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, 18:30 UTC
[Misc] Catching Ray Compiled Graph PP test failures for V1 (#14847)
Rui Qiao committed
Mar 16, 2025 at 15:46 UTC
b9b5bdfc7d5cd0f8610a4de7a79327d10a09dfab
1 file changed
+15
-5
tests/distributed/test_pipeline_parallel.py
+15
-5
@@ -350,6 +350,10 @@ def _compare_tp(
350
else:
351
pp_env = None
352
353
+ tp_env = {
354
+ "VLLM_USE_V1": vllm_major_version,
355
+ }
356
+
357
pp_args = [
358
*common_args,
359
"--pipeline-parallel-size",
@@ -374,14 +378,20 @@ def _compare_tp(
378
]
379
380
try:
377
- compare_two_settings(model_id, pp_args, tp_args, pp_env, method=method)
381
+ compare_two_settings(model_id,
382
+ pp_args,
383
+ tp_args,
384
+ pp_env,
385
+ tp_env,
386
+ method=method)
387
except Exception:
379
- if pp_env is None:
380
- raise
381
- else:
382
- # Ray Compiled Graph tests are flaky,
388
+ testing_ray_compiled_graph = pp_env is not None
389
+ if testing_ray_compiled_graph and vllm_major_version == "0":
390
+ # Ray Compiled Graph tests are flaky for V0,
391
# so we don't want to fail the test
392
logger.exception("Ray Compiled Graph tests failed")
393
+ else:
394
+ raise
395
396
397
@pytest.mark.parametrize(