@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / e2b1f863aa

Fork 0

A high-throughput and memory-efficient inference and serving engine for LLMs

Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git

Synced Nov 14, 05:30 UTC

[Doc]: fixing doc typos (#24635)

Signed-off-by: Didier Durand <durand.didier@gmail.com>

Didier Durand committed Sep 11, 2025 at 08:19 UTC e2b1f863aa0ca94532624fd2add48d61f8446c2d

9 files changed +10 -10

vllm/config/__init__.py

+1 -1

@@ -3881,7 +3881,7 @@ def contains_object_print(text):

3881 Check if the text looks like a printed Python object, e.g.

3882 contains any substring matching the pattern: "at 0xFFFFFFF>"

3883 We match against 0x followed by 2-16 hex chars (there's

3884 - a max of 16 on a 64 bit system).

3884 + a max of 16 on a 64-bit system).

3885

3886 Args:

3887 text (str): The text to check

vllm/entrypoints/openai/tool_parsers/internlm2_tool_parser.py

+1 -1

@@ -60,7 +60,7 @@ class Internlm2ToolParser(ToolParser):

60 if '<|action_start|>' not in current_text:

61 self.position = len(current_text)

62 return DeltaMessage(content=delta_text)

63 - # if the tool call is sended, return an empty delta message

63 + # if the tool call is sent, return an empty delta message

64 # to make sure the finish_reason will be sent correctly.

65 if self.current_tool_id > 0:

66 return DeltaMessage(content='')

vllm/model_executor/layers/mamba/ops/ssd_chunk_state.py

+1 -1

@@ -502,7 +502,7 @@ def _chunk_state_varlen_kernel(

502 dA_cumsum_ptrs += BLOCK_SIZE_K * stride_dA_cs_csize

503

504 # If the sequence starts after the last chunk idx, we don't need to add the contribution from the last chunk

505 - # If HAS_INITSTATES==True need to consider two possiblties

505 + # If HAS_INITSTATES==True need to consider two possibilities

506 # - if start_idx < pid_c * chunk_size, then we need to take the past_states_ptrs

507 # - if state_idx >= pid * chunk_size, then we need to insert initstates

508 if ((start_idx < pid_c * chunk_size) # first chunk

vllm/model_executor/models/arcee.py

+1 -1

@@ -342,7 +342,7 @@ class ArceeModel(nn.Module):

342 class ArceeForCausalLM(nn.Module, SupportsLoRA, SupportsPP):

343 """Arcee Model for causal language modeling, integrated with vLLM

344 runtime."""

345 - # Map fused module names to their sub-module components

345 + # Map fused module names to their submodule components

346 # (for quantization and LoRA)

347 packed_modules_mapping = {

348 "qkv_proj": ["q_proj", "k_proj", "v_proj"],

vllm/model_executor/models/llava_onevision.py

+1 -1

@@ -835,7 +835,7 @@ class LlavaOnevisionForConditionalGeneration(nn.Module, SupportsMultiModal,

835 return None

836

837 # The result multimodal_embeddings is tuple of tensors, with each

838 - # tensor correspoending to a multimodal data item (image or video).

838 + # tensor corresponding to a multimodal data item (image or video).

839 multimodal_embeddings: tuple[torch.Tensor, ...] = ()

840

841 # NOTE: It is important to iterate over the keys in this dictionary

vllm/model_executor/models/phi4_multimodal.py

+1 -1

@@ -1350,7 +1350,7 @@ class Phi4MultimodalForCausalLM(nn.Module, SupportsLoRA, SupportsMultiModal):

1350 return None

1351

1352 # The result multimodal_embeddings is tuple of tensors, with each

1353 - # tensor correspoending to a multimodal data item (image or video).

1353 + # tensor corresponding to a multimodal data item (image or video).

1354 multimodal_embeddings: tuple[torch.Tensor, ...] = ()

1355

1356 # NOTE: It is important to iterate over the keys in this dictionary

vllm/model_executor/models/phi4mm_audio.py

+1 -1

@@ -100,7 +100,7 @@ class ConformerEncoderLayer(nn.Module):

100 activation function for glu used in the multihead attention,

101 default "swish".

102 activation_checkpointing: str, optional

103 - a dictionarry of {"module","interval","offload"}, where

103 + a dictionary of {"module","interval","offload"}, where

104 "module": str

105 accept ["transformer", "attention"] to select

106 which module should do activation checkpointing.

vllm/model_executor/models/qwen2_5_omni_thinker.py

+2 -2

@@ -846,7 +846,7 @@ class Qwen2_5OmniThinkerForConditionalGeneration(

846 return []

847

848 # The result multimodal_embeddings is tuple of tensors, with each

849 - # tensor correspoending to a multimodal data item (image or video).

849 + # tensor corresponding to a multimodal data item (image or video).

850 multimodal_embeddings: tuple[torch.Tensor, ...] = ()

851

852 # NOTE: It is important to iterate over the keys in this dictionary

@@ -873,7 +873,7 @@ class Qwen2_5OmniThinkerForConditionalGeneration(

873 if multimodal_embeddings is not None \

874 and len(multimodal_embeddings) != 0:

875

876 - # TODO (ywang96): support overlapping modalitiy embeddings so that

876 + # TODO (ywang96): support overlapping modality embeddings so that

877 # `use_audio_in_video` will work on V1.

878 inputs_embeds = merge_multimodal_embeddings(

879 input_ids, inputs_embeds, multimodal_embeddings, [

vllm/v1/attention/backends/mla/common.py

+1 -1

@@ -463,7 +463,7 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[M]):

463 self.dcp_world_size = 1

464 self.dcp_rank = 0

465

466 - # Dont try to access the runner on AMD

466 + # Don't try to access the runner on AMD

467 if self.aot_schedule:

468 self.page_size = self.kv_cache_spec.block_size

469