[CI] Run basic fullgraph correctness on one GPU (#51271)
Signed-off-by: mgoin <mgoin64@gmail.com> Co-authored-by: OpenAI Codex <codex@openai.com>
Michael Goin committed
Aug 6, 2026 at 12:09 UTC
566c80edf9e770524b1506a9d681922d4601c70c
2 files changed
-22
.buildkite/test_areas/distributed.yaml
-4
@@ -68,14 +68,12 @@ steps:
68
- vllm/worker/worker_base.py
69
- vllm/v1/engine/
70
- vllm/v1/worker/
71
- - tests/compile/fullgraph/test_basic_correctness.py
71
- tests/compile/test_wrapper.py
72
- tests/entrypoints/llm/test_collective_rpc.py
73
commands:
74
# https://github.com/NVIDIA/nccl/issues/1838
75
- export NCCL_CUMEM_HOST_ENABLE=0
76
- pytest -v -s entrypoints/llm/test_collective_rpc.py
78
- - pytest -v -s ./compile/fullgraph/test_basic_correctness.py
77
- pytest -v -s ./compile/test_wrapper.py
78
79
- label: Distributed Torchrun + Shutdown Tests (2 GPUs)
@@ -163,13 +161,11 @@ steps:
161
- vllm/distributed/
162
- tests/distributed/test_pynccl
163
- tests/distributed/test_events
166
- - tests/compile/fullgraph/test_basic_correctness.py
164
- tests/distributed/test_symm_mem_allreduce.py
165
- tests/distributed/test_multiproc_executor.py
166
commands:
167
# https://github.com/NVIDIA/nccl/issues/1838
168
- export NCCL_CUMEM_HOST_ENABLE=0
172
- - pytest -v -s compile/fullgraph/test_basic_correctness.py
169
- pytest -v -s distributed/test_pynccl.py
170
- pytest -v -s distributed/test_events.py
171
- pytest -v -s distributed/test_symm_mem_allreduce.py
tests/compile/fullgraph/test_basic_correctness.py
-18
@@ -27,15 +27,6 @@ class TestSetting:
27
@pytest.mark.parametrize(
28
"test_setting",
29
[
30
- # basic llama model
31
- TestSetting(
32
- model="meta-llama/Llama-3.2-1B-Instruct",
33
- model_args=["--max-model-len", "2048"],
34
- pp_size=2,
35
- tp_size=2,
36
- attn_backend=ATTN_BACKEND,
37
- method="generate",
38
- ),
30
# MoE model
31
TestSetting(
32
model="ibm-granite/granite-3.0-1b-a400m-instruct",
@@ -68,21 +59,12 @@ class TestSetting:
59
def test_compile_correctness(
60
test_setting: TestSetting,
61
):
71
- # this test is run under multiple suits, with different GPUs.
72
- # make sure we only run the test with correct CUDA devices.
73
- # don't use "<", as it will duplicate the tests.
62
model = test_setting.model
63
model_args = test_setting.model_args
64
pp_size = test_setting.pp_size
65
tp_size = test_setting.tp_size
66
attn_backend = test_setting.attn_backend
67
method = test_setting.method
80
- if current_platform.device_count() < pp_size * tp_size:
81
- pytest.skip(
82
- f"Need at least {pp_size}*{tp_size} CUDA gpus but got "
83
- f"{current_platform.device_count()}"
84
- )
85
-
68
final_args = [
69
*model_args,
70
"-pp",