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 12, 04:45 UTC
[CI][ROCm][tests/v1/e2e] Fix multiprocessing launch for the test (#29123)
Signed-off-by: Divakar Verma <divakar.verma@amd.com>
Divakar Verma committed
Dec 2, 2025 at 14:46 UTC
afb1e5b380ff623e478d19a246b42b2903b9331f
1 file changed
+19
-3
tests/v1/e2e/test_kv_sharing_fast_prefill.py
+19
-3
@@ -7,6 +7,7 @@ import pytest
7
8
from vllm import LLM, SamplingParams
9
from vllm.config import CompilationConfig, CompilationMode
10
+from vllm.platforms import current_platform
11
12
from ...utils import check_answers, fork_new_process_for_each_test, prep_prompts
13
@@ -43,15 +44,26 @@ def test_prompts():
44
return prompts
45
46
46
-@fork_new_process_for_each_test
47
+use_fork_for_test = (
48
+ fork_new_process_for_each_test if not current_platform.is_rocm() else lambda x: x
49
+)
50
+
51
+
52
+@use_fork_for_test
53
@pytest.mark.parametrize("kv_sharing_fast_prefill", [False, True])
54
@pytest.mark.parametrize("enforce_eager", [True, False])
55
def test_kv_sharing_fast_prefill(
56
monkeypatch: pytest.MonkeyPatch,
57
kv_sharing_fast_prefill: bool,
58
enforce_eager: bool,
53
- test_prompts: list[str],
59
):
60
+ if not enforce_eager and current_platform.is_rocm():
61
+ # Relevant context: https://github.com/vllm-project/vllm/pull/29244
62
+ pytest.skip(
63
+ "ROCm: torch.compile produces incorrect output for gemma-3n's GELU "
64
+ "with tanh approximation. Use enforce_eager=True instead."
65
+ )
66
+
67
sampling_params = SamplingParams(temperature=0.0, max_tokens=100)
68
compilation_config = CompilationConfig(
69
# This allows vLLM compilation backend to handle allocating and
@@ -65,7 +77,11 @@ def test_kv_sharing_fast_prefill(
77
78
with monkeypatch.context() as m:
79
# Make scheduling deterministic for reproducibility
68
- m.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
80
+ if current_platform.is_rocm():
81
+ # Use spawn to prevent cuda re-initialization error
82
+ m.setenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
83
+ else:
84
+ m.setenv("VLLM_ENABLE_V1_MULTIPROCESSING", "0")
85
86
prompts, answer, indices = prep_prompts(batch_size)
87