@tduncan / Nvidia-Optimization-Stack-VLLM
Code Commits Issues Pulls Cloud Agent CI/CD
main
main
py 106 lines 3.08 KB
Raw
1 # SPDX-License-Identifier: Apache-2.0
2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
4 # NOTE To avoid overloading the CI pipeline, this test script will not
5 # be triggered on CI and is primarily intended for local testing and verification.
6
7 import vllm
8 from vllm.assets.image import ImageAsset
9 from vllm.lora.request import LoRARequest
10
11 from ..utils import multi_gpu_test
12
13 MODEL_PATH = "google/gemma-4-E2B-it"
14
15 PROMPT_TEMPLATE = """<bos><|turn>user
16 <|image|>What is in the image?<turn|>
17 <|turn>model
18 """
19
20 TEST_IMAGES = [
21 ImageAsset("stop_sign"),
22 ImageAsset("cherry_blossom"),
23 ]
24
25 EXPECTED_OUTPUTS_VISION = [
26 "A red stop sign stands prominently in the foreground.",
27 "A majestic skyscraper stands tall, partially obscured by a vibrant "
28 "canopy of cherry blossoms, against a clear blue sky.",
29 ]
30
31
32 def generate_and_test(llm: vllm.LLM, lora_path: str, lora_id: int) -> None:
33 prompts = [
34 {
35 "prompt": PROMPT_TEMPLATE,
36 "multi_modal_data": {"image": asset.pil_image},
37 }
38 for asset in TEST_IMAGES
39 ]
40 sampling_params = vllm.SamplingParams(temperature=0, max_tokens=128)
41 outputs = llm.generate(
42 prompts,
43 sampling_params,
44 lora_request=LoRARequest(str(lora_id), lora_id, lora_path),
45 )
46
47 generated_texts = [output.outputs[0].text.strip() for output in outputs]
48 for generated, expected in zip(generated_texts, EXPECTED_OUTPUTS_VISION):
49 assert generated.startswith(expected), (
50 f"Generated text {generated!r} does not match expected {expected!r}"
51 )
52
53
54 def test_gemma4_lora(gemma4_vision_lora_files):
55 llm = vllm.LLM(
56 MODEL_PATH,
57 max_model_len=1024,
58 enable_lora=True,
59 enforce_eager=True,
60 max_loras=4,
61 trust_remote_code=True,
62 limit_mm_per_prompt={"image": 1},
63 mm_processor_cache_gb=0,
64 enable_tower_connector_lora=True,
65 )
66
67 generate_and_test(llm, gemma4_vision_lora_files, lora_id=1)
68 generate_and_test(llm, gemma4_vision_lora_files, lora_id=2)
69
70
71 @multi_gpu_test(num_gpus=2)
72 def test_gemma4_lora_tp2(gemma4_vision_lora_files):
73 llm = vllm.LLM(
74 MODEL_PATH,
75 max_model_len=1024,
76 enforce_eager=True,
77 enable_lora=True,
78 max_loras=4,
79 trust_remote_code=True,
80 tensor_parallel_size=2,
81 limit_mm_per_prompt={"image": 1},
82 mm_processor_cache_gb=0,
83 enable_tower_connector_lora=True,
84 )
85
86 generate_and_test(llm, gemma4_vision_lora_files, lora_id=1)
87 generate_and_test(llm, gemma4_vision_lora_files, lora_id=2)
88
89
90 @multi_gpu_test(num_gpus=4)
91 def test_gemma4_lora_tp4(gemma4_vision_lora_files):
92 llm = vllm.LLM(
93 MODEL_PATH,
94 max_model_len=1024,
95 enforce_eager=True,
96 enable_lora=True,
97 max_loras=4,
98 trust_remote_code=True,
99 tensor_parallel_size=4,
100 limit_mm_per_prompt={"image": 1},
101 mm_processor_cache_gb=0,
102 enable_tower_connector_lora=True,
103 )
104
105 generate_and_test(llm, gemma4_vision_lora_files, lora_id=1)
106 generate_and_test(llm, gemma4_vision_lora_files, lora_id=2)