@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 41ae4a1eab

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 various files (#24798)

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

Didier Durand committed Sep 13, 2025 at 09:43 UTC 41ae4a1eab22b332cd7d8f233f09611a5da53b1c

10 files changed +13 -13

examples/offline_inference/tpu.py

+1 -1

@@ -42,7 +42,7 @@ def main():

42 llm_args["model"] = "meta-llama/Llama-3.1-8B-Instruct"

44 # Set `enforce_eager=True` to avoid ahead-of-time compilation.

45 - # In real workloads, `enforace_eager` should be `False`.

45 + # In real workloads, `enforce_eager` should be `False`.

46 llm = LLM(**llm_args)

47 outputs = llm.generate(prompts, sampling_params)

48 print("-" * 50)

vllm/core/block/naive_block.py

+1 -1

@@ -182,7 +182,7 @@ class NaiveBlockAllocator(BlockAllocator):

182 # Increment refcount for each block.

183 assert block.block_id is not None

184 refcount = self._refcounter.incr(block.block_id)

185 - assert refcount != 1, "can't fork free'd block"

185 + assert refcount != 1, "can't fork freed block"

186

187 forked_block = self._block_pool.init_block(

188 prev_block=prev_block,

vllm/core/evictor.py

+1 -1

@@ -58,7 +58,7 @@ class Evictor(ABC):

59 class BlockMetaData:

60 """Data structure for storing key data describe cached block, so that

61 - evitor could use to make its decision which one to choose for eviction

61 + evictor could use to make its decision which one to choose for eviction

63 Here we use physical block id as the dict key, as there maybe several

64 blocks with the same content hash, but their physical id is unique.

vllm/engine/metrics.py

+2 -2

@@ -379,7 +379,7 @@ class LoggingStatLogger(StatLoggerBase):

379 if local_interval_elapsed(stats.now, self.last_local_log,

380 self.local_interval):

381 # Compute summary metrics for tracked stats (and log them

382 - # to promethus if applicable).

382 + # to prometheus if applicable).

383 prompt_throughput = get_throughput(self.num_prompt_tokens,

384 now=stats.now,

385 last_log=self.last_local_log)

@@ -432,7 +432,7 @@ class LoggingStatLogger(StatLoggerBase):

432

433

434 class PrometheusStatLogger(StatLoggerBase):

435 - """PrometheusStatLogger is used LLMEngine to log to Promethus."""

435 + """PrometheusStatLogger is used LLMEngine to log to Prometheus."""

436 _metrics_cls = Metrics

437 _gauge_cls = prometheus_client.Gauge

438

vllm/model_executor/layers/linear.py

+2 -2

@@ -740,7 +740,7 @@ class MergedColumnParallelLinear(ColumnParallelLinear):

740 """

741 Handle special case for models where MLP layers are already

742 fused on disk. In this case, we have no shard id. This function

743 - determmines the shard id by splitting these layers and then calls

743 + determines the shard id by splitting these layers and then calls

744 the weight loader using the shard id.

745

746 An example of a model with these fused layers:

@@ -914,7 +914,7 @@ class QKVParallelLinear(ColumnParallelLinear):

914 """

915 Handle special case for models where QKV layers are already

916 fused on disk. In this case, we have no shard id. This function

917 - determmines the shard id by splitting these layers and then calls

917 + determines the shard id by splitting these layers and then calls

918 the weight loader using the shard id.

919

920 An example of a model with these fused layers:

vllm/model_executor/layers/vocab_parallel_embedding.py

+1 -1

@@ -258,7 +258,7 @@ class VocabParallelEmbedding(CustomOp):

258

259 if params_dtype is None:

260 params_dtype = torch.get_default_dtype()

261 - # Divide the weight matrix along the vocaburaly dimension.

261 + # Divide the weight matrix along the vocabulary dimension.

262 self.num_added_embeddings = self.num_embeddings - self.org_vocab_size

263 self.num_embeddings_per_partition = divide(self.num_embeddings_padded,

264 self.tp_size)

vllm/model_executor/models/ernie45_vl.py

+1 -1

@@ -1446,7 +1446,7 @@ class Ernie4_5_VLMoeForConditionalGeneration(nn.Module, SupportsMultiModal,

1446 return None

1447

1448 # The result multimodal_embeddings is tuple of tensors, with each

1449 - # tensor correspoending to a multimodal data item (image or video).

1449 + # tensor corresponding to a multimodal data item (image or video).

1450 multimodal_embeddings: tuple[torch.Tensor, ...] = ()

1451

1452 # NOTE: It is important to iterate over the keys in this dictionary

vllm/model_executor/models/gemma3n_mm.py

+2 -2

@@ -586,10 +586,10 @@ class Gemma3nForConditionalGeneration(nn.Module, SupportsMultiModal,

586

587 # ruff: noqa

588 # The Gemma3nProcessor expects all audio will be 30s in length and inserts 188 audio soft tokens into the

589 - # text to account for this. However, the audio preprocessing and encoder do not gurarantee they will

589 + # text to account for this. However, the audio preprocessing and encoder do not guarantee they will

590 # produce 188 soft tokens; they will produce at most that many tokens, but they may produce fewer tokens

591 # depending on the length of the longest audio input in the batch. When we encounter this situation, we pad

592 - # the audio feature out to 188 soft tokens with the emebedding of the last token in the embed_audio vocab.

592 + # the audio feature out to 188 soft tokens with the embedding of the last token in the embed_audio vocab.

593 # TODO precompute and cache padding

594 audio_padding_toks = torch.tensor([[self.vocab_size - 1]],

595 dtype=torch.long,

vllm/model_executor/models/nemotron_vl.py

+1 -1

@@ -560,7 +560,7 @@ class LlamaNemotronVLChatModel(nn.Module, SupportsMultiModal, SupportsPP,

560 return []

561

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

563 - # tensor correspoending to a multimodal data item (image).

563 + # tensor corresponding to a multimodal data item (image).

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

565

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

vllm/model_executor/models/phi4mm.py

+1 -1

@@ -1154,7 +1154,7 @@ class Phi4MMForCausalLM(nn.Module, SupportsLoRA, SupportsMultiModal):

1154 return None

1155

1156 # The result multimodal_embeddings is tuple of tensors, with each

1157 - # tensor correspoending to a multimodal data item (image or video).

1157 + # tensor corresponding to a multimodal data item (image or video).

1158 multimodal_embeddings: tuple[torch.Tensor, ...] = ()

1159

1160 # NOTE: It is important to iterate over the keys in this dictionary