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, 18:30 UTC
[Bugfix][CI] Skip flaky `test_eagle` test (#38566)
Signed-off-by: NickLucche <nlucches@redhat.com>
Nicolò Lucchesi committed
Mar 31, 2026 at 15:42 UTC
743038966947513019ec0ba623e687062e75286e
1 file changed
+8
-1
tests/v1/distributed/test_eagle_dp.py
+8
-1
@@ -23,9 +23,16 @@ else:
23
24
@pytest.mark.asyncio
25
@pytest.mark.parametrize("attn_backend", ATTN_BACKENDS)
26
+@pytest.mark.xfail(
27
+ not current_platform.is_rocm(),
28
+ reason="EAGLE + DP > 1 produces wrong outputs when async spec decode "
29
+ "correction is active. Root cause under investigation. "
30
+ "See: https://github.com/vllm-project/vllm/issues/31913",
31
+ strict=False,
32
+)
33
@pytest.mark.xfail(
34
current_platform.is_rocm(),
28
- reason="Test may fail on ROCm until batch invariance is enabled."
35
+ reason="Test may fail on ROCm until batch invariance is enabled. "
36
"See: https://github.com/vllm-project/vllm/issues/27433",
37
strict=False,
38
)