| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """This script contains: |
| 4 | 1. test lora with speculative decoding for batch inference |
| 5 | """ |
| 6 | |
| 7 | import pytest |
| 8 | import torch |
| 9 | |
| 10 | import vllm.envs as envs |
| 11 | from vllm import SamplingParams |
| 12 | from vllm.config import CompilationConfig |
| 13 | from vllm.lora.request import LoRARequest |
| 14 | from vllm.platforms import current_platform |
| 15 | from vllm.utils.torch_utils import set_random_seed |
| 16 | |
| 17 | from ..utils import assert_request_outputs_match |
| 18 | |
| 19 | LORA_TEST_PROMPT_MAP: dict[str, str] = {} |
| 20 | |
| 21 | LORA_TEST_PROMPT_MAP["premjatin/qwen-linear-algebra-coder"] = """ |
| 22 | ### INSTRUCTION: |
| 23 | You are an AI assistant that generates Python code to solve linear |
| 24 | algebra problems. |
| 25 | |
| 26 | ### PROBLEM: |
| 27 | Find the eigenvalues and eigenvectors of the following 3x3 matrix: |
| 28 | [[3, 2, 0], |
| 29 | [2, 3, 0], |
| 30 | [0, 0, 2]] |
| 31 | |
| 32 | ### OUTPUT FORMAT (STRICT): |
| 33 | Numbers should be represented as integers only. |
| 34 | |
| 35 | ### PYTHON SOLUTION: |
| 36 | """ |
| 37 | |
| 38 | SEED = 42 |
| 39 | |
| 40 | |
| 41 | @pytest.mark.parametrize( |
| 42 | "model_setup", |
| 43 | [ |
| 44 | ( |
| 45 | "eagle3", |
| 46 | "Qwen/Qwen3-1.7B", |
| 47 | "AngelSlim/Qwen3-1.7B_eagle3", |
| 48 | "premjatin/qwen-linear-algebra-coder", |
| 49 | 1, |
| 50 | ) |
| 51 | ], |
| 52 | ) |
| 53 | @pytest.mark.skipif( |
| 54 | not current_platform.is_cuda_alike(), reason="Requires CUDA or ROCm" |
| 55 | ) |
| 56 | def test_batch_inference_correctness( |
| 57 | monkeypatch: pytest.MonkeyPatch, |
| 58 | model_setup: tuple[str, str, str, str, int], |
| 59 | vllm_runner, |
| 60 | ): |
| 61 | """Compare the outputs of a LLM with only Lora and a LLM with both SD and Lora. |
| 62 | Should be the same and no failure when doing batch inference. |
| 63 | model_setup: (method, model_name, spec_model_name, lora_path, tp_size) |
| 64 | """ |
| 65 | with monkeypatch.context() as m: |
| 66 | m.setenv("VLLM_BATCH_INVARIANT", "1") |
| 67 | m.setattr(envs, "VLLM_BATCH_INVARIANT", True) |
| 68 | set_random_seed(SEED) |
| 69 | m.setattr(torch.backends.cudnn, "benchmark", False) |
| 70 | m.setattr(torch.backends.cudnn, "deterministic", True) |
| 71 | |
| 72 | method, model_name, spec_model_name, lora_path, tp_size = model_setup |
| 73 | |
| 74 | prompts = [LORA_TEST_PROMPT_MAP[lora_path]] * 100 |
| 75 | lora_request = LoRARequest("adapter", 1, lora_path) |
| 76 | sampling_params = SamplingParams( |
| 77 | temperature=0.0, top_p=1.0, top_k=-1, seed=SEED, max_tokens=128 |
| 78 | ) |
| 79 | |
| 80 | # without speculative decoding |
| 81 | with vllm_runner( |
| 82 | model_name, |
| 83 | block_size=None, |
| 84 | trust_remote_code=True, |
| 85 | tensor_parallel_size=tp_size, |
| 86 | max_model_len=2048, |
| 87 | max_num_seqs=4, |
| 88 | enable_lora=True, |
| 89 | max_loras=1, |
| 90 | max_cpu_loras=1, |
| 91 | max_lora_rank=16, |
| 92 | enable_chunked_prefill=None, |
| 93 | compilation_config=CompilationConfig(), |
| 94 | ) as ref_runner: |
| 95 | ref_outputs = ref_runner.llm.generate( |
| 96 | prompts, sampling_params, lora_request=lora_request |
| 97 | ) |
| 98 | |
| 99 | with vllm_runner( |
| 100 | model_name, |
| 101 | block_size=None, |
| 102 | trust_remote_code=True, |
| 103 | tensor_parallel_size=tp_size, |
| 104 | speculative_config={ |
| 105 | "method": method, |
| 106 | "model": spec_model_name, |
| 107 | "num_speculative_tokens": 3, |
| 108 | "max_model_len": 2048, |
| 109 | }, |
| 110 | max_model_len=2048, |
| 111 | max_num_seqs=4, |
| 112 | enable_lora=True, |
| 113 | max_loras=1, |
| 114 | max_cpu_loras=1, |
| 115 | max_lora_rank=16, |
| 116 | enable_chunked_prefill=None, |
| 117 | compilation_config=CompilationConfig(), |
| 118 | ) as spec_runner: |
| 119 | lora_spec_outputs = spec_runner.llm.generate( |
| 120 | prompts, sampling_params, lora_request=lora_request |
| 121 | ) |
| 122 | |
| 123 | # Under greedy verification, the spec-decode output should equal the |
| 124 | # non-spec output (modulo FP noise from the target's verify-path matmul |
| 125 | # running at seqlen num_speculative_tokens+1 vs 1). 90% leaves slack. |
| 126 | assert_request_outputs_match( |
| 127 | ref_outputs, |
| 128 | lora_spec_outputs, |
| 129 | required_matches=int(0.90 * len(ref_outputs)) + 1, |
| 130 | context=( |
| 131 | f"LoRA + {method}: target={model_name}, draft={spec_model_name}, " |
| 132 | f"adapter={lora_path}" |
| 133 | ), |
| 134 | ) |