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 11, 08:45 UTC
[CI] Skip ci failure test (#30804)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Dec 16, 2025 at 17:47 UTC
b6ec077e058e15e5b853793924e6643ec6c579aa
1 file changed
+3
-1
tests/compile/distributed/test_fusions_e2e.py
+3
-1
@@ -523,6 +523,8 @@ CUSTOM_OPS_QUANT_RMS_NORM = ["+quant_fp8,+rms_norm"]
523
list[tuple[Any, ...]](flat_product(MODELS_GROUP_FP8, CUSTOM_OPS_QUANT_RMS_NORM)),
524
)
525
@pytest.mark.parametrize("inductor_graph_partition", [True, False])
526
+# TODO: remove skip after we fix the fusion thoroughly
527
+@pytest.mark.skipif(is_blackwell(), reason="Temporarily disabled on Blackwell")
528
def test_rms_group_quant(
529
model_name: str,
530
model_kwargs: dict[str, Any],
@@ -562,7 +564,7 @@ def test_rms_group_quant(
564
splitting_ops=splitting_ops,
565
# Common
566
mode=CompilationMode.VLLM_COMPILE,
565
- pass_config=PassConfig(eliminate_noops=True, enable_fusion=True),
567
+ pass_config=PassConfig(eliminate_noops=True, fuse_norm_quant=True),
568
# Inductor caches custom passes by default as well via uuid
569
inductor_compile_config={"force_disable_caches": True},
570
)