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 15, 04:45 UTC
Fixing Chunked Prefill Test. (#19762)
Signed-off-by: Alexei V. Ivanov <alexei.ivanov@amd.com>
Alexei-V-Ivanov-AMD committed
Jun 19, 2025 at 03:36 UTC
4719460644b4629db2b6dbf12be331d0b34b4b6f
2 files changed
+17
-3
.buildkite/test-pipeline.yaml
+1
-1
@@ -89,7 +89,7 @@ steps:
89
- VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest -v -s basic_correctness/test_preemption.py
90
91
- label: Chunked Prefill Test
92
- mirror_hardwares: [amdexperimental]
92
+ mirror_hardwares: [amdexperimental, amdproduction]
93
source_file_dependencies:
94
- vllm/
95
- tests/basic_correctness/test_chunked_prefill
tests/basic_correctness/test_chunked_prefill.py
+16
-2
@@ -49,7 +49,13 @@ def use_v0_only(monkeypatch: pytest.MonkeyPatch):
49
# NOTE: Increasing this in this suite will fail CI because we currently cannot
50
# reset distributed env properly. Use a value > 1 just when you test.
51
@pytest.mark.parametrize("tensor_parallel_size", [1])
52
-@pytest.mark.parametrize("attention_backend", ["FLASHINFER", "FLASH_ATTN"])
52
+@pytest.mark.parametrize("attention_backend", [
53
+ pytest.param("FLASHINFER",
54
+ marks=pytest.mark.skipif(
55
+ current_platform.is_rocm(),
56
+ reason="FLASHINFER isn't supported on ROCm")),
57
+ "FLASH_ATTN"
58
+])
59
def test_models(
60
hf_runner: HfRunner,
61
vllm_runner: VllmRunner,
@@ -99,7 +105,13 @@ def test_models(
105
@multi_gpu_test(num_gpus=2)
106
@pytest.mark.parametrize("distributed_executor_backend", ["ray", "mp"])
107
@pytest.mark.parametrize("model", MODELS)
102
-@pytest.mark.parametrize("attention_backend", ["FLASHINFER", "FLASH_ATTN"])
108
+@pytest.mark.parametrize("attention_backend", [
109
+ pytest.param("FLASHINFER",
110
+ marks=pytest.mark.skipif(
111
+ current_platform.is_rocm(),
112
+ reason="FLASHINFER isn't supported on ROCm")),
113
+ "FLASH_ATTN"
114
+])
115
def test_models_distributed(
116
hf_runner: HfRunner,
117
vllm_runner: VllmRunner,
@@ -172,6 +184,8 @@ def test_models_distributed(
184
# Due to low-precision numerical divergence, this test is too sensitive to
185
# the async postprocessor
186
@pytest.mark.parametrize("disable_async_output_proc", [True])
187
+@pytest.mark.skipif(current_platform.is_rocm(),
188
+ reason="machete_prepack_B isn't supported on ROCm")
189
def test_models_with_fp8_kv_cache(
190
vllm_runner: VllmRunner,
191
example_prompts,