@tduncan / Nvidia-Optimization-Stack-VLLM / commits / 7b44aad5db

[MyPy] Fix mypy errors in `vllm/model_executor/models/[kK]*` (#59402)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

Harry Mellor committed Oct 1, 2026 at 20:32 UTC 7b44aad5db820f8914810ae9f238ffb871e86d44

10 files changed +106 -100

tools/pre_commit/mypy.py

-8

@@ -109,10 +109,6 @@ SEPARATE_GROUPS = [

109 "tests/v1/worker",

110 ]

111

112 -EXCLUDE = [

113 - r"vllm/model_executor/models/[kK]",

114 -]

115 -

112

113 def group_files(changed_files: list[str]) -> dict[str, list[str]]:

114 """Group changed files into different mypy calls.

@@ -124,16 +120,12 @@ def group_files(changed_files: list[str]) -> dict[str, list[str]]:

120 A dictionary mapping file group names to lists of changed files.

121

122 """

127 - exclude_pattern = re.compile(f"^{'|'.join(EXCLUDE)}.*")

123 silent_pattern = re.compile(f"^({'|'.join(SILENT_GROUPS)}).*")

124 file_groups: dict[str, list[str]] = {"": []}

125 file_groups.update({k: [] for k in SEPARATE_GROUPS})

126 # Longest path first so a sub-directory is not shadowed by its parent

127 separate_groups = sorted(SEPARATE_GROUPS, key=len, reverse=True)

128 for changed_file in changed_files:

134 - # Skip files which should be ignored completely

135 - if exclude_pattern.match(changed_file):

136 - continue

129 # Already-fixed paths go in the default group, which runs at the

130 # stricter follow_imports setting from pyproject.toml

131 if silent_pattern.match(changed_file):

vllm/model_executor/models/k2_horizon.py

+28 -17

@@ -502,6 +502,7 @@ class K2HorizonSparseMoeBlock(nn.Module):

502 self.gate.bias = nn.Parameter(self.gate.bias.float(), requires_grad=False)

503

504 self.num_shared_experts = config.num_shared_experts

505 + self.shared_experts: K2HorizonMLP | None

506 if config.num_shared_experts > 0:

507 self.shared_experts = K2HorizonMLP(

508 hidden_size=config.hidden_size,

@@ -644,6 +645,14 @@ class K2HorizonAttention(nn.Module):

645 rope_parameters=rope_parameters,

646 dual_chunk_attention_config=dual_chunk_attention_config,

647 )

648 + attn_kwargs: dict[str, Any] = (

649 + {

650 + "layer_idx": extract_layer_index(prefix),

651 + "dual_chunk_attention_config": dual_chunk_attention_config,

652 + }

653 + if dual_chunk_attention_config

654 + else {}

655 + )

656 self.attn = Attention(

657 self.num_heads,

658 self.head_dim,

@@ -652,12 +661,7 @@ class K2HorizonAttention(nn.Module):

661 cache_config=cache_config,

662 quant_config=quant_config,

663 prefix=f"{prefix}.attn",

655 - **{

656 - "layer_idx": extract_layer_index(prefix),

657 - "dual_chunk_attention_config": dual_chunk_attention_config,

658 - }

659 - if dual_chunk_attention_config

660 - else {},

664 + **attn_kwargs,

665 )

666

667 self.query_key_norm = query_key_norm

@@ -842,6 +846,14 @@ class K2HorizonMoVAAttention(nn.Module):

846 rope_parameters=rope_parameters,

847 dual_chunk_attention_config=dual_chunk_attention_config,

848 )

849 + attn_kwargs: dict[str, Any] = (

850 + {

851 + "layer_idx": extract_layer_index(prefix),

852 + "dual_chunk_attention_config": dual_chunk_attention_config,

853 + }

854 + if dual_chunk_attention_config

855 + else {}

856 + )

857 self.attn = Attention(

858 self.num_heads,

859 self.head_dim,

@@ -850,12 +862,7 @@ class K2HorizonMoVAAttention(nn.Module):

862 cache_config=cache_config,

863 quant_config=quant_config,

864 prefix=f"{prefix}.attn",

853 - **{

854 - "layer_idx": extract_layer_index(prefix),

855 - "dual_chunk_attention_config": dual_chunk_attention_config,

856 - }

857 - if dual_chunk_attention_config

858 - else {},

865 + **attn_kwargs,

866 )

867

868 self.query_key_norm = query_key_norm

@@ -1228,9 +1235,10 @@ class K2HorizonModel(nn.Module, EagleModelMixin):

1235 expert_params_mapping = self.get_expert_mapping()

1236 for name, loaded_weight in weights:

1237 if "scale" in name or "zero_point" in name:

1231 - name = maybe_remap_kv_scale_name(name, params_dict)

1232 - if name is None:

1238 + remapped_name = maybe_remap_kv_scale_name(name, params_dict)

1239 + if remapped_name is None:

1240 continue

1241 + name = remapped_name

1242

1243 # QK norm weights

1244 if name.endswith(".self_attn.q_norm.weight") or name.endswith(

@@ -1267,7 +1275,9 @@ class K2HorizonModel(nn.Module, EagleModelMixin):

1275 kv_rank

1276 ]

1277

1270 - weight_loader = getattr(param, "weight_loader", default_weight_loader)

1278 + weight_loader: Callable[..., Any] = getattr(

1279 + param, "weight_loader", default_weight_loader

1280 + )

1281 weight_loader(param, loaded_weight)

1282 loaded_params.add(name)

1283 continue

@@ -1380,9 +1390,10 @@ class K2HorizonModel(nn.Module, EagleModelMixin):

1390 continue

1391 if name.endswith("scale"):

1392 # Remapping the name of FP8 kv-scale.

1383 - name = maybe_remap_kv_scale_name(name, params_dict)

1384 - if name is None:

1393 + remapped_name = maybe_remap_kv_scale_name(name, params_dict)

1394 + if remapped_name is None:

1395 continue

1396 + name = remapped_name

1397 if name not in params_dict:

1398 continue

1399

vllm/model_executor/models/kanana_v.py

+16 -15

@@ -2,7 +2,7 @@

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

3 from collections.abc import Iterable, Mapping, Sequence

4 from functools import partial

5 -from typing import Annotated, Literal, TypeAlias

5 +from typing import Annotated, Any, Literal, TypeAlias, cast

7 import numpy as np

8 import regex as re

@@ -288,7 +288,7 @@ class CustomQwen2VLVE(Qwen2VisionTransformer):

288 """Drop-in replacement for the HF `_from_config` constructor."""

289 return cls(config)

290

291 - def forward(

291 + def forward( # type: ignore[override]

292 self,

293 pixel_values: torch.Tensor,

294 grid_thw: torch.Tensor,

@@ -331,12 +331,12 @@ class CustomQwen2VLVE(Qwen2VisionTransformer):

331 # Pre-compute seqlens for attention backend.

332 max_seqlen = self.compute_attn_mask_seqlen(cu_seqlens)

333

334 - encoder_states = () if output_hidden_states else None

334 + encoder_states: tuple[torch.Tensor, ...] = ()

335

336 for blk in self.blocks:

337 if output_hidden_states:

338 # Store patch-level states (S, D).

339 - encoder_states = encoder_states + (x.squeeze(1),)

339 + encoder_states += (x.squeeze(1),)

340

341 x = blk(

342 x,

@@ -348,14 +348,15 @@ class CustomQwen2VLVE(Qwen2VisionTransformer):

348

349 # Final hidden state at patch level (S, D).

350 hidden_states = x.squeeze(1)

351 + all_hidden_states = None

352 if output_hidden_states:

352 - encoder_states = encoder_states + (hidden_states,)

353 + all_hidden_states = encoder_states + (hidden_states,)

354

355 if not return_dict:

355 - return tuple(v for v in [hidden_states, encoder_states] if v is not None)

356 + return tuple(v for v in [hidden_states, all_hidden_states] if v is not None)

357 return BaseModelOutput(

358 last_hidden_state=hidden_states,

358 - hidden_states=encoder_states,

359 + hidden_states=all_hidden_states,

360 )

361

362 def get_num_tokens(self) -> int:

@@ -479,7 +480,7 @@ class KananaVMultiModalProcessor(BaseMultiModalProcessor[KananaVProcessingInfo])

480 assert isinstance(prompt_text, str)

481

482 # Images

482 - image_inputs = hf_data.get("images", [])

483 + image_inputs = cast(list[Any], hf_data.get("images", []))

484 pixel_sizes = []

485 if not isinstance(image_inputs[0], Image.Image):

486 image_inputs = [Image.fromarray(image) for image in image_inputs]

@@ -487,9 +488,9 @@ class KananaVMultiModalProcessor(BaseMultiModalProcessor[KananaVProcessingInfo])

488 image_processor = self.info.get_hf_processor().image_processor

489 processor_output = [image_processor(image) for image in image_inputs]

490 pixel_values = [o["pixel_values"] for o in processor_output]

490 - image_meta = [o["image_meta"] for o in processor_output]

491 + image_meta_list = [o["image_meta"] for o in processor_output]

492 # list of dict -> dict of list

492 - image_meta = {k: [d[k] for d in image_meta] for k in image_meta[0]}

493 + image_meta = {k: [d[k] for d in image_meta_list] for k in image_meta_list[0]}

494

495 for pixel_value in pixel_values:

496 pixel_sizes.append(pixel_value.shape[0])

@@ -545,7 +546,7 @@ class KananaVMultiModalProcessor(BaseMultiModalProcessor[KananaVProcessingInfo])

546 ) -> Sequence[PromptUpdate]:

547 tokenizer = self.info.get_tokenizer()

548

548 - def get_replacement(idx: int) -> Sequence[int]:

549 + def get_replacement(idx: int) -> list[int]:

550 out_item = out_mm_kwargs["image"][idx]

551 image_token_thw = out_item["image_token_thw"].data

552 assert isinstance(image_token_thw, torch.Tensor)

@@ -683,7 +684,7 @@ class KananaVForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP)

684 def _get_visual_feature_at(

685 self,

686 v_output: Sequence[torch.Tensor],

686 - layer_index: int | Sequence[int],

687 + layer_index: int | list[int] | tuple[int, ...],

688 ) -> torch.Tensor:

689 if isinstance(layer_index, (list, tuple)):

690 visual_features = torch.stack(v_output, dim=1)[

@@ -696,7 +697,7 @@ class KananaVForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP)

697 def forward_vision(

698 self,

699 pixel_values: torch.Tensor,

699 - image_metas: dict | None = None,

700 + image_metas: dict,

701 ) -> torch.Tensor:

702 vision_model_args = {

703 "pixel_values": pixel_values,

@@ -714,7 +715,7 @@ class KananaVForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP)

715 def forward_projector(

716 self,

717 visual_features: torch.Tensor,

717 - image_metas: dict | None = None,

718 + image_metas: dict,

719 ) -> torch.Tensor:

720 visual_embeds = self.abstractor(

721 visual_features,

@@ -725,7 +726,7 @@ class KananaVForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP)

726 def forward_and_project_vision(

727 self,

728 pixel_values: torch.Tensor,

728 - image_metas: dict | None = None,

729 + image_metas: dict,

730 ) -> torch.Tensor:

731 assert pixel_values is not None

732 visual_features = self.forward_vision(pixel_values, image_metas=image_metas)

vllm/model_executor/models/keye.py

+29 -34

@@ -208,8 +208,8 @@ class KeyeVisionEmbeddings(nn.Module):

208

209 self.num_patches = (self.image_size // self.patch_size) ** 2

210 self.num_positions = self.num_patches

211 - self.cache_position_embedding = dict()

212 - self.cache_position_count = dict()

211 + self.cache_position_embedding: dict[tuple[int, int], torch.Tensor] = {}

212 + self.cache_position_count: dict[tuple[int, int], int] = {}

213 self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)

214 self.packing_position_embedding = nn.Embedding(32768, self.embed_dim)

215

@@ -264,7 +264,7 @@ class KeyeVisionEmbeddings(nn.Module):

264 if len(self.cache_position_embedding) >= max_cache:

265 min_hit_grid = min(

266 self.cache_position_count,

267 - key=self.cache_position_count.get,

267 + key=self.cache_position_count.__getitem__,

268 )

269 self.cache_position_count.pop(min_hit_grid)

270 self.cache_position_embedding.pop(min_hit_grid)

@@ -278,8 +278,7 @@ class KeyeVisionEmbeddings(nn.Module):

278 self,

279 pixel_values: torch.FloatTensor,

280 position_ids: torch.Tensor | None = None,

281 - image_grid_thw: list[tuple[int, int, int] | list[tuple[int, int, int]]]

282 - | None = None,

281 + image_grid_thw: list[tuple[int, int, int]] | None = None,

282 interpolate_pos_encoding=False,

283 ) -> torch.Tensor:

284 if pixel_values.dim() == 4:

@@ -408,7 +407,7 @@ class KeyeSiglipAttention(nn.Module):

407 hidden_states: torch.Tensor,

408 attention_mask: torch.Tensor | None = None,

409 output_attentions: bool | None = False,

411 - cu_seqlens: list[torch.Tensor] | None = None,

410 + cu_seqlens: torch.Tensor | None = None,

411 rope_emb: tuple[torch.Tensor, torch.Tensor] | None = None,

412 ) -> torch.Tensor:

413 qkv, _ = self.qkv_proj(hidden_states)

@@ -417,6 +416,8 @@ class KeyeSiglipAttention(nn.Module):

416 dim=-1,

417 )

418

419 + if cu_seqlens is None:

420 + raise ValueError("cu_seqlens cannot be None.")

421 max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()

422

423 if rope_emb is None:

@@ -432,8 +433,6 @@ class KeyeSiglipAttention(nn.Module):

433 self.head_dim,

434 )

435 else:

435 - if cu_seqlens is None:

436 - raise ValueError("cu_seqlens cannot be None when rope_emb is not None.")

436 cos, sin = rope_emb

437 q = q.view(*q.shape[:-1], self.num_heads, self.head_dim)

438 k = k.view(

@@ -511,7 +510,7 @@ class KeyeSiglipEncoderLayer(nn.Module):

510 hidden_states: torch.Tensor,

511 attention_mask: torch.Tensor,

512 output_attentions: bool | None = False,

514 - cu_seqlens: list[torch.Tensor] | None = None,

513 + cu_seqlens: torch.Tensor | None = None,

514 rope_emb: tuple[torch.Tensor, torch.Tensor] | None = None,

515 ) -> tuple[torch.FloatTensor]:

516 residual = hidden_states

@@ -576,13 +575,12 @@ class KeyeSiglipEncoder(nn.Module):

575 attention_mask: torch.Tensor | None = None,

576 output_attentions: bool | None = None,

577 output_hidden_states: bool | None = None,

579 - cu_seqlens: list[torch.Tensor] | None = None,

580 - image_grid_thw: list[tuple[int, int, int] | list[tuple[int, int, int]]]

581 - | None = None,

578 + cu_seqlens: torch.Tensor | None = None,

579 + image_grid_thw: list[tuple[int, int, int]] | None = None,

580 height_position_ids: torch.Tensor | None = None,

581 width_position_ids: torch.Tensor | None = None,

582 use_rope: bool | None = False,

585 - window_size: bool | None = -1,

583 + window_size: int = -1,

584 vision_or_text: str = "vision",

585 ) -> BaseModelOutput:

586 device = inputs_embeds.device

@@ -662,14 +660,13 @@ class KeyeSiglipVisionTransformer(nn.Module):

660 position_ids: torch.Tensor | None = None,

661 height_position_ids: torch.Tensor | None = None,

662 width_position_ids: torch.Tensor | None = None,

665 - cu_seqlens: list[torch.Tensor] | None = None,

663 + cu_seqlens: torch.Tensor | None = None,

664 padding_mask: torch.Tensor | None = None,

665 vision_return_embed_list: bool | None = False,

668 - image_grid_thw: list[tuple[int, int, int] | list[tuple[int, int, int]]]

669 - | None = None,

666 + image_grid_thw: list[tuple[int, int, int]] | None = None,

667 return_pooler_output: bool | None = True,

668 use_rope: bool | None = False,

672 - window_size: bool | None = -1,

669 + window_size: int = -1,

670 ) -> BaseModelOutputWithPooling:

671 hidden_states = self.embeddings(

672 pixel_values,

@@ -757,12 +754,11 @@ class KeyeSiglipVisionModel(nn.Module):

754 interpolate_pos_encoding: bool = False,

755 position_ids: torch.Tensor | None = None,

756 vision_return_embed_list: bool | None = False,

760 - image_grid_thw: list[tuple[int, int, int] | list[tuple[int, int, int]]]

761 - | None = None,

762 - cu_seqlens: list[torch.Tensor] | None = None,

757 + image_grid_thw: list[tuple[int, int, int]] | None = None,

758 + cu_seqlens: torch.Tensor | None = None,

759 return_pooler_output: bool | None = True,

760 use_rope: bool | None = False,

765 - window_size: bool | None = -1,

761 + window_size: int = -1,

762 ) -> BaseModelOutputWithPooling:

763 return self.vision_model(

764 pixel_values=pixel_values,

@@ -1337,7 +1333,7 @@ class BaseKeyeModule(nn.Module, SupportsMultiModal):

1333 def _process_video_embeds(

1334 self,

1335 video_type: Literal["video_embeds", "pixel_values_videos"],

1340 - video_grid_thw: list[torch.Tensor],

1336 + video_grid_thw: torch.Tensor,

1337 pixel_values_videos: torch.Tensor | None = None,

1338 ) -> torch.Tensor | list[torch.Tensor]:

1339 siglip_position_ids = list()

@@ -1361,6 +1357,7 @@ class BaseKeyeModule(nn.Module, SupportsMultiModal):

1357 "Video embeddings are not supported for this processing path."

1358 )

1359 else:

1360 + assert pixel_values_videos is not None

1361 pixel_values_videos = pixel_values_videos.type(self.visual.dtype)

1362 # These are host-built; concat straight into pinned buffers so

1363 # the H2D copies stay non-blocking.

@@ -1535,12 +1532,11 @@ class KeyeForConditionalGeneration(

1532 image_grid_thw=image_grid_thw,

1533 )

1534

1538 - if image_embeds is not None:

1539 - return KeyeImageEmbeddingInputs(

1540 - type="image_embeds",

1541 - image_embeds=image_embeds,

1542 - image_grid_thw=image_grid_thw,

1543 - )

1535 + return KeyeImageEmbeddingInputs(

1536 + type="image_embeds",

1537 + image_embeds=image_embeds,

1538 + image_grid_thw=image_grid_thw,

1539 + )

1540

1541 def _parse_and_validate_video_input(

1542 self, **kwargs: object

@@ -1559,12 +1555,11 @@ class KeyeForConditionalGeneration(

1555 video_grid_thw=video_grid_thw,

1556 )

1557

1562 - if video_embeds is not None:

1563 - return KeyeVideoEmbeddingInputs(

1564 - type="video_embeds",

1565 - video_embeds=video_embeds,

1566 - video_grid_thw=video_grid_thw,

1567 - )

1558 + return KeyeVideoEmbeddingInputs(

1559 + type="video_embeds",

1560 + video_embeds=video_embeds,

1561 + video_grid_thw=video_grid_thw,

1562 + )

1563

1564 def _process_video_input(

1565 self, video_input: KeyeVideoInputs

vllm/model_executor/models/keye_vl1_5.py

+13 -15

@@ -75,7 +75,7 @@ def split_thw(grid_thw: torch.Tensor) -> torch.Tensor:

76 def get_num_patches(

77 grid_thw: torch.Tensor, num_frames: list[int] | torch.Tensor

78 -) -> list[int]:

78 +) -> torch.Tensor:

79 """Return num_patches per video.

81 Args:

@@ -84,7 +84,7 @@ def get_num_patches(

84 num_frames: List or tensor indicating the number of frames per video

86 Returns:

87 - List of ints representing the number of patches for each video

87 + Tensor of the number of patches for each video

89 Examples:

90 >>> # Suppose there are 2 videos with a total of 3 grids

@@ -538,12 +538,11 @@ class KeyeVL1_5ForConditionalGeneration(

538 image_grid_thw=image_grid_thw,

539 )

540

541 - if image_embeds is not None:

542 - return KeyeVL1_5ImageEmbeddingInputs(

543 - type="image_embeds",

544 - image_embeds=image_embeds,

545 - image_grid_thw=image_grid_thw,

546 - )

541 + return KeyeVL1_5ImageEmbeddingInputs(

542 + type="image_embeds",

543 + image_embeds=image_embeds,

544 + image_grid_thw=image_grid_thw,

545 + )

546

547 def _parse_and_validate_video_input(

548 self, **kwargs: object

@@ -564,13 +563,12 @@ class KeyeVL1_5ForConditionalGeneration(

563 num_frames=num_frames,

564 )

565

567 - if video_embeds is not None:

568 - return KeyeVL1_5VideoEmbeddingInputs(

569 - type="video_embeds",

570 - video_embeds=video_embeds,

571 - video_grid_thw=video_grid_thw,

572 - num_frames=num_frames,

573 - )

566 + return KeyeVL1_5VideoEmbeddingInputs(

567 + type="video_embeds",

568 + video_embeds=video_embeds,

569 + video_grid_thw=video_grid_thw,

570 + num_frames=num_frames,

571 + )

572

573 def _process_video_input(

574 self, video_input: KeyeVL1_5VideoInputs

vllm/model_executor/models/kimi_audio.py

+3 -1

@@ -18,6 +18,7 @@ from vllm.config.speech_to_text import SpeechToTextParams

18 from vllm.inputs import PromptType, TokensPrompt

19 from vllm.model_executor.model_loader import DefaultModelLoader

20 from vllm.model_executor.models.interfaces import (

21 + MultiModalEmbeddings,

22 SupportsMultiModal,

23 SupportsPP,

24 SupportsTranscription,

@@ -240,6 +241,7 @@ class KimiAudioMultiModalProcessor(BaseMultiModalProcessor[KimiAudioProcessingIn

241 # KimiAudioProcessor expects raw numpy arrays

242 mm_data = hf_inputs.hf_data

243 if audios := mm_data.pop("audio", []):

244 + assert isinstance(audios, Sequence)

245 audio_arrays = []

246 for aud in audios:

247 if isinstance(aud, (tuple, list)) and len(aud) == 2:

@@ -485,7 +487,7 @@ class KimiAudioForConditionalGeneration(

487 def embed_input_ids(

488 self,

489 input_ids: torch.Tensor,

488 - multimodal_embeddings: tuple[torch.Tensor, ...] | None = None,

490 + multimodal_embeddings: MultiModalEmbeddings | None = None,

491 *,

492 is_multimodal: torch.Tensor | None = None,

493 ) -> torch.Tensor:

vllm/model_executor/models/kimi_k25.py

+12 -6

@@ -136,9 +136,9 @@ class KimiK25ProcessingInfo(BaseProcessingInfo):

136 # may remap token IDs vs config.json.

137 config_token_id = hf_config.media_placeholder_token_id

138 resolved_token_id = tokenizer.convert_tokens_to_ids("<|media_pad|>")

139 + unk_token_id = getattr(tokenizer, "unk_token_id", None)

140 is_valid_resolved = isinstance(resolved_token_id, int) and (

140 - tokenizer.unk_token_id is None

141 - or resolved_token_id != tokenizer.unk_token_id

141 + unk_token_id is None or resolved_token_id != unk_token_id

142 )

143 if is_valid_resolved and resolved_token_id != config_token_id:

144 logger.warning_once(

@@ -165,7 +165,7 @@ class KimiK25ProcessingInfo(BaseProcessingInfo):

165 )

166 self.media_tokens_calculator = image_processor.media_tokens_calculator

167

168 - def get_hf_processor(self):

168 + def get_hf_processor(self, **kwargs: object):

169 return self.hf_processor

170

171 def get_hf_config(self):

@@ -191,7 +191,11 @@ class KimiK25DummyInputsBuilder(BaseDummyInputsBuilder[KimiK25ProcessingInfo]):

191 )

192

193 video_chunk_dummy_item = VisionChunkVideo(

194 - type="video_chunk", video_chunk=dummy_videos

194 + type="video_chunk",

195 + video_chunk=dummy_videos,

196 + uuid=None,

197 + prompt="",

198 + video_idx=0,

199 )

200 video_chunk_num_tokens = self.info.media_tokens_calculator(

201 video_chunk_dummy_item

@@ -204,6 +208,7 @@ class KimiK25DummyInputsBuilder(BaseDummyInputsBuilder[KimiK25ProcessingInfo]):

208 width=MaxImageTokenMeta.width,

209 num_images=1,

210 )[0],

211 + uuid=None,

212 )

213 image_num_tokens = self.info.media_tokens_calculator(image_dummy_item)

214 # return the larger one

@@ -377,7 +382,7 @@ class KimiK25ForConditionalGeneration(

382 )

383 self.media_placeholder: int = self.config.media_placeholder_token_id

384

380 - def _maybe_ignore_quant_config(self, quant_config: QuantizationConfig):

385 + def _maybe_ignore_quant_config(self, quant_config: QuantizationConfig | None):

386 if isinstance(quant_config, compressed_tensors.CompressedTensorsConfig):

387 return None

388 return quant_config

@@ -392,6 +397,7 @@ class KimiK25ForConditionalGeneration(

397

398 if isinstance(pixel_values, list):

399 pixel_values = torch.cat(pixel_values, dim=0)

400 + assert isinstance(pixel_values, torch.Tensor)

401

402 if len(pixel_values.shape) == 5 or len(pixel_values.shape) == 3:

403 pixel_values = pixel_values.reshape(

@@ -612,7 +618,7 @@ class KimiK25ForConditionalGeneration(

618

619 grid_thw_list = [[1, ho * kh, wo * kw] for _ in range(max_batch_size)]

620

615 - ps = self.vision_tower.patch_size

621 + ps: int | tuple[int, int] = self.vision_tower.patch_size

622 if isinstance(ps, int):

623 ps = (ps, ps)

624 total_patches = max_batch_size * ho * kh * wo * kw

vllm/model_executor/models/kimi_k25_vit.py

+1 -1

@@ -804,7 +804,7 @@ class MoonViT3dPretrainedModel(nn.Module):

804 self,

805 grid_thw_list: list[list[int]],

806 *,

807 - max_batch_size: int,

807 + max_batch_size: int | None,

808 max_seqlen_override: int | None = None,

809 device: torch.device,

810 ) -> dict[str, torch.Tensor | None]:

vllm/model_executor/models/kimi_vl.py

+3 -2

@@ -268,6 +268,7 @@ class KimiVLMultiModalProcessor(BaseMultiModalProcessor[KimiVLProcessingInfo]):

268 if isinstance(images, ImageEmbeddingItems):

269 num_image_tokens = images.get_feature_size(item_idx)

270 else:

271 + assert isinstance(images, ImageProcessorItems)

272 image_size = images.get_image_size(item_idx)

273 num_image_tokens = self.info.get_num_image_tokens(

274 image_width=image_size.width,

@@ -317,7 +318,7 @@ class KimiVLForConditionalGeneration(

318

319 assert isinstance(config.vision_config, MoonViTConfig)

320 self.use_data_parallel = (

320 - model_config.multimodal_config.mm_encoder_tp_mode == "data"

321 + model_config.get_multimodal_config().mm_encoder_tp_mode == "data"

322 )

323 self.hidden_size = config.text_config.hidden_size

324

@@ -469,7 +470,7 @@ class KimiVLForConditionalGeneration(

470 )

471 grid_hws_list = [(ho * kh, wo * kw) for _ in range(max_batch_size)]

472

472 - patch_size = self.config.vision_config.patch_size

473 + patch_size: int | tuple[int, int] = self.config.vision_config.patch_size

474 if isinstance(patch_size, int):

475 patch_size = (patch_size, patch_size)

476

vllm/transformers_utils/configs/kimi_vl.py

+1 -1

@@ -30,7 +30,7 @@ class KimiVLConfig(PreTrainedConfig):

30 text_config = DeepseekV2Config()

31 elif isinstance(text_config, dict):

32 text_config = DeepseekV2Config(**text_config)

33 - self.text_config = text_config

33 + self.text_config: DeepseekV2Config = text_config

35 self.ignore_index = ignore_index

36 self.media_placeholder_token_id = media_placeholder_token_id