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 9, 17:06 UTC
[Tests] Clarify pytest skip reasons with actionable context (#32981)
Signed-off-by: 7. Sun <jhao.sun@gmail.com>
7. Sun committed
Jan 24, 2026 at 06:38 UTC
0b9a735e113ea97d3a75d5491b6514c03f6081e3
2 files changed
+7
-3
tests/samplers/test_beam_search.py
+2
-2
@@ -20,7 +20,7 @@ MM_BEAM_WIDTHS = [2]
20
MODELS = ["TinyLlama/TinyLlama-1.1B-Chat-v1.0"]
21
22
23
-@pytest.mark.skip_v1 # FIXME: This fails on V1 right now.
23
+@pytest.mark.skip_v1 # V1 engine does not yet support beam search
24
@pytest.mark.parametrize("model", MODELS)
25
@pytest.mark.parametrize("dtype", ["half"])
26
@pytest.mark.parametrize("max_tokens", MAX_TOKENS)
@@ -62,7 +62,7 @@ def test_beam_search_single_input(
62
)
63
64
65
-@pytest.mark.skip_v1 # FIXME: This fails on V1 right now.
65
+@pytest.mark.skip_v1 # V1 engine does not yet support beam search
66
@pytest.mark.parametrize("model", MODELS)
67
@pytest.mark.parametrize("dtype", ["half"])
68
@pytest.mark.parametrize("max_tokens", MAX_TOKENS)
tests/v1/sample/test_topk_topp_sampler.py
+5
-1
@@ -48,7 +48,11 @@ def test_topk_impl_equivalence():
48
assert torch.allclose(result1, result2)
49
50
51
-@pytest.mark.skip(reason="FIXME: This test is failing right now.")
51
+@pytest.mark.skip(
52
+ reason="FlashInfer top-k/top-p renorm comparison fails; "
53
+ "needs investigation of tolerance threshold or "
54
+ "interface differences between Python and FlashInfer implementations"
55
+)
56
def test_flashinfer_sampler():
57
"""
58
This test verifies that the FlashInfer top-k and top-p sampling