| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Tests for Mistral3's multimodal preprocessing kwargs.""" |
| 4 | |
| 5 | import pytest |
| 6 | import torch |
| 7 | import torch.nn.functional as F |
| 8 | from PIL import Image |
| 9 | from transformers import AutoProcessor, BatchFeature |
| 10 | from transformers.models.pixtral import PixtralProcessor |
| 11 | |
| 12 | from vllm.model_executor.layers.fusion.mm_input_norm import ( |
| 13 | FusedMMInputNorm, |
| 14 | IdentityInputNorm, |
| 15 | build_mm_input_norm, |
| 16 | ) |
| 17 | from vllm.model_executor.models.lightonocr import ( |
| 18 | LightOnOCRForConditionalGeneration, |
| 19 | LightOnOCRProcessingInfo, |
| 20 | ) |
| 21 | from vllm.model_executor.models.mistral3 import Mistral3HFEncoderInfo |
| 22 | from vllm.model_executor.models.pixtral import ( |
| 23 | PixtralHFEncoderInfo, |
| 24 | pixtral_patch_embed, |
| 25 | ) |
| 26 | from vllm.multimodal import MULTIMODAL_REGISTRY |
| 27 | from vllm.multimodal.inputs import MultiModalKwargsItems |
| 28 | from vllm.platforms import current_platform |
| 29 | from vllm.triton_utils import HAS_TRITON |
| 30 | |
| 31 | from ...utils import build_model_context |
| 32 | |
| 33 | # This repo ships both params.json (Mistral) and config.json (HF). Auto config |
| 34 | # selects PixtralForConditionalGeneration; force HF to exercise Mistral3. |
| 35 | _MODEL_CONFIG_KWARGS = {"config_format": "hf"} |
| 36 | _MODEL_ID = "mistralai/Mistral-Small-3.1-24B-Instruct-2503" |
| 37 | _LIGHTON_MODEL_ID = "lightonai/LightOnOCR-1B-1025" |
| 38 | |
| 39 | _RGB_MEAN = [0.48145466, 0.4578275, 0.40821073] |
| 40 | _RGB_STD = [0.26862954, 0.26130258, 0.27577711] |
| 41 | _RGB_RESCALE = 1.0 / 255.0 |
| 42 | |
| 43 | |
| 44 | @pytest.mark.usefixtures("default_vllm_config") |
| 45 | @pytest.mark.skipif( |
| 46 | not HAS_TRITON or not current_platform.is_cuda(), reason="CUDA Triton kernel" |
| 47 | ) |
| 48 | @pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16]) |
| 49 | @pytest.mark.parametrize("normalize", [False, True]) |
| 50 | @pytest.mark.parametrize(("channels", "patch_shape"), [(3, (4, 4)), (2, (2, 4))]) |
| 51 | def test_pixtral_patch_embed_preserves_image_order_and_strides( |
| 52 | dtype: torch.dtype, |
| 53 | normalize: bool, |
| 54 | channels: int, |
| 55 | patch_shape: tuple[int, int], |
| 56 | ): |
| 57 | """Packed projection matches CHW normalization and patch convolution.""" |
| 58 | device = torch.device(current_platform.device_type) |
| 59 | norm = ( |
| 60 | FusedMMInputNorm( |
| 61 | _RGB_MEAN[:channels], |
| 62 | _RGB_STD[:channels], |
| 63 | _RGB_RESCALE, |
| 64 | channel=channels, |
| 65 | ).to(device) |
| 66 | if normalize |
| 67 | else IdentityInputNorm() |
| 68 | ) |
| 69 | weight = torch.randn(32, channels, *patch_shape, device=device, dtype=dtype) |
| 70 | images = [ |
| 71 | torch.randint(0, 256, (channels, 17, 18), device=device, dtype=torch.uint8), |
| 72 | torch.randint(0, 256, (channels, 25, 17), device=device, dtype=torch.uint8)[ |
| 73 | :, 1:17, 1:17 |
| 74 | ], |
| 75 | torch.randint(0, 256, (channels, 19, 20), device=device, dtype=torch.uint8)[ |
| 76 | :, 1:17, 2:18 |
| 77 | ], |
| 78 | ] |
| 79 | reference = [] |
| 80 | for image in images: |
| 81 | if normalize: |
| 82 | normalized = ( |
| 83 | image.float() * norm.weight[:, None, None] + norm.bias[:, None, None] |
| 84 | ) |
| 85 | else: |
| 86 | normalized = image |
| 87 | reference.append( |
| 88 | F.conv2d(normalized.to(dtype).unsqueeze(0), weight, stride=patch_shape) |
| 89 | ) |
| 90 | packed, actual = pixtral_patch_embed(images, weight, norm) |
| 91 | for expected, result in zip(reference, actual): |
| 92 | torch.testing.assert_close(result, expected, rtol=0.01, atol=0.02) |
| 93 | expected_packed = torch.cat( |
| 94 | [result.flatten(2).transpose(1, 2) for result in reference], dim=1 |
| 95 | ) |
| 96 | torch.testing.assert_close(packed, expected_packed, rtol=0.01, atol=0.02) |
| 97 | |
| 98 | |
| 99 | def _process_images_with_hf( |
| 100 | hf_processor, |
| 101 | images: list[Image.Image], |
| 102 | mm_processor_kwargs: dict[str, object], |
| 103 | ) -> tuple[list[torch.Tensor], list[int]]: |
| 104 | """Process images and placeholders through the public HF processor.""" |
| 105 | hf_out = hf_processor( |
| 106 | text=hf_processor.image_token * len(images), |
| 107 | images=images, |
| 108 | return_tensors="pt", |
| 109 | **mm_processor_kwargs, |
| 110 | ) |
| 111 | pixel_values = hf_out["pixel_values"] |
| 112 | image_sizes = hf_out["image_sizes"] |
| 113 | unpadded = [p[:, :h, :w] for p, (h, w) in zip(pixel_values, image_sizes)] |
| 114 | |
| 115 | special_token_ids = { |
| 116 | hf_processor.image_token_id, |
| 117 | hf_processor.image_break_token_id, |
| 118 | hf_processor.image_end_token_id, |
| 119 | } |
| 120 | placeholder_tokens = [ |
| 121 | token_id |
| 122 | for token_id in hf_out["input_ids"][0].tolist() |
| 123 | if token_id in special_token_ids |
| 124 | ] |
| 125 | return unpadded, placeholder_tokens |
| 126 | |
| 127 | |
| 128 | def _placeholder_tokens_from_prompt_updates( |
| 129 | processor, |
| 130 | images: list[Image.Image], |
| 131 | pixel_values: list[torch.Tensor], |
| 132 | mm_processor_kwargs: dict[str, object], |
| 133 | ) -> list[int]: |
| 134 | hf_inputs = BatchFeature({"pixel_values": pixel_values}) |
| 135 | fields_config = processor._get_mm_fields_config(hf_inputs, mm_processor_kwargs) |
| 136 | out_mm_kwargs = MultiModalKwargsItems.from_hf_inputs(hf_inputs, fields_config) |
| 137 | # Prompt updates use raw PIL sizes and must predict the processed grid. |
| 138 | mm_items = processor.info.parse_mm_data({"image": images}) |
| 139 | updates = processor._get_prompt_updates( |
| 140 | mm_items, mm_processor_kwargs, out_mm_kwargs |
| 141 | ) |
| 142 | placeholder_tokens: list[int] = [] |
| 143 | for item_idx in range(len(images)): |
| 144 | details = updates[0].resolve(item_idx).content |
| 145 | placeholder_tokens.extend(details.full) |
| 146 | return placeholder_tokens |
| 147 | |
| 148 | |
| 149 | def _expected_placeholder_tokens_per_image( |
| 150 | hf_processor, |
| 151 | pixel_values: torch.Tensor, |
| 152 | ) -> int: |
| 153 | """Count projected tokens from the actual HF-processed H×W.""" |
| 154 | image_h, image_w = pixel_values.shape[-2:] |
| 155 | patch_size = hf_processor.image_processor.patch_size |
| 156 | if isinstance(patch_size, dict): |
| 157 | patch_h = patch_size["height"] |
| 158 | patch_w = patch_size["width"] |
| 159 | else: |
| 160 | patch_h = patch_w = int(patch_size) |
| 161 | |
| 162 | spatial_merge_size = getattr(hf_processor, "spatial_merge_size", 1) |
| 163 | merged_patch_h = patch_h * spatial_merge_size |
| 164 | merged_patch_w = patch_w * spatial_merge_size |
| 165 | assert image_h % merged_patch_h == 0 |
| 166 | assert image_w % merged_patch_w == 0 |
| 167 | |
| 168 | return (image_h // merged_patch_h) * (image_w // merged_patch_w) |
| 169 | |
| 170 | |
| 171 | @pytest.mark.parametrize("model_id", [_MODEL_ID]) |
| 172 | @pytest.mark.parametrize( |
| 173 | ("mm_processor_kwargs", "image_size", "expected_toks_per_img"), |
| 174 | [ |
| 175 | ({}, (448, 448), 256), |
| 176 | ({"size": {"longest_edge": 1008}}, (1540, 1540), 1296), |
| 177 | ( |
| 178 | {"images_kwargs": {"size": {"longest_edge": 1008}}}, |
| 179 | (1540, 1540), |
| 180 | 1296, |
| 181 | ), |
| 182 | ({"size": {"longest_edge": 1288}}, (1536, 1187), 1656), |
| 183 | ({"size": {"longest_edge": 1008}}, (29, 29), 4), |
| 184 | ({"size": {"longest_edge": 1000}}, (1540, 1700), 1188), |
| 185 | ], |
| 186 | ) |
| 187 | @pytest.mark.parametrize("num_imgs", [1, 2]) |
| 188 | @pytest.mark.parametrize("kwargs_on_init", [True, False]) |
| 189 | def test_processor_size_override( |
| 190 | model_id: str, |
| 191 | mm_processor_kwargs: dict[str, object], |
| 192 | image_size: tuple[int, int], |
| 193 | expected_toks_per_img: int, |
| 194 | num_imgs: int, |
| 195 | kwargs_on_init: bool, |
| 196 | ): |
| 197 | ctx = build_model_context( |
| 198 | model_id, |
| 199 | mm_processor_kwargs=mm_processor_kwargs if kwargs_on_init else None, |
| 200 | limit_mm_per_prompt={"image": num_imgs}, |
| 201 | model_config_kwargs=_MODEL_CONFIG_KWARGS, |
| 202 | ) |
| 203 | processor = MULTIMODAL_REGISTRY.create_processor(ctx.model_config) |
| 204 | hf_processor_mm_kwargs = {} if kwargs_on_init else mm_processor_kwargs |
| 205 | vllm_hf_processor = processor.info.get_hf_processor(**hf_processor_mm_kwargs) |
| 206 | assert isinstance(vllm_hf_processor, PixtralProcessor) |
| 207 | |
| 208 | hf_processor = AutoProcessor.from_pretrained( |
| 209 | model_id, |
| 210 | fix_mistral_regex=True, |
| 211 | ) |
| 212 | |
| 213 | dummy_image = Image.new("RGB", image_size, color=(127, 127, 127)) |
| 214 | images = [dummy_image] * num_imgs |
| 215 | merged_mm_kwargs = processor.info.ctx.get_merged_mm_kwargs(hf_processor_mm_kwargs) |
| 216 | pixel_values, hf_placeholder_tokens = _process_images_with_hf( |
| 217 | hf_processor, images, merged_mm_kwargs |
| 218 | ) |
| 219 | |
| 220 | prompt_update_tokens = _placeholder_tokens_from_prompt_updates( |
| 221 | processor, images, pixel_values, hf_processor_mm_kwargs |
| 222 | ) |
| 223 | expected_from_pixel_values = _expected_placeholder_tokens_per_image( |
| 224 | hf_processor, pixel_values[0] |
| 225 | ) |
| 226 | assert expected_from_pixel_values == expected_toks_per_img |
| 227 | assert hf_placeholder_tokens.count(hf_processor.image_token_id) == ( |
| 228 | expected_from_pixel_values * num_imgs |
| 229 | ) |
| 230 | assert prompt_update_tokens == hf_placeholder_tokens |
| 231 | |
| 232 | |
| 233 | @pytest.mark.usefixtures("default_vllm_config") |
| 234 | def test_mm_device_do_normalize(): |
| 235 | ctx = build_model_context( |
| 236 | _MODEL_ID, |
| 237 | limit_mm_per_prompt={"image": 2}, |
| 238 | model_config_kwargs=_MODEL_CONFIG_KWARGS, |
| 239 | ) |
| 240 | assert ctx.model_config.multimodal_config.mm_device_do_normalize |
| 241 | |
| 242 | ctx.model_config.multimodal_config.mm_device_do_normalize = False |
| 243 | processor = MULTIMODAL_REGISTRY.create_processor(ctx.model_config) |
| 244 | images = [ |
| 245 | Image.new("RGB", (31, 47), color=(17, 89, 231)), |
| 246 | Image.new("RGB", (48, 32), color=(201, 13, 127)), |
| 247 | ] |
| 248 | prompt = [processor.info.get_hf_config().image_token_index] * len(images) |
| 249 | mm_items = processor.info.parse_mm_data({"image": images}) |
| 250 | |
| 251 | normalized_inputs = processor(prompt, mm_items=mm_items) |
| 252 | normalized_values = normalized_inputs["mm_kwargs"].get_data()["pixel_values"] |
| 253 | |
| 254 | ctx.model_config.multimodal_config.mm_device_do_normalize = True |
| 255 | raw_inputs = processor(prompt, mm_items=mm_items) |
| 256 | raw_values = raw_inputs["mm_kwargs"].get_data()["pixel_values"] |
| 257 | assert all(value.dtype == torch.uint8 for value in raw_values) |
| 258 | |
| 259 | input_norm = build_mm_input_norm(ctx.model_config) |
| 260 | patch_size = processor.info.get_hf_config().vision_config.patch_size |
| 261 | |
| 262 | def pack_patches(image: torch.Tensor) -> torch.Tensor: |
| 263 | channels, height, width = image.shape |
| 264 | rows = height // patch_size |
| 265 | cols = width // patch_size |
| 266 | return ( |
| 267 | image[:, : rows * patch_size, : cols * patch_size] |
| 268 | .reshape(channels, rows, patch_size, cols, patch_size) |
| 269 | .permute(1, 3, 0, 2, 4) |
| 270 | .reshape(-1, channels * patch_size**2) |
| 271 | ) |
| 272 | |
| 273 | for raw, normalized in zip(raw_values, normalized_values): |
| 274 | output = input_norm(pack_patches(raw), normalized.dtype) |
| 275 | torch.testing.assert_close( |
| 276 | output, pack_patches(normalized), rtol=1e-5, atol=1e-6 |
| 277 | ) |
| 278 | |
| 279 | |
| 280 | def test_scoped_request_size_overrides_configured_flat_size(): |
| 281 | ctx = build_model_context( |
| 282 | _MODEL_ID, |
| 283 | mm_processor_kwargs={"size": {"longest_edge": 1008}}, |
| 284 | limit_mm_per_prompt={"image": 1}, |
| 285 | model_config_kwargs=_MODEL_CONFIG_KWARGS, |
| 286 | ) |
| 287 | processor = MULTIMODAL_REGISTRY.create_processor(ctx.model_config) |
| 288 | hf_processor = AutoProcessor.from_pretrained( |
| 289 | _MODEL_ID, |
| 290 | fix_mistral_regex=True, |
| 291 | ) |
| 292 | |
| 293 | hf_processor_mm_kwargs: dict[str, object] = { |
| 294 | "images_kwargs": {"size": {"longest_edge": 1288}} |
| 295 | } |
| 296 | image = Image.new("RGB", (1536, 1187), color=(127, 127, 127)) |
| 297 | |
| 298 | merged_mm_kwargs = processor.info.ctx.get_merged_mm_kwargs(hf_processor_mm_kwargs) |
| 299 | pixel_values, hf_placeholder_tokens = _process_images_with_hf( |
| 300 | hf_processor, [image], merged_mm_kwargs |
| 301 | ) |
| 302 | prompt_update_tokens = _placeholder_tokens_from_prompt_updates( |
| 303 | processor, [image], pixel_values, hf_processor_mm_kwargs |
| 304 | ) |
| 305 | |
| 306 | expected_from_pixel_values = _expected_placeholder_tokens_per_image( |
| 307 | hf_processor, pixel_values[0] |
| 308 | ) |
| 309 | assert expected_from_pixel_values == 1656 |
| 310 | assert hf_placeholder_tokens.count(hf_processor.image_token_id) == 1656 |
| 311 | assert prompt_update_tokens == hf_placeholder_tokens |
| 312 | |
| 313 | |
| 314 | def test_lightonocr_keeps_vision_config_image_size(): |
| 315 | assert not LightOnOCRForConditionalGeneration.supports_mm_device_do_normalize |
| 316 | |
| 317 | ctx = build_model_context( |
| 318 | _LIGHTON_MODEL_ID, |
| 319 | mm_processor_kwargs={"size": {"longest_edge": 1008}}, |
| 320 | model_config_kwargs=_MODEL_CONFIG_KWARGS, |
| 321 | ) |
| 322 | processor = MULTIMODAL_REGISTRY.create_processor(ctx.model_config) |
| 323 | |
| 324 | assert isinstance(processor.info, LightOnOCRProcessingInfo) |
| 325 | encoder_info = processor.info.get_vision_encoder_info() |
| 326 | assert isinstance(encoder_info, PixtralHFEncoderInfo) |
| 327 | assert not isinstance(encoder_info, Mistral3HFEncoderInfo) |
| 328 | assert encoder_info.get_image_size() == ( |
| 329 | processor.info.get_hf_config().vision_config.image_size |
| 330 | ) |