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 16, 18:15 UTC
[aDAG] Unflake aDAG + PP tests (#7600)
SangBin Cho committed
Aug 16, 2024 at 20:49 UTC
4706eb628e0830f8ab4db5e9f5ede3bc4eee2b76
3 files changed
+9
-4
.buildkite/test-pipeline.yaml
+2
-2
@@ -314,11 +314,11 @@ steps:
314
num_gpus: 4
315
source_file_dependencies:
316
- vllm/
317
- - tests/distributed/test_pipeline_parallel
317
- tests/distributed/test_pp_cudagraph.py
318
+ - tests/distributed/test_pipeline_parallel
319
commands:
320
- - pytest -v -s distributed/test_pipeline_parallel.py
320
- pytest -v -s distributed/test_pp_cudagraph.py
321
+ - pytest -v -s distributed/test_pipeline_parallel.py
322
323
- label: LoRA Long Context (Distributed) # 11min
324
# This test runs llama 13B, so it is required to run on 4 GPUs.
tests/distributed/test_pipeline_parallel.py
+4
@@ -80,6 +80,10 @@ def test_compare_tp(TP_SIZE, PP_SIZE, EAGER_MODE, CHUNKED_PREFILL, MODEL_NAME,
80
"VLLM_USE_RAY_SPMD_WORKER": "1",
81
"VLLM_USE_RAY_COMPILED_DAG_NCCL_CHANNEL": "1",
82
}
83
+ # Temporary. Currently when zeromq + SPMD is used, it does not properly
84
+ # terminate because of aDAG issue.
85
+ pp_args.append("--disable-frontend-multiprocessing")
86
+ tp_args.append("--disable-frontend-multiprocessing")
87
88
try:
89
compare_two_settings(MODEL_NAME, pp_args, tp_args, pp_env)
tests/utils.py
+3
-2
@@ -384,6 +384,7 @@ def fork_new_process_for_each_test(
384
os.setpgrp()
385
from _pytest.outcomes import Skipped
386
pid = os.fork()
387
+ print(f"Fork a new process to run a test {pid}")
388
if pid == 0:
389
try:
390
f(*args, **kwargs)
@@ -401,11 +402,11 @@ def fork_new_process_for_each_test(
402
pgid = os.getpgid(pid)
403
_pid, _exitcode = os.waitpid(pid, 0)
404
# ignore SIGTERM signal itself
404
- old_singla_handler = signal.signal(signal.SIGTERM, signal.SIG_IGN)
405
+ old_signal_handler = signal.signal(signal.SIGTERM, signal.SIG_IGN)
406
# kill all child processes
407
os.killpg(pgid, signal.SIGTERM)
408
# restore the signal handler
408
- signal.signal(signal.SIGTERM, old_singla_handler)
409
+ signal.signal(signal.SIGTERM, old_signal_handler)
410
assert _exitcode == 0, (f"function {f} failed when called with"
411
f" args {args} and kwargs {kwargs}")
412