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 13, 15:45 UTC
Add runai model streamer e2e test for GCS (#28079)
Signed-off-by: Alexis MacAskill <amacaskill@google.com>
Alexis MacAskill committed
Nov 6, 2025 at 19:07 UTC
a47d94f18cf0c4d919dee5ac7ee399880190a41d
2 files changed
+20
.buildkite/test-pipeline.yaml
+3
@@ -546,8 +546,11 @@ steps:
546
547
- label: Model Executor Test # 23min
548
timeout_in_minutes: 35
549
+ torch_nightly: true
550
mirror_hardwares: [amdexperimental]
551
source_file_dependencies:
552
+ - vllm/engine/arg_utils.py
553
+ - vllm/config/model.py
554
- vllm/model_executor
555
- tests/model_executor
556
- tests/entrypoints/openai/test_tensorizer_entrypoint.py
tests/model_executor/model_loader/runai_model_streamer/test_runai_model_streamer_loader.py
+17
@@ -1,12 +1,16 @@
1
# SPDX-License-Identifier: Apache-2.0
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4
+import pytest
5
+
6
from vllm import SamplingParams
7
from vllm.config.load import LoadConfig
8
from vllm.model_executor.model_loader import get_model_loader
9
10
load_format = "runai_streamer"
11
test_model = "openai-community/gpt2"
12
+# TODO(amacaskill): Replace with a GKE owned GCS bucket.
13
+test_gcs_model = "gs://vertex-model-garden-public-us/codegemma/codegemma-2b/"
14
15
prompts = [
16
"Hello, my name is",
@@ -32,3 +36,16 @@ def test_runai_model_loader_download_files(vllm_runner):
36
with vllm_runner(test_model, load_format=load_format) as llm:
37
deserialized_outputs = llm.generate(prompts, sampling_params)
38
assert deserialized_outputs
39
+
40
+
41
+def test_runai_model_loader_download_files_gcs(
42
+ vllm_runner, monkeypatch: pytest.MonkeyPatch
43
+):
44
+ monkeypatch.setenv("GOOGLE_CLOUD_PROJECT", "fake-project")
45
+ monkeypatch.setenv("RUNAI_STREAMER_GCS_USE_ANONYMOUS_CREDENTIALS", "true")
46
+ monkeypatch.setenv(
47
+ "CLOUD_STORAGE_EMULATOR_ENDPOINT", "https://storage.googleapis.com"
48
+ )
49
+ with vllm_runner(test_gcs_model, load_format=load_format) as llm:
50
+ deserialized_outputs = llm.generate(prompts, sampling_params)
51
+ assert deserialized_outputs