@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 83609ca91d

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 (#24173)

Signed-off-by: Didier Durand <durand.didier@gmail.com> Co-authored-by: Russell Bryant <rbryant@redhat.com> Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>

Didier Durand committed Sep 4, 2025 at 17:52 UTC 83609ca91d42c8847d1b4c272b011a0b6c27319e

12 files changed +13 -13

benchmarks/benchmark_dataset.py

+1 -1

@@ -403,7 +403,7 @@ class RandomDataset(BenchmarkDataset):

403 # [6880, 6881] -> ['Ġcalls', 'here'] ->

404 # [1650, 939, 486] -> ['Ġcall', 'sh', 'ere']

405 # To avoid uncontrolled change of the prompt length,

406 - # the encoded sequence is truncated before being decode again.

406 + # the encoded sequence is truncated before being decoded again.

407 total_input_len = prefix_len + int(input_lens[i])

408 re_encoded_sequence = tokenizer.encode(prompt, add_special_tokens=False)[

409 :total_input_len

benchmarks/kernels/benchmark_lora.py

+1 -1

@@ -637,7 +637,7 @@ def bench_optype(

637 # Clear LoRA optimization hash-maps.

638 _LORA_A_PTR_DICT.clear()

639 _LORA_B_PTR_DICT.clear()

640 - # Run bench function so that _LORA_A_PTR_DICT and _LORA_B_PTR_DICT are setup

640 + # Run bench function so that _LORA_A_PTR_DICT and _LORA_B_PTR_DICT are set up

641 for kwargs in kwargs_list:

642 op_type.bench_fn()(**kwargs)

643 torch.cuda.synchronize()

benchmarks/multi_turn/benchmark_serving_multi_turn.py

+1 -1

@@ -962,7 +962,7 @@ async def main_mp(

962

963 # At this point all the clients finished,

964 # collect results (TTFT, TPOT, etc.) from all the clients.

965 - # This needs to happens before calling join on the clients

965 + # This needs to happen before calling join on the clients

966 # (result_queue should be emptied).

967 while not result_queue.empty():

968 client_metrics.append(result_queue.get())

examples/offline_inference/audio_language.py

+1 -1

@@ -117,7 +117,7 @@ def run_gemma3n(question: str, audio_count: int) -> ModelRequestData:

117

118 # Granite Speech

119 def run_granite_speech(question: str, audio_count: int) -> ModelRequestData:

120 - # NOTE - the setting in this example are somehat different than what is

120 + # NOTE - the setting in this example are somewhat different from what is

121 # optimal for granite speech, and it is generally recommended to use beam

122 # search. Check the model README for suggested settings.

123 # https://huggingface.co/ibm-granite/granite-speech-3.3-8b

tests/models/multimodal/generation/vlm_utils/builders.py

+1 -1

@@ -250,7 +250,7 @@ def build_video_inputs_from_test_info(

250

251 def apply_image_size_scaling(image, size: Union[float, tuple[int, int]],

252 size_type: SizeType):

253 - """Applies a size scaler to one image; this can be a an image size factor,

253 + """Applies a size scaler to one image; this can be an image size factor,

254 which scales the image while maintaining the aspect ratio"""

255 # Special case for embeddings; if it's a tensor, it's only valid if we

256 # are considering size factors at constant scale, i.e., we just clone

tests/models/multimodal/generation/vlm_utils/case_filtering.py

+1 -1

@@ -42,7 +42,7 @@ def get_filtered_test_settings(

42 else:

43 assert test_info.prompt_formatter is not None

45 - # Everything looks okay; keep if this is has correct proc handling

45 + # Everything looks okay; keep if this is correct proc handling

46 if (test_info.distributed_executor_backend

47 is not None) == new_proc_per_test:

48 matching_tests[test_name] = test_info

vllm/attention/backends/mla/common.py

+1 -1

@@ -822,7 +822,7 @@ class MLACommonMetadataBuilder(AttentionMetadataBuilder[T], Generic[T]):

822 and context_lens_tensor is not None \

823 and context_lens_tensor[:self.num_prefills].max() > 0:

824

825 - # NOTE: it is recommend you read the `Chunked Prefill` section in

825 + # NOTE: it is recommended you read the `Chunked Prefill` section in

826 # the comment at the top of the file before trying to understand

827 # the following code

828

vllm/engine/async_llm_engine.py

+1 -1

@@ -717,7 +717,7 @@ class AsyncLLMEngine(EngineClient):

717 # Stop the execute model loop in parallel workers until there

718 # are more requests to process. This avoids waiting

719 # indefinitely in torch.distributed ops which may otherwise

720 - # timeout, and unblocks the RPC thread in the workers so that

720 + # time out, and unblocks the RPC thread in the workers so that

721 # they can process any other queued control plane messages,

722 # such as add/remove lora adapters.

723 await engine.engine.stop_remote_worker_execution_loop_async()

vllm/engine/multiprocessing/client.py

+1 -1

@@ -270,7 +270,7 @@ class MQLLMEngineClient(EngineClient):

270 queue.put_nowait(request_output)

271

272 async def setup(self):

273 - """Setup the client before it starts sending server requests."""

273 + """Set up the client before it starts sending server requests."""

274

275 # Start output_loop

276 if self.output_loop is None:

vllm/model_executor/layers/quantization/awq_triton.py

+1 -1

@@ -19,7 +19,7 @@ def awq_dequantize_kernel(

19 num_rows, # input num rows in qweight

20 BLOCK_SIZE_X: tl.constexpr,

21 BLOCK_SIZE_Y: tl.constexpr):

22 - # Setup the pids.

22 + # Set up the pids.

23 pid_x = tl.program_id(axis=0)

24 pid_y = tl.program_id(axis=1)

vllm/model_executor/layers/quantization/base_config.py

+1 -1

@@ -128,7 +128,7 @@ class QuantizationConfig(ABC):

128 @staticmethod

129 def get_from_keys_or(config: dict[str, Any], keys: list[str],

130 default: Any) -> Any:

131 - """Get a optional value from the model's quantization config."""

131 + """Get an optional value from the model's quantization config."""

132 try:

133 return QuantizationConfig.get_from_keys(config, keys)

134 except ValueError:

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

+2 -2

@@ -401,7 +401,7 @@ M = TypeVar("M", bound=MLACommonMetadata)

401

402

403 def use_flashinfer_prefill() -> bool:

404 - # For blackwell default to flashinfer prefill if its available since

404 + # For blackwell default to flashinfer prefill if it's available since

405 # it is faster than FA2.

406 return (flashinfer_available and not envs.VLLM_USE_CUDNN_PREFILL

407 and current_platform.is_device_capability(100))

@@ -1018,7 +1018,7 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):

1018 return layer.weight

1019

1020 # we currently do not have quantized bmm's which are needed for

1021 - # `W_UV` and `W_UK_T`, we we just store fp16/bf16 copies and perform

1021 + # `W_UV` and `W_UK_T`, we just store fp16/bf16 copies and perform

1022 # the bmm's in 16-bit, the extra memory overhead of this is fairly low

1023 kv_b_proj_weight = get_and_maybe_dequant_weights(self.kv_b_proj).T

1024 assert kv_b_proj_weight.shape == (