9a8966bcc2 · tduncan/Nvidia-Optimization-Stack-VLLM

vllm/distributed/eplb/eplb_state.py

+4 -4

@@ -337,11 +337,11 @@ class EplbState:

337 Args:

338 model (MixtureOfExperts): The MoE model.

339 is_dummy (bool): If `True`, this is a dummy step and the load

340 - metrics recorded in this forward pass will not count. Defaults

341 - to `False`.

340 + metrics recorded in this forward pass will not count. Defaults

341 + to `False`.

342 is_profile (bool): If `True`, perform a dummy rearrangement

343 - with maximum communication cost. This is used in `profile_run`

344 - to reserve enough memory for the communication buffer.

343 + with maximum communication cost. This is used in `profile_run`

344 + to reserve enough memory for the communication buffer.

345 log_stats (bool): If `True`, log the expert load metrics.

346

347 # Stats

vllm/distributed/eplb/rebalance_algo.py

+3 -3

@@ -102,14 +102,14 @@ def rebalance_experts_hierarchical(

102 num_groups: int,

103 num_nodes: int,

104 num_gpus: int,

105 -):

105 +) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:

106 """

107 Parameters:

108 weight: [num_moe_layers, num_logical_experts]

109 num_physical_experts: number of physical experts after replication

110 num_groups: number of expert groups

111 - num_nodes: number of server nodes, where the intra-node network

112 - (e.g, NVLink) is faster

111 + num_nodes: number of server nodes, where the intra-node network

112 + (e.g, NVLink) is faster

113 num_gpus: number of GPUs, must be a multiple of `num_nodes`

114

115 Returns:

vllm/distributed/kv_transfer/kv_connector/v1/base.py

+3 -2

@@ -149,7 +149,7 @@ class KVConnectorBase_V1(ABC):

149

150 @abstractmethod

151 def start_load_kv(self, forward_context: "ForwardContext",

152 - **kwargs) -> None:

152 + **kwargs: Any) -> None:

153 """

154 Start loading the KV cache from the connector to vLLM's paged

155 KV buffer. This is called from the forward context before the

@@ -182,7 +182,8 @@ class KVConnectorBase_V1(ABC):

182

183 @abstractmethod

184 def save_kv_layer(self, layer_name: str, kv_layer: torch.Tensor,

185 - attn_metadata: "AttentionMetadata", **kwargs) -> None:

185 + attn_metadata: "AttentionMetadata",

186 + **kwargs: Any) -> None:

187 """

188 Start saving a layer of KV cache from vLLM's paged buffer

189 to the connector. This is called from within attention layer to

vllm/distributed/kv_transfer/kv_connector/v1/lmcache_connector.py

+3 -2

@@ -30,7 +30,7 @@ class LMCacheConnectorV1(KVConnectorBase_V1):

30 # Worker-side methods

31 # ==============================

32 def start_load_kv(self, forward_context: "ForwardContext",

33 - **kwargs) -> None:

33 + **kwargs: Any) -> None:

34 """

35 Start loading the KV cache from the connector to vLLM's paged

36 KV buffer. This is called from the forward context before the

@@ -61,7 +61,8 @@ class LMCacheConnectorV1(KVConnectorBase_V1):

61 self._lmcache_engine.wait_for_layer_load(layer_name)

62

63 def save_kv_layer(self, layer_name: str, kv_layer: torch.Tensor,

64 - attn_metadata: "AttentionMetadata", **kwargs) -> None:

64 + attn_metadata: "AttentionMetadata",

65 + **kwargs: Any) -> None:

66 """

67 Start saving the a layer of KV cache from vLLM's paged buffer

68 to the connector. This is called from within attention layer to

vllm/distributed/kv_transfer/kv_connector/v1/p2p/p2p_nccl_connector.py

+4 -3

@@ -91,7 +91,7 @@ class P2pNcclConnector(KVConnectorBase_V1):

91 # ==============================

92

93 def start_load_kv(self, forward_context: "ForwardContext",

94 - **kwargs) -> None:

94 + **kwargs: Any) -> None:

95 """Start loading the KV cache from the connector buffer to vLLM's

96 paged KV buffer.

97

@@ -212,7 +212,8 @@ class P2pNcclConnector(KVConnectorBase_V1):

212 return

213

214 def save_kv_layer(self, layer_name: str, kv_layer: torch.Tensor,

215 - attn_metadata: "AttentionMetadata", **kwargs) -> None:

215 + attn_metadata: "AttentionMetadata",

216 + **kwargs: Any) -> None:

217 """Start saving the KV cache of the layer from vLLM's paged buffer

218 to the connector.

219

@@ -278,7 +279,7 @@ class P2pNcclConnector(KVConnectorBase_V1):

279

280 def get_finished(

281 self, finished_req_ids: set[str],

281 - **kwargs) -> tuple[Optional[set[str]], Optional[set[str]]]:

282 + **kwargs: Any) -> tuple[Optional[set[str]], Optional[set[str]]]:

283 """

284 Notifies worker-side connector ids of requests that have

285 finished generating tokens.

vllm/distributed/kv_transfer/kv_connector/v1/p2p/tensor_memory_pool.py

+3 -2

@@ -218,8 +218,9 @@ class TensorMemoryPool:

218

219 return addr

220

221 - def load_tensor(self, addr: int, dtype: torch.dtype,

222 - shape: tuple[int, ...], device) -> torch.Tensor:

221 + def load_tensor(self, addr: int, dtype: torch.dtype, shape: tuple[int,

222 + ...],

223 + device: torch.device) -> torch.Tensor:

224 """Loads a tensor from pinned host memory to the specified device.

225

226 Args:

vllm/distributed/kv_transfer/kv_connector/v1/shared_storage_connector.py

+4 -3

@@ -3,7 +3,7 @@

3 import hashlib

4 import os

5 from dataclasses import dataclass

6 -from typing import TYPE_CHECKING, Optional

6 +from typing import TYPE_CHECKING, Any, Optional

7

8 import safetensors

9 import torch

@@ -90,7 +90,7 @@ class SharedStorageConnector(KVConnectorBase_V1):

90 logger.info("Shared storage path is %s", self._storage_path)

91

92 def start_load_kv(self, forward_context: "ForwardContext",

93 - **kwargs) -> None:

93 + **kwargs: Any) -> None:

94 """Start loading the KV cache from the connector buffer to vLLM's

95 paged KV buffer.

96

@@ -191,7 +191,8 @@ class SharedStorageConnector(KVConnectorBase_V1):

191 return

192

193 def save_kv_layer(self, layer_name: str, kv_layer: torch.Tensor,

194 - attn_metadata: "AttentionMetadata", **kwargs) -> None:

194 + attn_metadata: "AttentionMetadata",

195 + **kwargs: Any) -> None:

196 """Start saving the KV cache of the layer from vLLM's paged buffer

197 to the connector.

198

vllm/distributed/kv_transfer/kv_pipe/pynccl_pipe.py

+2 -2

@@ -251,8 +251,8 @@ class PyNcclPipe(KVPipeBase):

251 """

252 Receives a tensor and its metadata from the source rank. Blocking call.

253

254 - Args:

255 - tensor: The received tensor, or `None` if no tensor is received.

254 + Returns:

255 + The received tensor, or `None` if no tensor is received.

256 """

257 if self.transport_thread is None:

258 self.transport_thread = ThreadPoolExecutor(max_workers=1)

vllm/model_executor/models/interfaces.py

+1 -1

@@ -823,7 +823,7 @@ class SupportsEagle3(Protocol):

823

824 Args:

825 layers: Tuple of layer indices that should output auxiliary

826 - hidden states.

826 + hidden states.

827 """

828 ...

829

vllm/model_executor/models/keye.py

+3 -9

@@ -1520,15 +1520,9 @@ class BaseKeyeModule(nn.Module):

1520 batch.

1521 **NOTE**: If mrope is enabled (default setting for Qwen2-VL

1522 opensource models), the shape will be `(3, seq_len)`,

1523 - otherwise it will be `(seq_len,).

1524 - pixel_values: Pixel values to be fed to a model.

1525 - `None` if no images are passed.

1526 - image_grid_thw: Tensor `(n_images, 3)` of image 3D grid in LLM.

1527 - `None` if no images are passed.

1528 - pixel_values_videos: Pixel values of videos to be fed to a model.

1529 - `None` if no videos are passed.

1530 - video_grid_thw: Tensor `(n_videos, 3)` of video 3D grid in LLM.

1531 - `None` if no videos are passed.

1523 + otherwise it will be `(seq_len,)`.

1524 + intermediate_tensors: Intermediate tensors from prior forward pass.

1525 + inputs_embeds: Optional tensor of input embeddings.

1526 """

1527 if intermediate_tensors is not None:

1528 inputs_embeds = None

vllm/model_executor/models/keye_vl1_5.py

+6 -5

@@ -58,17 +58,18 @@ def split_thw(grid_thw: torch.Tensor) -> torch.Tensor:

58 return torch.cat([ones, h_w], dim=1).repeat_interleave(t, dim=0)

59

60

61 -def get_num_patches(grid_thw: torch.Tensor, num_frames: Union[list[int],

62 - torch.Tensor]):

61 +def get_num_patches(grid_thw: torch.Tensor,

62 + num_frames: Union[list[int], torch.Tensor]) -> list[int]:

63 """

64 Return num_patches per video.

65

66 Args:

67 - t: tensor with shape [N, ...] where each item is a list/tensor

68 - cu_seqlens: list indicating the boundaries of groups

67 + grid_thw: Tensor with shape [N, 3] containing temporal, height, width

68 + dimensions

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

70

71 Returns:

71 - list of ints representing the sum of products for each group

72 + List of ints representing the number of patches for each video

73

74 Examples:

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

vllm/model_executor/models/llava.py

+3 -1

@@ -732,7 +732,9 @@ class LlavaForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP):

732 Args:

733 input_ids: Flattened (concatenated) input_ids corresponding to a

734 batch.

735 - pixel_values: The pixels in each input image.

735 + positions: Position indices for the input tokens.

736 + intermediate_tensors: Intermediate tensors from prior forward pass.

737 + inputs_embeds: Optional tensor of input embeddings.

738

739 Info:

740 [LlavaImageInputs][]

vllm/model_executor/models/llava_next.py

+3 -2

@@ -535,8 +535,9 @@ class LlavaNextForConditionalGeneration(nn.Module, SupportsMultiModal,

535 Args:

536 input_ids: Flattened (concatenated) input_ids corresponding to a

537 batch.

538 - pixel_values: The pixels in each grid patch for each input image.

539 - image_sizes: The original `(height, width)` for each input image.

538 + positions: Position indices for the input tokens.

539 + intermediate_tensors: Intermediate tensors from prior forward pass.

540 + inputs_embeds: Optional tensor of input embeddings.

541

542 Info:

543 [LlavaNextImageInputs][]

vllm/model_executor/models/mistral3.py

+3 -1

@@ -578,7 +578,9 @@ class Mistral3ForConditionalGeneration(nn.Module, SupportsLoRA,

578 Args:

579 input_ids: Flattened (concatenated) input_ids corresponding to a

580 batch.

581 - pixel_values: The pixels in each input image.

581 + positions: Position indices for the input tokens.

582 + intermediate_tensors: Intermediate tensors from prior forward pass.

583 + inputs_embeds: Optional tensor of input embeddings.

584

585 Info:

586 [Mistral3ImagePixelInputs][]

vllm/model_executor/models/mllama4.py

+4 -5

@@ -387,11 +387,10 @@ class Llama4VisionEncoder(nn.Module):

387 ) -> torch.Tensor:

388 r"""

389 Args:

390 - inputs_embeds (`torch.FloatTensor` of shape

391 - `(batch_size, sequence_length, hidden_size)`):

392 - Optionally, instead of passing `input_ids` you can choose to

393 - directly pass an embedded representation. This is useful if you

394 - want more control over how to convert `input_ids` indices into

390 + hidden_states: Input tensor of shape

391 + (batch_size, sequence_length, hidden_size).

392 + Hidden states from the model embeddings, representing

393 + the input tokens.

394 associated vectors than the model's internal embedding

395 lookup matrix.

396 """

vllm/model_executor/models/moonvit.py

+14 -4

@@ -70,11 +70,15 @@ def multihead_attention(

70 v: torch.Tensor,

71 q_cu_seqlens: Optional[torch.Tensor] = None,

72 k_cu_seqlens: Optional[torch.Tensor] = None,

73 -):

73 +) -> torch.Tensor:

74 """Multi-head attention using flash attention 2.

75

76 Args:

77 - q, k, v: tensor of shape (batch_size, seqlen, num_heads, head_dim),

77 + q: Query tensor of shape (batch_size, seqlen, num_heads, head_dim),

78 + or (tot_seqlens, num_heads, head_dim) if packing.

79 + k: Key tensor of shape (batch_size, seqlen, num_heads, head_dim),

80 + or (tot_seqlens, num_heads, head_dim) if packing.

81 + v: Value tensor of shape (batch_size, seqlen, num_heads, head_dim),

82 or (tot_seqlens, num_heads, head_dim) if packing.

83 q_cu_seqlens (torch.Tensor): cumulative sequence lengths of q.

84 The first element should be 0 and the last element should be q.shape[0].

@@ -123,8 +127,14 @@ def sdpa_attention(

127 """SDPA attention.

128

129 Args:

126 - q, k, v: tensor of shape (batch_size, seqlen, num_heads, head_dim),

130 + q: Query tensor of shape (batch_size, seqlen, num_heads, head_dim),

131 + or (tot_seqlens, num_heads, head_dim) if packing.

132 + k: Key tensor of shape (batch_size, seqlen, num_heads, head_dim),

133 + or (tot_seqlens, num_heads, head_dim) if packing.

134 + v: Value tensor of shape (batch_size, seqlen, num_heads, head_dim),

135 or (tot_seqlens, num_heads, head_dim) if packing.

136 + q_cu_seqlens: Optional cumulative sequence lengths of q.

137 + k_cu_seqlens: Optional cumulative sequence lengths of k.

138 """

139 seq_length = q.shape[0]

140 attention_mask = torch.zeros([1, seq_length, seq_length],

@@ -387,7 +397,7 @@ class MLP2(nn.Module):

397 def __init__(self,

398 dims: list[int],

399 activation,

390 - bias=True,

400 + bias: bool = True,

401 prefix: str = "",

402 use_data_parallel: bool = False):

403 super().__init__()

vllm/model_executor/models/phi4_multimodal.py

+2 -2

@@ -374,8 +374,8 @@ class Phi4MMAudioMeanVarianceNormLayer(nn.Module):

374 Typically used as a very first layer in a model.

375

376 Args:

377 - input_size: int

378 - layer input size.

377 + config: [Phi4MultimodalAudioConfig](https://huggingface.co/docs/transformers/model_doc/phi4_multimodal#transformers.Phi4MultimodalAudioConfig)

378 + object containing model parameters.

379 """

380

381 def __init__(self, config: Phi4MultimodalAudioConfig):

vllm/model_executor/models/qwen2_vl.py

+3 -9

@@ -1372,15 +1372,9 @@ class Qwen2VLForConditionalGeneration(nn.Module, SupportsMultiModal,

1372 batch.

1373 **NOTE**: If mrope is enabled (default setting for Qwen2-VL

1374 opensource models), the shape will be `(3, seq_len)`,

1375 - otherwise it will be `(seq_len,).

1376 - pixel_values: Pixel values to be fed to a model.

1377 - `None` if no images are passed.

1378 - image_grid_thw: Tensor `(n_images, 3)` of image 3D grid in LLM.

1379 - `None` if no images are passed.

1380 - pixel_values_videos: Pixel values of videos to be fed to a model.

1381 - `None` if no videos are passed.

1382 - video_grid_thw: Tensor `(n_videos, 3)` of video 3D grid in LLM.

1383 - `None` if no videos are passed.

1375 + otherwise it will be `(seq_len,)`.

1376 + intermediate_tensors: Intermediate tensors from prior forward pass.

1377 + inputs_embeds: Optional tensor of input embeddings.

1378 """

1379

1380 if intermediate_tensors is not None:

vllm/model_executor/models/siglip2navit.py

+8 -19

@@ -390,12 +390,9 @@ class Siglip2EncoderLayer(nn.Module):

390 position_embeddings: torch.Tensor) -> tuple[torch.FloatTensor]:

391 """

392 Args:

393 - hidden_states (`torch.FloatTensor`):

394 - Input to the layer of shape `(batch, seq_len, embed_dim)`.

395 - output_attentions (`bool`, *optional*, defaults to `False`):

396 - Whether or not to return the attentions tensors of all

397 - attention layers. See `attentions` under

398 - returned tensors for more detail.

393 + hidden_states: Input tensor of shape (batch, seq_len, embed_dim).

394 + cu_seqlens: Cumulative sequence lengths tensor.

395 + position_embeddings: Position embeddings tensor.

396 """

397 residual = hidden_states

398

@@ -534,19 +531,11 @@ class Siglip2Encoder(nn.Module):

531 ) -> torch.Tensor:

532 r"""

533 Args:

537 - inputs_embeds (`torch.FloatTensor` of shape

538 - `(batch_size, sequence_length, hidden_size)`):

539 - Optionally, instead of passing `input_ids` you can choose to

540 - directly pass an embedded representation. This is useful if

541 - you want more control over how to convert `input_ids` indices

542 - into associated vectors than the model's internal embedding

543 - lookup matrix.

544 - grid_thws (`torch.LongTensor`):

545 - grid shape (num_patches, 3)

546 - output_hidden_states (`bool`, *optional*):

547 - Whether or not to return the hidden states of all layers. See

548 - `hidden_states` under returned tensors for more detail.

549 - return_dict (`bool`, *optional*):

534 + inputs_embeds: Input tensor of shape

535 + (batch_size, sequence_length, hidden_size).

536 + Embedded representation of the input tokens.

537 + grid_thws: Grid tensor of shape (num_patches, 3)

538 + containing grid dimensions.

539 Whether or not to return a [`~utils.ModelOutput`] instead of

540 a plain tuple.

541 """

vllm/model_executor/models/ultravox.py

+5 -4

@@ -597,10 +597,11 @@ class UltravoxModel(nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA):

597 with the `input_ids`.

598

599 Args:

600 - audio_features: A batch of audio input chunks [B, N, 80, M].

601 - audio_lens: Length of audio frames for each audio chunk [B].

602 - audio_token_len: Length of audio tokens for each audio chunk [B'].

603 - Note: batch dim is different from batch dim in audio chunks.

600 + input_ids: Flattened (concatenated) input_ids corresponding to a

601 + batch.

602 + positions: Position indices for the input tokens.

603 + intermediate_tensors: Intermediate tensors from prior forward pass.

604 + inputs_embeds: Optional tensor of input embeddings.

605

606 """

607

vllm/model_executor/models/zamba2.py

+2 -2

@@ -909,8 +909,8 @@ class Zamba2ForCausalLM(nn.Module, HasInnerState, IsHybrid):

909 prefix: Optional prefix for parameter names

910

911 Raises:

912 - AssertionError: If prefix caching is enabled

913 - (not supported by Mamba)

912 + AssertionError: If prefix caching is enabled

913 + (not supported by Mamba)

914 """

915 config = vllm_config.model_config.hf_config

916 cache_config = vllm_config.cache_config

vllm/transformers_utils/config.py

+7 -6

@@ -679,20 +679,21 @@ def get_hf_file_to_dict(file_name: str,

679

680

681 @cache

682 -def get_pooling_config(model: str, revision: Optional[str] = 'main'):

682 +def get_pooling_config(model: str,

683 + revision: Optional[str] = 'main') -> Optional[dict]:

684 """

685 This function gets the pooling and normalize

686 config from the model - only applies to

687 sentence-transformers models.

688

689 Args:

689 - model (str): The name of the Hugging Face model.

690 - revision (str, optional): The specific version

691 - of the model to use. Defaults to 'main'.

690 + model: The name of the Hugging Face model.

691 + revision: The specific version of the model to use.

692 + Defaults to 'main'.

693

694 Returns:

694 - dict: A dictionary containing the pooling

695 - type and whether normalization is used.

695 + A dictionary containing the pooling type and whether

696 + normalization is used, or None if no pooling configuration is found.

697 """

698

699 modules_file_name = "modules.json"

vllm/transformers_utils/configs/jais.py

+4 -4

@@ -74,10 +74,10 @@ class JAISConfig(PretrainedConfig):

74 use_cache (`bool`, *optional*, defaults to `True`):

75 Whether or not the model should return the last key/values

76 attentions (not used by all models).

77 - scale_attn_by_inverse_layer_idx (`bool`, *optional*,

78 - defaults to `False`):

79 - Whether to additionally scale attention weights by

80 - `1 / layer_idx + 1`.

77 + scale_attn_by_inverse_layer_idx

78 + (`bool`, *optional*, defaults to `False`):

79 + Whether to additionally scale attention weights

80 + by `1 / layer_idx + 1`.

81 reorder_and_upcast_attn (`bool`, *optional*, defaults to `False`):

82 Whether to scale keys (K) prior to computing attention

83 (dot-product)

vllm/transformers_utils/configs/ultravox.py

-4

@@ -37,10 +37,6 @@ class UltravoxConfig(transformers.PretrainedConfig):

37 The initialization value for the layer normalization.

38 projector_act (`str`, *optional*, defaults to `"swiglu"`):

39 The activation function used by the multimodal projector.

40 - text_model_lora_config (`LoraConfigSimplified`, *optional*):

41 - The LoRA configuration for finetuning the text model.

42 - audio_model_lora_config (`LoraConfigSimplified`, *optional*):

43 - The LoRA configuration for finetuning the audio model.

40 projector_ln_mid (`bool`, *optional*, defaults to `False`):

41 Whether to apply layer normalization at the middle of the

42 projector or at the end. Versions v0.4.1 and below

vllm/transformers_utils/processors/deepseek_vl2.py

+4 -5

@@ -25,6 +25,7 @@

25 # CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

26

27 import math

28 +from typing import Any

29

30 import torch

31 import torchvision.transforms as T

@@ -178,17 +179,15 @@ class DeepseekVLV2Processor(ProcessorMixin):

179 prompt: str,

180 images: list[Image.Image],

181 inference_mode: bool = True,

181 - **kwargs,

182 + **kwargs: Any,

183 ):

184 """

185

186 Args:

187 prompt (str): the formatted prompt;

187 - conversations (list[dict]): conversations with a list of messages;

188 images (list[ImageType]): the list of images;

189 inference_mode (bool): if True, then remove the last eos token;

190 - system_prompt (str): the system prompt;

191 - **kwargs:

190 + **kwargs: Additional keyword arguments.

191

192 Returns:

193 outputs (BaseProcessorOutput): the output of the processor,

@@ -259,7 +258,7 @@ class DeepseekVLV2Processor(ProcessorMixin):

258 text: str,

259 images: list[Image.Image],

260 inference_mode: bool = True,

262 - **kwargs,

261 + **kwargs: Any,

262 ):

263 """

264

vllm/transformers_utils/runai_utils.py

+1 -3

@@ -33,7 +33,6 @@ def list_safetensors(path: str = "") -> list[str]:

33

34 Args:

35 path: The object storage path to list from.

36 - allow_pattern: A list of patterns of which files to pull.

36

37 Returns:

38 list[str]: List of full object storage paths allowed by the pattern

@@ -54,8 +53,7 @@ class ObjectStorageModel:

53 dir: The temporary created directory.

54

55 Methods:

57 - pull_files(): Pull model from object storage to the temporary

58 - directory.

56 + pull_files(): Pull model from object storage to the temporary directory.

57 """

58

59 def __init__(self) -> None:

vllm/transformers_utils/s3_utils.py

+3 -3

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

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

3

4 import fnmatch

5 -from typing import Optional

5 +from typing import Any, Optional

6

7 from vllm.utils import PlaceholderModule

8

@@ -26,7 +26,7 @@ def _filter_ignore(paths: list[str], patterns: list[str]) -> list[str]:

26 ]

27

28

29 -def glob(s3=None,

29 +def glob(s3: Optional[Any] = None,

30 path: str = "",

31 allow_pattern: Optional[list[str]] = None) -> list[str]:

32 """

@@ -51,7 +51,7 @@ def glob(s3=None,

51

52

53 def list_files(

54 - s3,

54 + s3: Any,

55 path: str,

56 allow_pattern: Optional[list[str]] = None,

57 ignore_pattern: Optional[list[str]] = None

[Docs] Fix warnings in mkdocs build (continued) (#24791)