vllm.models.kimi_k3.common.mm_preprocess
¶
Shared Kimi-K3 multimodal preprocessing.
Classes:
-
KimiK3DummyInputsBuilder–Builds image-based dummy inputs for K3 profiling.
-
KimiK3MultiModalProcessor–Image-only multi-modal processor for Kimi-K3.
-
KimiK3ProcessingInfo–Processing information for the image-only Kimi-K3 model.
KimiK3DummyInputsBuilder
¶
Bases: BaseDummyInputsBuilder[KimiK3ProcessingInfo]
Builds image-based dummy inputs for K3 profiling.
The dummy text is made of <|kimi_image_placeholder|> tokens — exactly
the placeholder that K3's _get_prompt_updates expands — and the dummy
mm data is a plain list of PIL images under the image key.
Source code in vllm/models/kimi_k3/common/mm_preprocess.py
KimiK3MultiModalProcessor
¶
Bases: BaseMultiModalProcessor[KimiK3ProcessingInfo]
Image-only multi-modal processor for Kimi-K3.
Source code in vllm/models/kimi_k3/common/mm_preprocess.py
_get_mm_fields_config(hf_inputs, hf_processor_mm_kwargs)
¶
Slice the flattened patch tensor back into per-image items.
pixel_values holds all patches from every image concatenated; each
image's patch count is prod(grid_thws[i]). grid_thws is one
[N_t, N_h, N_w] row per image.
Source code in vllm/models/kimi_k3/common/mm_preprocess.py
_get_prompt_updates(mm_items, hf_processor_mm_kwargs, out_mm_kwargs)
¶
Expand each K3 image placeholder into a resolution-aware update.
K3's prompt carries a single <|kimi_image_placeholder|> token per
image. This replaces that token with
<|media_begin|>image {w}x{h}<|media_content|>{pads}<|media_end|>,
embedding the per-image resolution in the prompt and marking only the
<|media_pad|> positions as embedding slots (the number of pads is
the feature size returned by media_tokens_calculator).
Source code in vllm/models/kimi_k3/common/mm_preprocess.py
KimiK3ProcessingInfo
¶
Bases: BaseProcessingInfo
Processing information for the image-only Kimi-K3 model.
K3 uses the standard image modality (unlike K2.5's unified
vision_chunk), so it builds its own KimiK3Processor wrapper around
the checkpoint's image processor and resolves the <|media_pad|> token
id the same way K2.5 does.
Source code in vllm/models/kimi_k3/common/mm_preprocess.py
165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 | |