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 5, 21:00 UTC
[ROCm][CI] Enable hybrid chunked prefill test (#38317)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Andreas Karatzas committed
Mar 29, 2026 at 21:30 UTC
4f2ed5fddb52b45ea8a8b0dad4a354802a6b3bea
2 files changed
+21
-2
.buildkite/test-amd.yaml
+13
@@ -1801,6 +1801,19 @@ steps:
1801
- tests/v1/e2e
1802
commands:
1803
- pytest -v -s v1/e2e/spec_decode/test_spec_decode.py -k "eagle_correctness_heavy"
1804
+
1805
+
1806
+- label: V1 e2e (4xH100-4xMI325) # TBD
1807
+ timeout_in_minutes: 180
1808
+ mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi325]
1809
+ agent_pool: mi325_4
1810
+ optional: true
1811
+ source_file_dependencies:
1812
+ - vllm/v1/attention/backends/utils.py
1813
+ - vllm/v1/worker/gpu_model_runner.py
1814
+ - tests/v1/e2e/test_hybrid_chunked_prefill.py
1815
+ commands:
1816
+ - pytest -v -s v1/e2e/test_hybrid_chunked_prefill.py
1817
1818
1819
- label: V1 Spec Decode # TBD
tests/v1/e2e/test_hybrid_chunked_prefill.py
+8
-2
@@ -36,14 +36,20 @@ MESSAGES = [
36
]
37
38
39
-@pytest.mark.skipif(not current_platform.is_cuda(), reason="CUDA not available")
39
@pytest.mark.parametrize(
40
"model_name",
41
[
42
pytest.param("Qwen/Qwen3.5-4B", marks=[large_gpu_mark(min_gb=40)]),
43
pytest.param(
44
"nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8",
46
- marks=[large_gpu_mark(min_gb=80)] + multi_gpu_marks(num_gpus=4),
45
+ marks=[large_gpu_mark(min_gb=80)]
46
+ + multi_gpu_marks(num_gpus=4)
47
+ + [
48
+ pytest.mark.skipif(
49
+ not current_platform.is_cuda(),
50
+ reason="modelopt quantization is supported only on CUDA",
51
+ )
52
+ ],
53
),
54
],
55
)