| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | |
| 5 | import pytest |
| 6 | import torch |
| 7 | |
| 8 | from vllm.multimodal.image import rescale_image_size |
| 9 | |
| 10 | from ...conftest import IMAGE_ASSETS, ImageTestAssets, VllmRunner |
| 11 | from ..utils import check_logprobs_close |
| 12 | |
| 13 | HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts( |
| 14 | { |
| 15 | "stop_sign": "<|im_start|>User\n<image>\nWhat's the content in the center of the image?<|im_end|>\n<|im_start|>Assistant\n", # noqa: E501 |
| 16 | "cherry_blossom": "<|im_start|>User\n<image>\nWhat is the season?<|im_end|>\n<|im_start|>Assistant\n", # noqa: E501 |
| 17 | } |
| 18 | ) |
| 19 | |
| 20 | IMAGE_SIZE_FACTOR_GROUPS = ( |
| 21 | # Single-scale |
| 22 | (1.0,), |
| 23 | # Single-scale, batched |
| 24 | (1.0, 1.0, 1.0), |
| 25 | # Multi-scale |
| 26 | (0.25, 0.5, 1.0), |
| 27 | ) |
| 28 | |
| 29 | |
| 30 | def run_awq_test( |
| 31 | vllm_runner: type[VllmRunner], |
| 32 | image_assets: ImageTestAssets, |
| 33 | source_model: str, |
| 34 | quant_model: str, |
| 35 | *, |
| 36 | size_factor_groups: tuple[tuple[float, ...], ...], |
| 37 | dtype: str, |
| 38 | max_tokens: int, |
| 39 | num_logprobs: int, |
| 40 | tensor_parallel_size: int, |
| 41 | distributed_executor_backend: str | None = None, |
| 42 | ): |
| 43 | images = [asset.pil_image for asset in image_assets] |
| 44 | |
| 45 | inputs_per_image_and_size_group = [ |
| 46 | ( |
| 47 | [prompt for _ in size_factors], |
| 48 | [rescale_image_size(image, factor) for factor in size_factors], |
| 49 | ) |
| 50 | for size_factors in size_factor_groups |
| 51 | for image, prompt in zip(images, HF_IMAGE_PROMPTS) |
| 52 | ] |
| 53 | |
| 54 | # NOTE: take care of the order. run vLLM first, and then run HF. |
| 55 | # vLLM needs a fresh new process without cuda initialization. |
| 56 | # if we run HF first, the cuda initialization will be done and it |
| 57 | # will hurt multiprocessing backend with fork method (the default method). |
| 58 | |
| 59 | # max_model_len should be greater than image_feature_size |
| 60 | with vllm_runner( |
| 61 | source_model, |
| 62 | max_model_len=4096, |
| 63 | dtype=dtype, |
| 64 | tensor_parallel_size=tensor_parallel_size, |
| 65 | distributed_executor_backend=distributed_executor_backend, |
| 66 | enforce_eager=True, |
| 67 | default_torch_num_threads=1, |
| 68 | ) as vllm_model: |
| 69 | source_outputs_per_image = [ |
| 70 | vllm_model.generate_greedy_logprobs( |
| 71 | prompts, max_tokens, num_logprobs=num_logprobs, images=images |
| 72 | ) |
| 73 | for prompts, images in inputs_per_image_and_size_group |
| 74 | ] |
| 75 | |
| 76 | with vllm_runner( |
| 77 | quant_model, |
| 78 | quantization="awq", |
| 79 | max_model_len=4096, |
| 80 | dtype=dtype, |
| 81 | tensor_parallel_size=tensor_parallel_size, |
| 82 | distributed_executor_backend=distributed_executor_backend, |
| 83 | enforce_eager=True, |
| 84 | default_torch_num_threads=1, |
| 85 | ) as vllm_model: |
| 86 | quant_outputs_per_image = [ |
| 87 | vllm_model.generate_greedy_logprobs( |
| 88 | prompts, max_tokens, num_logprobs=num_logprobs, images=images |
| 89 | ) |
| 90 | for prompts, images in inputs_per_image_and_size_group |
| 91 | ] |
| 92 | |
| 93 | for source_outputs, quant_outputs in zip( |
| 94 | source_outputs_per_image, quant_outputs_per_image |
| 95 | ): |
| 96 | # TODO: Check whether using original CLIPVisionModel can improve |
| 97 | # consistency against HF |
| 98 | check_logprobs_close( |
| 99 | outputs_0_lst=source_outputs, |
| 100 | outputs_1_lst=quant_outputs, |
| 101 | name_0="source", |
| 102 | name_1="awq", |
| 103 | ) |
| 104 | |
| 105 | |
| 106 | @pytest.mark.parametrize( |
| 107 | ("model", "quantization", "dtype"), |
| 108 | [ |
| 109 | ("mattbucci/gemma-4-26B-AWQ", "awq", "float16"), |
| 110 | ("cyankiwi/gemma-4-26B-A4B-it-AWQ-4bit", "compressed-tensors", "bfloat16"), |
| 111 | ], |
| 112 | ids=[ |
| 113 | "gemma4-moe-standard-awq-dot-suffix", |
| 114 | "gemma4-moe-compressed-tensors-underscore-suffix", |
| 115 | ], |
| 116 | ) |
| 117 | @torch.inference_mode() |
| 118 | def test_awq_load( |
| 119 | vllm_runner: type[VllmRunner], |
| 120 | example_prompts: list[str], |
| 121 | model: str, |
| 122 | quantization: str, |
| 123 | dtype: str, |
| 124 | ) -> None: |
| 125 | """Regression test: AWQ weight loading must not KeyError.""" |
| 126 | with vllm_runner( |
| 127 | model, |
| 128 | quantization=quantization, |
| 129 | dtype=dtype, |
| 130 | max_model_len=128, |
| 131 | enforce_eager=True, |
| 132 | ) as vllm_model: |
| 133 | outputs = vllm_model.generate_greedy(example_prompts[:2], max_tokens=32) |
| 134 | assert len(outputs) == 2 |
| 135 | |
| 136 | |
| 137 | @pytest.mark.parametrize( |
| 138 | ("source_model", "quant_model"), |
| 139 | [("OpenGVLab/InternVL2-2B", "OpenGVLab/InternVL2-2B-AWQ")], |
| 140 | ) |
| 141 | @pytest.mark.parametrize("dtype", ["half"]) |
| 142 | @pytest.mark.parametrize("max_tokens", [128]) |
| 143 | @pytest.mark.parametrize("num_logprobs", [5]) |
| 144 | @torch.inference_mode() |
| 145 | def test_awq_models( |
| 146 | vllm_runner, |
| 147 | image_assets, |
| 148 | source_model, |
| 149 | quant_model, |
| 150 | dtype, |
| 151 | max_tokens, |
| 152 | num_logprobs, |
| 153 | ) -> None: |
| 154 | run_awq_test( |
| 155 | vllm_runner, |
| 156 | image_assets, |
| 157 | source_model, |
| 158 | quant_model, |
| 159 | size_factor_groups=IMAGE_SIZE_FACTOR_GROUPS, |
| 160 | dtype=dtype, |
| 161 | max_tokens=max_tokens, |
| 162 | num_logprobs=num_logprobs, |
| 163 | tensor_parallel_size=1, |
| 164 | ) |