| 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 |
| 5 | # not be triggered on CI and is primarily intended for local testing |
| 6 | # and verification. |
| 7 | |
| 8 | import vllm |
| 9 | from vllm.lora.request import LoRARequest |
| 10 | |
| 11 | from ..utils import multi_gpu_test |
| 12 | |
| 13 | MODEL_PATH = "deepseek-ai/DeepSeek-V2-Lite-Chat" |
| 14 | |
| 15 | PROMPT_TEMPLATE = "<｜begin▁of▁sentence｜>You are a helpful assistant.\n\nUser: {context}\n\nAssistant:" # noqa: E501 |
| 16 | |
| 17 | |
| 18 | def generate_and_test(llm: vllm.LLM, lora_path: str, lora_id: int): |
| 19 | prompts = [ |
| 20 | PROMPT_TEMPLATE.format(context="Who are you?"), |
| 21 | ] |
| 22 | sampling_params = vllm.SamplingParams(temperature=0, max_tokens=64) |
| 23 | outputs = llm.generate( |
| 24 | prompts, |
| 25 | sampling_params, |
| 26 | lora_request=LoRARequest(str(lora_id), lora_id, lora_path) if lora_id else None, |
| 27 | ) |
| 28 | # Print the outputs. |
| 29 | generated_texts: list[str] = [] |
| 30 | for output in outputs: |
| 31 | prompt = output.prompt |
| 32 | generated_text = output.outputs[0].text.strip() |
| 33 | generated_texts.append(generated_text) |
| 34 | print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}") |
| 35 | # return generated_texts |
| 36 | expected_lora_output = [ |
| 37 | "I am \u5f20\u5b50\u8c6a, an AI assistant developed by \u9648\u58eb\u680b.", # noqa: E501 |
| 38 | ] |
| 39 | for i in range(len(expected_lora_output)): |
| 40 | assert generated_texts[i].startswith(expected_lora_output[i]) |
| 41 | |
| 42 | |
| 43 | def test_deepseekv2_lora(deepseekv2_lora_files): |
| 44 | # We enable enforce_eager=True here to reduce VRAM usage for lora-test CI, |
| 45 | # Otherwise, the lora-test will fail due to CUDA OOM. |
| 46 | llm = vllm.LLM( |
| 47 | MODEL_PATH, |
| 48 | max_model_len=1024, |
| 49 | enable_lora=True, |
| 50 | max_loras=4, |
| 51 | enforce_eager=True, |
| 52 | trust_remote_code=True, |
| 53 | enable_chunked_prefill=True, |
| 54 | ) |
| 55 | generate_and_test(llm, deepseekv2_lora_files, 1) |
| 56 | |
| 57 | |
| 58 | def test_deepseekv2(deepseekv2_lora_files): |
| 59 | # We enable enforce_eager=True here to reduce VRAM usage for lora-test CI, |
| 60 | # Otherwise, the lora-test will fail due to CUDA OOM. |
| 61 | llm = vllm.LLM( |
| 62 | MODEL_PATH, |
| 63 | max_model_len=1024, |
| 64 | enable_lora=True, |
| 65 | max_loras=4, |
| 66 | enforce_eager=True, |
| 67 | trust_remote_code=True, |
| 68 | ) |
| 69 | generate_and_test(llm, deepseekv2_lora_files, 1) |
| 70 | |
| 71 | |
| 72 | @multi_gpu_test(num_gpus=2) |
| 73 | def test_deepseekv2_tp2(deepseekv2_lora_files): |
| 74 | # We enable enforce_eager=True here to reduce VRAM usage for lora-test CI, |
| 75 | # Otherwise, the lora-test will fail due to CUDA OOM. |
| 76 | llm = vllm.LLM( |
| 77 | MODEL_PATH, |
| 78 | max_model_len=1024, |
| 79 | enable_lora=True, |
| 80 | max_loras=4, |
| 81 | enforce_eager=True, |
| 82 | trust_remote_code=True, |
| 83 | tensor_parallel_size=2, |
| 84 | ) |
| 85 | generate_and_test(llm, deepseekv2_lora_files, 2) |
| 86 | |
| 87 | |
| 88 | @multi_gpu_test(num_gpus=4) |
| 89 | def test_deepseekv2_tp4(deepseekv2_lora_files): |
| 90 | # We enable enforce_eager=True here to reduce VRAM usage for lora-test CI, |
| 91 | # Otherwise, the lora-test will fail due to CUDA OOM. |
| 92 | llm = vllm.LLM( |
| 93 | MODEL_PATH, |
| 94 | max_model_len=1024, |
| 95 | enable_lora=True, |
| 96 | max_loras=4, |
| 97 | enforce_eager=True, |
| 98 | trust_remote_code=True, |
| 99 | tensor_parallel_size=4, |
| 100 | ) |
| 101 | generate_and_test(llm, deepseekv2_lora_files, 2) |