A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 13, 15:45 UTC
[CI] Reduce Blackwell Fusion test runtime by filtering tests and only run all tests in nightly (#28074)
Copilot committed
Nov 7, 2025 at 15:58 UTC
a736e5ff770bd0e69492c24d24d04b7fd47789c0
2 files changed
+31
-8
.buildkite/test-pipeline.yaml
+26
-1
@@ -472,7 +472,9 @@ steps:
472
- tests/compile
473
commands:
474
- pytest -v -s compile/test_full_graph.py
475
- - pytest -v -s compile/test_fusions_e2e.py
475
+ # Limit to no custom ops to reduce running time
476
+ # Wrap with quotes to escape yaml and avoid starting -k string with a -
477
+ - "pytest -v -s compile/test_fusions_e2e.py -k 'TRITON and -quant_fp8'"
478
479
- label: Cudagraph test
480
timeout_in_minutes: 20
@@ -929,6 +931,29 @@ steps:
931
- pytest -v -s tests/compile/test_silu_mul_quant_fusion.py
932
# this runner has 2 GPUs available even though num_gpus=2 is not set
933
- pytest -v -s tests/compile/test_fusion_all_reduce.py
934
+ # Limit to Inductor partition, no custom ops, and allreduce & attn fusion to reduce running time
935
+ # Wrap with quotes to escape yaml
936
+ - "pytest -v -s tests/compile/test_fusions_e2e.py::test_tp2_attn_quant_allreduce_rmsnorm -k 'True and Llama-3.1 and -quant_fp8 and -rms_norm'"
937
+
938
+- label: Blackwell Fusion E2E Tests # 30 min
939
+ timeout_in_minutes: 40
940
+ working_dir: "/vllm-workspace/"
941
+ gpu: b200
942
+ optional: true
943
+ num_gpus: 2
944
+ source_file_dependencies:
945
+ - csrc/quantization/fp4/
946
+ - vllm/model_executor/layers/quantization/utils/flashinfer_utils.py
947
+ - vllm/v1/attention/backends/flashinfer.py
948
+ - vllm/compilation/
949
+ # can affect pattern matching
950
+ - vllm/model_executor/layers/layernorm.py
951
+ - vllm/model_executor/layers/activation.py
952
+ - vllm/model_executor/layers/quantization/input_quant_fp8.py
953
+ - tests/compile/test_fusions_e2e.py
954
+ commands:
955
+ - nvidia-smi
956
+ # Run all e2e fusion tests
957
- pytest -v -s tests/compile/test_fusions_e2e.py
958
959
- label: Blackwell GPT-OSS Eval
tests/compile/test_fusions_e2e.py
+5
-7
@@ -54,11 +54,11 @@ if current_platform.is_cuda():
54
55
MODELS_FP4 = [
56
ModelBackendTestCase(
57
- model_name="nvidia/Llama-4-Scout-17B-16E-Instruct-FP4",
57
+ model_name="nvidia/Llama-3.1-8B-Instruct-FP4",
58
model_kwargs=dict(max_model_len=1024, kv_cache_dtype="fp8"),
59
backend=_Backend.FLASHINFER,
60
- attention_fusions=48,
61
- allreduce_fusions=96,
60
+ attention_fusions=32,
61
+ allreduce_fusions=65,
62
),
63
]
64
@@ -95,8 +95,7 @@ elif current_platform.is_rocm():
95
),
96
]
97
98
-# TODO(luka) test both in nightly
99
-CUSTOM_OPS_FP8 = ["-quant_fp8"] # , "+quant_fp8"]
98
+CUSTOM_OPS_FP8 = ["-quant_fp8", "+quant_fp8"]
99
100
101
@pytest.mark.parametrize(
@@ -171,8 +170,7 @@ def test_attn_quant(
170
assert int(matches[0]) == attention_fusions
171
172
174
-# TODO(luka) test both in nightly
175
-CUSTOM_OPS_RMS_NORM = ["-rms_norm"] # , "+rms_norm"]
173
+CUSTOM_OPS_RMS_NORM = ["-rms_norm", "+rms_norm"]
174
175
176
def custom_ops_product(*custom_ops_lists: list[str]) -> Iterable[str]: