Skip to content

vllm.models.deepseek_v4.common.mm_preprocess

Multimodal preprocessing for the DeepSeek-V4 vision variants (DeepSeek-V4-Flash-Vision-Exp).

The image transform and sentinel-block construction are ported from the official repository's image_processor.py so that token counts bit-match the reference. Each <｜deepseek_image｜> placeholder in the prompt expands to a variable-length block of sentinel tokens; only positions with type == IMAGE receive vision embeddings, the other sentinels are looked up from learned embedding vectors in the model.

Unlike the reference (which uses out-of-vocab ids vocab_size + type), the sentinel block borrows five consecutive reserved tokenizer tokens (<|place_holder_mm_span_0431|> .. _0435|>): they are special tokens the tokenizer never emits from plain text, so the ids stay in-vocabulary and work with stock token validation, logprobs and detokenization. The ids are pure markers — every sentinel position's embedding is overwritten with vision/sentinel vectors before the decoder layers see it, exactly like the reference's out-of-vocab scheme.

Classes:

Functions:

DeepseekV4VLImageProcessor

Per-image transform (the PIL-input equivalent of the reference load_image).

Source code in vllm/models/deepseek_v4/common/mm_preprocess.py 
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
class DeepseekV4VLImageProcessor:
    """Per-image transform (the PIL-input equivalent of the reference
    ``load_image``)."""

    def __init__(self, config: DeepseekV4Config) -> None:
        super().__init__()
        self.patch_size = config.vision_patch_size
        self.downsample_ratio = config.vision_downsample_ratio
        self.max_n_token = config.vision_max_n_token
        self.min_pixels = config.vision_min_pixels
        self.max_wh_ratio = config.vision_max_wh_ratio

    def __call__(self, image: Image.Image):
        return load_image(
            image,
            patch_size=self.patch_size,
            downsample_ratio=self.downsample_ratio,
            max_n_token=self.max_n_token,
            min_pixels=self.min_pixels,
            max_wh_ratio=self.max_wh_ratio,
        )

DeepseekV4VLProcessor

Minimal stand-in for the HF processor of DeepSeek-V4 vision models.

The official repository ships image preprocessing as plain functions in image_processor.py (no auto_map processor), so this class wraps their ports directly and the model loads without --trust-remote-code.

__call__ returns a BatchFeature with one entry per image (flattened across images):

  • patches: (sum(n_vit_h * n_vit_w), 3, p, p) bf16 ViT patches.
  • vit_grid: (num_images, 2) int64 [n_vit_h, n_vit_w].
  • llm_grid: (num_images, 2) int64 [n_llm_h, n_llm_w].
  • perm: concatenated per-image (n_llm_h * n_llm_w,) int64 index selecting aligner outputs into the final N-layout order.
  • types: concatenated per-image pad-free sentinel block types; block ids = IMAGE_SENTINEL_BASE_ID + types.
Source code in vllm/models/deepseek_v4/common/mm_preprocess.py 
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
class DeepseekV4VLProcessor:
    """Minimal stand-in for the HF processor of DeepSeek-V4 vision models.

    The official repository ships image preprocessing as plain functions in
    ``image_processor.py`` (no ``auto_map`` processor), so this class wraps
    their ports directly and the model loads without ``--trust-remote-code``.

    ``__call__`` returns a ``BatchFeature`` with one entry per image
    (flattened across images):

    - ``patches``: ``(sum(n_vit_h * n_vit_w), 3, p, p)`` bf16 ViT patches.
    - ``vit_grid``: ``(num_images, 2)`` int64 ``[n_vit_h, n_vit_w]``.
    - ``llm_grid``: ``(num_images, 2)`` int64 ``[n_llm_h, n_llm_w]``.
    - ``perm``: concatenated per-image ``(n_llm_h * n_llm_w,)`` int64 index
      selecting aligner outputs into the final N-layout order.
    - ``types``: concatenated per-image pad-free sentinel block types;
      ``block ids = IMAGE_SENTINEL_BASE_ID + types``.
    """

    def __init__(self, config: DeepseekV4Config) -> None:
        super().__init__()
        self.config = config
        self.image_processor = DeepseekV4VLImageProcessor(config)

    def __call__(
        self,
        text: str | None = None,
        images: Sequence[Image.Image] | None = None,
        return_tensors: str | None = None,
        **kwargs: Any,
    ) -> BatchFeature:
        patches_list = []
        vit_grid = []
        llm_grid = []
        perm_list = []
        types_list = []
        for image in images or []:
            patches, n_vit_h, n_vit_w, n_llm_h, n_llm_w = self.image_processor(image)
            types, perm = build_image_block_pad_free(n_llm_h, n_llm_w)
            patches_list.append(patches)
            vit_grid.append((n_vit_h, n_vit_w))
            llm_grid.append((n_llm_h, n_llm_w))
            perm_list.append(perm)
            types_list.append(types)

        if not patches_list:
            return BatchFeature({})

        return BatchFeature(
            {
                "patches": torch.cat(patches_list),
                "vit_grid": torch.tensor(vit_grid, dtype=torch.int64),
                "llm_grid": torch.tensor(llm_grid, dtype=torch.int64),
                "perm": torch.cat(perm_list),
                "types": torch.cat(types_list),
            }
        )

build_image_block(n_llm_h, n_llm_w, start_pos)

Builds the N-layout token types (final order) and the aligner-row order for IMAGE slots.

Source code in vllm/models/deepseek_v4/common/mm_preprocess.py 
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
def build_image_block(n_llm_h: int, n_llm_w: int, start_pos: int):
    """Builds the N-layout token types (final order) and the aligner-row order
    for IMAGE slots."""
    compress_pad = COMPRESS_PAD_TO - 1 - start_pos % COMPRESS_PAD_TO
    pad_h = n_llm_h % 2
    rows = n_llm_h + pad_h
    row_len = n_llm_w + 1
    pad_last = rows // 2 * row_len % 2 * 2
    types = torch.tensor(
        ([IMAGE] * n_llm_w + [IMAGE_NEW_LINE]) * n_llm_h
        + [IMAGE_PAD] * (row_len * pad_h),
        dtype=torch.int64,
    )
    order = torch.arange(rows * row_len).view(rows // 2, 2, row_len)
    order = order.transpose(1, 2).reshape(-1)
    image_idx = torch.full((rows * row_len,), -1, dtype=torch.int64)
    image_idx.view(rows, row_len)[:n_llm_h, :n_llm_w] = torch.arange(
        n_llm_h * n_llm_w
    ).view(n_llm_h, n_llm_w)
    perm = image_idx[order]
    perm = perm[perm >= 0]
    types = torch.cat(
        [
            torch.full((compress_pad,), IMAGE_PAD, dtype=torch.int64),
            torch.tensor([IMAGE_START]),
            types[order],
            torch.full((pad_last,), IMAGE_PAD, dtype=torch.int64),
            torch.tensor([IMAGE_END]),
        ]
    )
    return types, perm

build_image_block_pad_free(n_llm_h, n_llm_w)

build_image_block without the position-dependent compressor pad.

start_pos is chosen so that compress_pad == 0; the pad is instead prepended when the block is spliced into the final prompt, where its position is known.

Source code in vllm/models/deepseek_v4/common/mm_preprocess.py 
230
231
232
233
234
235
236
237
def build_image_block_pad_free(n_llm_h: int, n_llm_w: int):
    """``build_image_block`` without the position-dependent compressor pad.

    ``start_pos`` is chosen so that ``compress_pad == 0``; the pad is instead
    prepended when the block is spliced into the final prompt, where its
    position is known.
    """
    return build_image_block(n_llm_h, n_llm_w, COMPRESS_PAD_TO - 1)

grid_tokens(best_height, best_width, patch_size, downsample_ratio)

Number of LLM tokens the aligner grid occupies (N-layout, including row/align padding).

Source code in vllm/models/deepseek_v4/common/mm_preprocess.py 
 92
 93
 94
 95
 96
 97
 98
 99
100
101
def grid_tokens(best_height, best_width, patch_size, downsample_ratio):
    """Number of LLM tokens the aligner grid occupies (N-layout, including
    row/align padding)."""
    n_llm_h = math.ceil((best_height // patch_size) / downsample_ratio)
    n_llm_w = math.ceil((best_width // patch_size) / downsample_ratio)
    num_tokens = n_llm_h * (n_llm_w + 1) + 2
    if n_llm_h % 2 == 1:
        num_tokens += n_llm_w + 1
    num_tokens += (n_llm_h + 1) // 2 * (n_llm_w + 1) % 2 * 2
    return n_llm_h, n_llm_w, num_tokens

image_sentinel_mask(token_ids)

Boolean mask for image-block sentinel positions (in-vocab ids).

Source code in vllm/models/deepseek_v4/common/mm_preprocess.py 
73
74
75
76
77
def image_sentinel_mask(token_ids: torch.Tensor) -> torch.Tensor:
    """Boolean mask for image-block sentinel positions (in-vocab ids)."""
    return (token_ids >= IMAGE_SENTINEL_BASE_ID) & (
        token_ids < IMAGE_SENTINEL_BASE_ID + len(IMAGE_SENTINEL_TOKEN_NAMES)
    )

load_image(image, *, patch_size, downsample_ratio, max_n_token, min_pixels, max_wh_ratio)

Transform one PIL image into ViT patches.

Same math as the reference load_image, except the image is already decoded (vLLM supplies PIL images instead of a record dict).

Source code in vllm/models/deepseek_v4/common/mm_preprocess.py 
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
def load_image(
    image: Image.Image,
    *,
    patch_size: int,
    downsample_ratio: int,
    max_n_token: int,
    min_pixels: int,
    max_wh_ratio: float | None,
):
    """Transform one PIL image into ViT patches.

    Same math as the reference ``load_image``, except the image is already
    decoded (vLLM supplies PIL images instead of a record dict).
    """
    p = patch_size
    image = image.convert("RGB")
    width, height = image.size
    if max_wh_ratio is not None and width > height * max_wh_ratio:
        width = height * max_wh_ratio
    if 0 < width * height < min_pixels:
        ratio = (min_pixels / (width * height)) ** 0.5
        width = int(width * ratio)
        height = int(height * ratio)
    best_width = math.ceil(width / p) * p
    best_height = math.ceil(height / p) * p
    n_llm_h, n_llm_w, best_height, best_width = safe_resize(
        height, width, best_height, best_width, p, downsample_ratio, max_n_token
    )
    n_vit_h, n_vit_w = best_height // p, best_width // p
    if max_wh_ratio is not None and image.width >= max_wh_ratio * image.height:
        image = image.resize((best_width, best_height))
    else:
        image = ImageOps.pad(image, (best_width, best_height), color=(127, 127, 127))
    x = torch.from_numpy(np.asarray(image, dtype=np.float32)).permute(2, 0, 1) / 255
    x = ((x - 0.5) / 0.5).to(torch.bfloat16)
    patches = (
        x.reshape(3, n_vit_h, p, n_vit_w, p)
        .permute(1, 3, 0, 2, 4)
        .reshape(n_vit_h * n_vit_w, 3, p, p)
    )
    return patches, n_vit_h, n_vit_w, n_llm_h, n_llm_w

validate_image_sentinel_ids(tokenizer)

Check the borrowed sentinel ids against the tokenizer.

Source code in vllm/models/deepseek_v4/common/mm_preprocess.py 
80
81
82
83
84
85
86
87
88
89
def validate_image_sentinel_ids(tokenizer) -> None:
    """Check the borrowed sentinel ids against the tokenizer."""
    for i, name in enumerate(IMAGE_SENTINEL_TOKEN_NAMES):
        token_id = tokenizer.convert_tokens_to_ids(name)
        if token_id != IMAGE_SENTINEL_BASE_ID + i:
            raise ValueError(
                f"Image sentinel token {name!r} has id {token_id}, expected "
                f"{IMAGE_SENTINEL_BASE_ID + i}; the DeepSeek-V4 vision "
                "sentinel block requires these consecutive reserved ids."
            )