@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 9701352e4b

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 (#24001)

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

Didier Durand committed Aug 31, 2025 at 10:21 UTC 9701352e4ba75e89b5d35a5cb74d0a55567319d9

10 files changed +14 -14

vllm/compilation/monitor.py

+1 -1

@@ -43,7 +43,7 @@ cudagraph_capturing_enabled: bool = True

45 def validate_cudagraph_capturing_enabled():

46 - # used to monitor whether an cudagraph capturing is legal at runtime.

46 + # used to monitor whether a cudagraph capturing is legal at runtime.

47 # should be called before any cudagraph capturing.

48 # if an illegal cudagraph capturing happens, raise an error.

49 global cudagraph_capturing_enabled

vllm/core/evictor.py

+1 -1

@@ -76,7 +76,7 @@ class LRUEvictor(Evictor):

76 that's recorded in the Block. If there are multiple blocks with

77 the same last_accessed time, then the one with the largest num_hashed_tokens

78 will be evicted. If two blocks each have the lowest last_accessed time and

79 - highest num_hashed_tokens value, then one will be chose arbitrarily

79 + highest num_hashed_tokens value, then one will be chosen arbitrarily

80 """

82 # CLEANUP_THRESHOLD determines the maximum allowable size of the priority

vllm/engine/llm_engine.py

+1 -1

@@ -1239,7 +1239,7 @@ class LLMEngine:

1239

1240 # Stop the execute model loop in parallel workers until there are

1241 # more requests to process. This avoids waiting indefinitely in

1242 - # torch.distributed ops which may otherwise timeout, and unblocks

1242 + # torch.distributed ops which may otherwise time out, and unblocks

1243 # the RPC thread in the workers so that they can process any other

1244 # queued control plane messages, such as add/remove lora adapters.

1245 logger.debug("Stopping remote worker execution loop.")

vllm/entrypoints/llm.py

+5 -5

@@ -329,7 +329,7 @@ class LLM:

329 Args:

330 prompts: The prompts to the LLM. You may pass a sequence of prompts

331 for batch inference. See [PromptType][vllm.inputs.PromptType]

332 - for more details about the format of each prompts.

332 + for more details about the format of each prompt.

333 sampling_params: The sampling parameters for text generation. If

334 None, we use the default sampling parameters.

335 When it is a single value, it is applied to every prompt.

@@ -853,7 +853,7 @@ class LLM:

853 Args:

854 prompts: The prompts to the LLM. You may pass a sequence of prompts

855 for batch inference. See [PromptType][vllm.inputs.PromptType]

856 - for more details about the format of each prompts.

856 + for more details about the format of each prompt.

857 pooling_params: The pooling parameters for pooling. If None, we

858 use the default pooling parameters.

859 use_tqdm: If `True`, shows a tqdm progress bar.

@@ -946,7 +946,7 @@ class LLM:

946 Args:

947 prompts: The prompts to the LLM. You may pass a sequence of prompts

948 for batch inference. See [PromptType][vllm.inputs.PromptType]

949 - for more details about the format of each prompts.

949 + for more details about the format of each prompt.

950 pooling_params: The pooling parameters for pooling. If None, we

951 use the default pooling parameters.

952 use_tqdm: If `True`, shows a tqdm progress bar.

@@ -994,7 +994,7 @@ class LLM:

994 Args:

995 prompts: The prompts to the LLM. You may pass a sequence of prompts

996 for batch inference. See [PromptType][vllm.inputs.PromptType]

997 - for more details about the format of each prompts.

997 + for more details about the format of each prompt.

998 use_tqdm: If `True`, shows a tqdm progress bar.

999 If a callable (e.g., `functools.partial(tqdm, leave=False)`),

1000 it is used to create the progress bar.

@@ -1038,7 +1038,7 @@ class LLM:

1038 Args:

1039 prompts: The prompts to the LLM. You may pass a sequence of prompts

1040 for batch inference. See [PromptType][vllm.inputs.PromptType]

1041 - for more details about the format of each prompts.

1041 + for more details about the format of each prompt.

1042 use_tqdm: If `True`, shows a tqdm progress bar.

1043 If a callable (e.g., `functools.partial(tqdm, leave=False)`),

1044 it is used to create the progress bar.

vllm/executor/mp_distributed_executor.py

+1 -1

@@ -101,7 +101,7 @@ class MultiprocessingDistributedExecutor(DistributedExecutorBase):

101 result_handler.start()

102 self.worker_monitor.start()

103

104 - # Set up signal handlers to shutdown the executor cleanly

104 + # Set up signal handlers to shut down the executor cleanly

105 # sometimes gc does not work well

106

107 self.driver_worker = WorkerWrapperBase(self.vllm_config, 0)

vllm/lora/layers.py

+1 -1

@@ -605,7 +605,7 @@ class ColumnParallelLinearWithLoRA(BaseLinearLayerWithLoRA):

605

606 class MergedColumnParallelLinearWithLoRA(ColumnParallelLinearWithLoRA):

607 """ColumnParallelLinear layer that is composed of 2 sublayers (slices)

608 - packed together (eg. gate_proj + up_proj -> gate_up_proj).

608 + packed together (e.g. gate_proj + up_proj -> gate_up_proj).

609

610 This means we have 2 LoRAs, each applied to one half of the layer.

611

vllm/platforms/interface.py

+1 -1

@@ -537,7 +537,7 @@ class Platform:

537

538 def get_global_graph_pool(self) -> Any:

539 """

540 - Return the global graph pool for the this platform.

540 + Return the global graph pool for this platform.

541 """

542 cls = self.__class__

543 if cls._global_graph_pool is None:

vllm/reasoning/hunyuan_a13b_reasoning_parser.py

+1 -1

@@ -30,7 +30,7 @@ class HunyuanA13BReasoningParser(ReasoningParser):

30 Key Features:

31 - For non-stream output , Recognizes and extracts reasoning ("think")

32 and answer ("answer") sections from text using regular expressions.

33 - - For stream process, it require a token id sequences to change the

33 + - For stream process, it requires a token id sequences to change the

34 reasoning state and other state so it maintains internal state to

35 manage parsing across multiple token.

vllm/v1/worker/gpu_model_runner.py

+1 -1

@@ -2734,7 +2734,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):

2734 layer_names)

2735 attn_backends = {}

2736 attn_backend_layers = defaultdict(list)

2737 - # Dedupe based on full class name; this is a bit safer than using

2737 + # Dedupe based on full class name; this is a bit safer than

2738 # using the class itself as the key because when we create dynamic

2739 # attention backend subclasses (e.g. ChunkedLocalAttention) unless

2740 # they are cached correctly, there will be different objects per

vllm/v1/worker/gpu_worker.py

+1 -1

@@ -224,7 +224,7 @@ class Worker(WorkerBase):

224 memory can be used for KV cache without OOMs.

225

226 The engine will first conduct a profiling of the existing memory usage.

227 - Then, it calculate the free memory that can be used for KV cache in

227 + Then, it calculates the free memory that can be used for KV cache in

228 bytes.

229

230 Tip: