@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 107284959a

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]: fix typos in Python comments (#24026)

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

Didier Durand committed Sep 1, 2025 at 11:38 UTC 107284959ac14d42aba09a03220ff21da1eafcb2

14 files changed +17 -17

examples/offline_inference/multilora_inference.py

+1 -1

@@ -23,7 +23,7 @@ def create_test_prompts(

23 2 requests for base model, 4 requests for the LoRA. We define 2

24 different LoRA adapters (using the same model for demo purposes).

25 Since we also set `max_loras=1`, the expectation is that the requests

26 - with the second LoRA adapter will be ran after all requests with the

26 + with the second LoRA adapter will be run after all requests with the

27 first adapter have finished.

28 """

29 return [

vllm/distributed/device_communicators/pynccl.py

+1 -1

@@ -31,7 +31,7 @@ class PyNcclCommunicator:

31 group: the process group to work on. If None, it will use the

32 default process group.

33 device: the device to bind the PyNcclCommunicator to. If None,

34 - it will be bind to f"cuda:{local_rank}".

34 + it will be bound to f"cuda:{local_rank}".

35 library_path: the path to the NCCL library. If None, it will

36 use the default library path.

37 It is the caller's responsibility to make sure each communicator

vllm/distributed/parallel_state.py

+2 -2

@@ -939,8 +939,8 @@ def get_pipeline_model_parallel_group():

939 def graph_capture(device: torch.device):

940 """

941 `graph_capture` is a context manager which should surround the code that

942 - is capturing the CUDA graph. Its main purpose is to ensure that the

943 - some operations will be run after the graph is captured, before the graph

942 + is capturing the CUDA graph. Its main purpose is to ensure that some

943 + operations will be run after the graph is captured, before the graph

944 is replayed. It returns a `GraphCaptureContext` object which contains the

945 necessary data for the graph capture. Currently, it only contains the

946 stream that the graph capture is running on. This stream is set to the

vllm/entrypoints/openai/tool_parsers/pythonic_tool_parser.py

+1 -1

@@ -165,7 +165,7 @@ class PythonicToolParser(ToolParser):

165 index] += delta.function.arguments

166

167 # HACK: serving_chat.py inspects the internal state of tool parsers

168 - # when determining it's final streaming delta, automatically

168 + # when determining its final streaming delta, automatically

169 # adding autocompleted JSON.

170 # These two lines avoid that nonsense while ensuring finish_reason

171 # is set to tool_calls when at least one tool is called.

vllm/model_executor/layers/fused_moe/moe_pallas.py

+1 -1

@@ -7,7 +7,7 @@ import torch.nn.functional as F

8 def _histogram(input: torch.Tensor, min: int, max: int) -> torch.Tensor:

9 """

10 - Compute the histogram of a int32 tensor. The bin edges are defined by the

10 + Compute the histogram of an int32 tensor. The bin edges are defined by the

11 min and max values, with step = 1.

12 """

13 assert input.dtype == torch.int32, "input must be of torch.int32 dtype."

vllm/model_executor/models/ovis.py

+1 -1

@@ -544,7 +544,7 @@ class Ovis(nn.Module, SupportsMultiModal, SupportsPP):

544 vision_embeddings)

545 input_ids = None

546

547 - # up until here we have a inputs_embeds 100% numerical identity

547 + # up until here we have an inputs_embeds 100% numerical identity

548 # between the OG HF Transformers implementation and ours

549 hidden_states = self.llm(

550 input_ids=input_ids,

vllm/model_executor/models/phi4mm_audio.py

+3 -3

@@ -43,7 +43,7 @@ class ConformerEncoderLayer(nn.Module):

43 if set different to 0, the number of

44 depthwise_seperable_out_channel will be used as a

45 channel_out of the second conv1d layer.

46 - otherwise, it equal to 0, the second conv1d layer is skipped.

46 + otherwise, it equals to 0, the second conv1d layer is skipped.

47 depthwise_multiplier: int

48 number of input_dim channels duplication. this value

49 will be used to compute the hidden channels of the Conv1D.

@@ -115,7 +115,7 @@ class ConformerEncoderLayer(nn.Module):

115 we recalculate activation in backward.

116 default "".

117 export: bool, optional

118 - if set to True, it remove the padding from convolutional layers

118 + if set to True, it removes the padding from convolutional layers

119 and allow the onnx conversion for inference.

120 default False.

121 use_pt_scaled_dot_product_attention: bool, optional

@@ -686,7 +686,7 @@ class ConformerEncoder(TransformerEncoderBase):

686 only work for glu_in_attention !=0

687 default "swish".

688 export: bool, optional

689 - if set to True, it remove the padding from convolutional layers

689 + if set to True, it removes the padding from convolutional layers

690 and allow the onnx conversion for inference.

691 default False.

692 activation_checkpointing: str, optional

vllm/model_executor/models/phi4mm_utils.py

+1 -1

@@ -258,7 +258,7 @@ class DepthWiseSeperableConv1d(nn.Module):

258 if set different to 0, the number of

259 depthwise_seperable_out_channel will be used as a channel_out

260 of the second conv1d layer.

261 - otherwise, it equal to 0, the second conv1d layer is skipped.

261 + otherwise, it equals to 0, the second conv1d layer is skipped.

262 kernel_size: int

263 kernel_size

264 depthwise_multiplier: int

vllm/third_party/pynvml.py

+1 -1

@@ -1022,7 +1022,7 @@ def _extractNVMLErrorsAsClasses():

1022 Each NVML Error gets a new NVMLError subclass. This way try,except blocks can filter appropriate

1023 exceptions more easily.

1024

1025 - NVMLError is a parent class. Each NVML_ERROR_* gets it's own subclass.

1025 + NVMLError is a parent class. Each NVML_ERROR_* gets its own subclass.

1026 e.g. NVML_ERROR_ALREADY_INITIALIZED will be turned into NVMLError_AlreadyInitialized

1027 '''

1028 this_module = sys.modules[__name__]

vllm/transformers_utils/configs/nemotron.py

+1 -1

@@ -26,7 +26,7 @@ logger = logging.get_logger(__name__)

26 class NemotronConfig(PretrainedConfig):

27 r"""

28 This is the configuration class to store the configuration of a

29 - [`NemotronModel`]. It is used to instantiate an Nemotron model

29 + [`NemotronModel`]. It is used to instantiate a Nemotron model

30 according to the specified arguments, defining the model architecture.

31 Instantiating a configuration with the defaults will yield a similar

32 configuration to that of the Nemotron-8B.

vllm/transformers_utils/configs/nemotron_h.py

+1 -1

@@ -38,7 +38,7 @@ class NemotronHConfig(PretrainedConfig):

38 passed when calling [`NemotronHModel`]

39 tie_word_embeddings (`bool`, *optional*, defaults to `False`):

40 Whether the model's input and output word embeddings should be

41 - tied. Note that this is only relevant if the model has a output

41 + tied. Note that this is only relevant if the model has an output

42 word embedding layer.

43 hidden_size (`int`, *optional*, defaults to 4096):

44 Dimension of the hidden representations.

vllm/transformers_utils/processors/ovis.py

+1 -1

@@ -55,7 +55,7 @@ class OvisProcessorKwargs(ProcessingKwargs, total=False): # type: ignore[call-

56 class OvisProcessor(ProcessorMixin):

57 r"""

58 - Constructs a Ovis processor which wraps a Ovis image processor and a Qwen2 tokenizer into a single processor.

58 + Constructs an Ovis processor which wraps an Ovis image processor and a Qwen2 tokenizer into a single processor.

59 [`OvisProcessor`] offers all the functionalities of [`Qwen2VLImageProcessor`] and [`Qwen2TokenizerFast`]. See the

60 [`~OvisProcessor.__call__`] and [`~OvisProcessor.decode`] for more information.

61 Args:

vllm/transformers_utils/processors/ovis2_5.py

+1 -1

@@ -41,7 +41,7 @@ class Ovis2_5ProcessorKwargs(ProcessingKwargs,

42 class Ovis2_5Processor(ProcessorMixin):

43 r"""

44 - Constructs a Ovis processor which wraps a Ovis image processor

44 + Constructs an Ovis processor which wraps an Ovis image processor

45 and a Qwen2 tokenizer into a single processor.

46 [`OvisProcessor`] offers all the functionalities of

47 [`Qwen2VLImageProcessor`] and [`Qwen2TokenizerFast`].

vllm/v1/spec_decode/ngram_proposer.py

+1 -1

@@ -107,7 +107,7 @@ def _find_longest_matched_ngram_and_propose_tokens(

107 longest_ngram = 0

108 position = 0

109

110 - # lps[0] always equal to 0, we starts with index 1

110 + # lps[0] always equal to 0, we start with index 1

111 prev_lps = 0

112 i = 1

113 while i < total_token: