(
"target_sparsity,thr_prefill,thr_decode",
[
(0.0, 0.0, 0.0),
(0.5, 85.97384174442398, 55.48258322852407),
(0.9, 1418.142868970396, 863.147841750025),
],
ids=[
"target_sparsity_0.0", "target_sparsity_0.5", "target_sparsity_0.9"
],
)
def test_skip_softmax_attention_2gpus(self, target_sparsity: float,
thr_prefill: float,
thr_decode: float):
sparse_attention_config = SkipSoftmaxAttentionConfig(
threshold_scale_factor={
"prefill": thr_prefill,
"decode": thr_decode,
})
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.75,
enable_block_reuse=False)

with LLM(self.MODEL_PATH,
attn_backend="TRTLLM",
max_batch_size=256,
max_num_tokens=100000,
tensor_parallel_size=2,
moe_expert_parallel_size=2,
enable_attention_dp=True,
kv_cache_config=kv_cache_config,
sparse_attention_config=sparse_attention_config) as llm:
task = LongBenchV1(self.MODEL_NAME)
task.evaluate(llm,
extra_acc_spec=f"target_sparsity={target_sparsity}")


class TestPhi4MiniInstruct(LlmapiAccuracyTestHarness):
MODEL_NAME = "microsoft/Phi-4-mini-instruct"
Expand Down
1 change: 0 additions & 1 deletion tests/integration/test_lists/test-db/l0_b200.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -55,7 +55,6 @@ l0_b200:
- 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_0.0]
- accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B_Instruct_2507::test_skip_softmax_attention[target_sparsity_0.5]
- accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B_Instruct_2507::test_skip_softmax_attention[target_sparsity_0.9]
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_nvfp4[tp1-cutlass]
Expand Down
3 changes: 3 additions & 0 deletions tests/integration/test_lists/test-db/l0_dgx_h100.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -47,6 +47,9 @@ l0_dgx_h100:
- accuracy/test_llm_api_autodeploy.py::TestLlama3_1_8B::test_auto_dtype[False-2]
# llmapi
- unittest/llmapi/test_mpi_session.py::test_llmapi_launch_multiple_tasks
# ------------- Skip softmax attention tests ---------------
- accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B_Instruct_2507::test_skip_softmax_attention_2gpus[target_sparsity_0.5]
- accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B_Instruct_2507::test_skip_softmax_attention_2gpus[target_sparsity_0.9]
- condition:
ranges:
system_gpu_count:
Expand Down
4 changes: 0 additions & 4 deletions tests/integration/test_lists/test-db/l0_h100.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -78,10 +78,6 @@ l0_h100:
- 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
# Waive known failures in https://nvbugs/5774869
# - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B_Instruct_2507::test_skip_softmax_attention[target_sparsity_0.0] TIMEOUT (90)
Copy link
Collaborator

@SimengLiu-nv SimengLiu-nv Jan 20, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@bobboli Would you fill a new PR to add back the tests?

# - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B_Instruct_2507::test_skip_softmax_attention[target_sparsity_0.5] TIMEOUT (90)
# - accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B_Instruct_2507::test_skip_softmax_attention[target_sparsity_0.9] TIMEOUT (90)
- accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_eagle3[enable_chunked_prefill=False-eagle3_one_model=False]
- accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_eagle3[enable_chunked_prefill=True-eagle3_one_model=True]
- accuracy/test_llm_api_pytorch.py::TestQwen3_8B::test_eagle3[enable_chunked_prefill=False-eagle3_one_model=True]
Expand Down
3 changes: 0 additions & 3 deletions tests/integration/test_lists/waives.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -337,9 +337,6 @@ accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_fp8_eagle3_tp8[
accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_nvfp4_tp4[torch_compile=False] SKIP (https://nvbugs/5794796)
accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_fp8_tp4[torch_compile=False] SKIP (https://nvbugs/5794796)
accuracy/test_llm_api_pytorch.py::TestQwen3_235B_A22B::test_fp8[throughput_latency] SKIP (https://nvbugs/5775544)
accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B_Instruct_2507::test_skip_softmax_attention[target_sparsity_0.5] SKIP (https://nvbugs/5774869)
accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B_Instruct_2507::test_skip_softmax_attention[target_sparsity_0.0] SKIP (https://nvbugs/5774869)
accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B_Instruct_2507::test_skip_softmax_attention[target_sparsity_0.9] SKIP (https://nvbugs/5774869)
triton_server/test_triton.py::test_llava_onevision[llava_onevision] SKIP (https://nvbugs/5775205)
triton_server/test_triton.py::test_gpt_ib_lad[gpt-ib-lad] SKIP (https://nvbugs/5775223)
unittest/_torch/modules/test_fused_moe.py::test_fused_moe_fp8_blockwise_cute_dsl_multi_gpu[MoEWeightLoadingMode.FUSED_GATE_UP_PROJ-DefaultMoeRoutingMethod-1] SKIP (https://nvbugs/5775256)
Expand Down
Loading