d7e1e59972 · tduncan/Nvidia-Optimization-Stack-VLLM

tests/core/test_scheduler.py

+1 -1

@@ -641,7 +641,7 @@ def test_schedule_decode_blocks_to_copy_update():

641 # Nothing is preempted.

642 assert output.blocks_to_swap_out == []

643 # Since append_slot returns the source -> dist mapping, it should

644 - # applied.

644 + # be applied.

645 assert output.blocks_to_copy == [(2, 3)]

646

647

tests/entrypoints/openai/correctness/test_transcription_api_correctness.py

+1 -1

@@ -32,7 +32,7 @@ def to_bytes(y, sr):

32

33 async def transcribe_audio(client, tokenizer, y, sr):

34 # Send loaded audio directly instead of loading from disk,

35 - # dont account for that time though

35 + # don't account for that time though

36 with to_bytes(y, sr) as f:

37 start_time = time.perf_counter()

38 transcription = await client.audio.transcriptions.create(

tests/entrypoints/openai/test_return_token_ids.py

+1 -1

@@ -224,7 +224,7 @@ async def test_comparison_with_prompt_logprobs_and_logprobs(server):

224 logprobs_token_ids.append(token_id)

225

226 # When echo=True, the logprobs include both prompt and response tokens

227 - # The token_ids field should match the the suffix of response portion

227 + # The token_ids field should match the suffix of response portion

228 # The prompt_token_ids should match the prompt portion

229 assert len(completion.choices[0].token_ids) < len(logprobs_token_ids)

230 response_token_ids_length = len(completion.choices[0].token_ids)

tests/entrypoints/openai/test_serving_chat.py

+1 -1

@@ -313,7 +313,7 @@ async def test_serving_chat_did_set_correct_cache_salt(model_type):

313 }],

314 )

315

316 - # By default cache_salt in the engine prompt is not set

316 + # By default, cache_salt in the engine prompt is not set

317 with suppress(Exception):

318 await serving_chat.create_chat_completion(req)

319 assert "cache_salt" not in mock_engine.generate.call_args.args[0]

tests/kernels/utils.py

+1 -1

@@ -1236,7 +1236,7 @@ def baseline_scaled_mm(a: torch.Tensor,

1236 bias: Optional[torch.Tensor] = None) -> torch.Tensor:

1237

1238 # We treat N-dimensional group scaling as extended numpy-style broadcasting

1239 - # in numpy simply stretches dimensions with an extent of 1 to match the

1239 + # in numpy simply stretches dimensions with an extent of 1 to match

1240 # the target shape by repeating the data along that dimension (broadcasting)

1241 # , we extend these semantics to say if the extent of a dimension in the

1242 # source shape is not 1 and does not match the target shape we repeat each

tests/multimodal/test_utils.py

+2 -2

@@ -458,7 +458,7 @@ def run_dp_sharded_vision_model_vs_direct(local_rank: int, world_size: int,

458 with torch.inference_mode():

459 sharded_output = run_dp_sharded_vision_model(image_input, vision_model)

460

461 - # Check that the world size is setup correctly

461 + # Check that the world size is set up correctly

462 assert get_tensor_model_parallel_world_size() == world_size

463

464 # Check that the outputs have the same shape

@@ -642,7 +642,7 @@ def run_dp_sharded_mrope_vision_model_vs_direct(local_rank: int,

642 rope_type="rope_3d")

643 sharded_output = torch.cat(sharded_output, dim=0)

644

645 - # Check that the world size is setup correctly

645 + # Check that the world size is set up correctly

646 assert get_tensor_model_parallel_world_size() == world_size

647

648 # Compare outputs (only on rank 0)

tests/v1/e2e/test_spec_decode.py

+1 -1

@@ -83,7 +83,7 @@ def test_ngram_correctness(

83 model_name: str,

84 ):

85 '''

86 - Compare the outputs of a original LLM and a speculative LLM

86 + Compare the outputs of an original LLM and a speculative LLM

87 should be the same when using ngram speculative decoding.

88 '''

89 with monkeypatch.context() as m:

tests/v1/kv_connector/unit/test_remote_decode_lifecycle.py

+2 -2

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

42 engine_core_outputs = scheduler.update_from_output(scheduler_output,

43 model_runner_output)

44

45 - # Ensure the request is finished after 1 tokens.

45 + # Ensure the request is finished after 1 token.

46 assert request.is_finished()

47 assert request.status == RequestStatus.FINISHED_LENGTH_CAPPED

48 output = engine_core_outputs[0].outputs[0]

@@ -141,7 +141,7 @@ def test_short_prompt_lifecycle():

141

142

143 def test_prefix_cache_lifecycle():

144 - """Test that remote decode params still works with a prefix cache hit."""

144 + """Test that remote decode params still work with a prefix cache hit."""

145

146 vllm_config = create_vllm_config()

147 scheduler = create_scheduler(vllm_config)

tests/v1/spec_decode/test_tree_attention.py

+2 -2

@@ -187,7 +187,7 @@ def test_tree_attn_correctness() -> None:

187 dtype=torch.bfloat16,

188 )

189

190 - # Setup the block table and KV cache for paged KV.

190 + # Set up the block table and KV cache for paged KV.

191 assert max_sequence_length % block_size == 0

192 max_blocks_per_batch = max_sequence_length // block_size

193 kv_cache = torch.randn(

@@ -222,7 +222,7 @@ def test_tree_attn_correctness() -> None:

222 num_alloc_blocks_per_batch] = block_ids.view(

223 -1, num_alloc_blocks_per_batch)

224

225 - # Setup the slot mapping for the input KVs.

225 + # Set up the slot mapping for the input KVs.

226 tree_positions = sequence_position + torch.arange(

227 0,

228 tree_size_q,

vllm/lora/utils.py

+1 -1

@@ -239,7 +239,7 @@ def get_adapter_absolute_path(lora_path: str) -> str:

239 except (HfHubHTTPError, RepositoryNotFoundError, EntryNotFoundError,

240 HFValidationError):

241 # Handle errors that may occur during the download

242 - # Return original path instead instead of throwing error here

242 + # Return original path instead of throwing error here

243 logger.exception("Error downloading the HuggingFace model")

244 return lora_path

245

vllm/model_executor/layers/quantization/compressed_tensors/utils.py

+1 -1

@@ -94,7 +94,7 @@ def find_matched_target(

94 config that a layer corresponds to.

95

96 Recall that a compressed-tensors configs has a concept of

97 - config_groups, where each layer can be quantized with with a different

97 + config_groups, where each layer can be quantized with a different

98 scheme.

99

100 targets in each config_group will be a list of either layer names

vllm/multimodal/utils.py

+5 -5

@@ -213,7 +213,7 @@ class MediaConnector:

213 image_mode: str = "RGB",

214 ) -> Image.Image:

215 """

216 - Load a PIL image from a HTTP or base64 data URL.

216 + Load a PIL image from an HTTP or base64 data URL.

217

218 By default, the image is converted into RGB format.

219 """

@@ -237,7 +237,7 @@ class MediaConnector:

237 image_mode: str = "RGB",

238 ) -> Image.Image:

239 """

240 - Asynchronously load a PIL image from a HTTP or base64 data URL.

240 + Asynchronously load a PIL image from an HTTP or base64 data URL.

241

242 By default, the image is converted into RGB format.

243 """

@@ -261,7 +261,7 @@ class MediaConnector:

261 image_mode: str = "RGB",

262 ) -> tuple[npt.NDArray, dict[str, Any]]:

263 """

264 - Load video from a HTTP or base64 data URL.

264 + Load video from an HTTP or base64 data URL.

265 """

266 image_io = ImageMediaIO(image_mode=image_mode,

267 **self.media_io_kwargs.get("image", {}))

@@ -281,7 +281,7 @@ class MediaConnector:

281 image_mode: str = "RGB",

282 ) -> tuple[npt.NDArray, dict[str, Any]]:

283 """

284 - Asynchronously load video from a HTTP or base64 data URL.

284 + Asynchronously load video from an HTTP or base64 data URL.

285

286 By default, the image is converted into RGB format.

287 """

@@ -370,7 +370,7 @@ def group_mm_inputs_by_modality(

370

371 def modality_group_func(

372 mm_input: MultiModalKwargsItems) -> Union[str, int]:

373 - # If the input has multiple modalities, return a id as the unique key

373 + # If the input has multiple modalities, return an id as the unique key

374 # for the mm_input input.

375 if len(mm_input) > 1:

376 return id(mm_input)

vllm/v1/attention/backends/utils.py

+1 -1

@@ -709,7 +709,7 @@ def reorder_batch_to_split_decodes_and_prefills(

709

710 for i, req_id in enumerate(input_batch.req_ids):

711 num_tokens = scheduler_output.num_scheduled_tokens[req_id]

712 - # for now treat 1 scheduled token as "decode" even if its not,

712 + # for now treat 1 scheduled token as "decode" even if it's not,

713 # we should update this to something like < 8 in the future but

714 # currently the TritonMLA._forward_decode only supports

715 # num_tokens = 1

vllm/v1/structured_output/utils.py

+2 -2

@@ -65,9 +65,9 @@ def get_outlines_cache_path() -> str:

65 elif xdg_cache_home:

66 return os.path.join(xdg_cache_home, ".cache", "outlines")

67 # If homedir is "/", we may be inside a container, and thus writing to

68 - # root would be problematic, so we fallback to using a tempfile.

68 + # root would be problematic, so we fall back to using a tempfile.

69 # Also validate the path exists, since os.path.expanduser does

70 - # not garuntee existence.

70 + # not guarantee existence.

71 elif os.path.isdir(home_dir) and home_dir != "/":

72 # Default Unix fallback: ~/.cache/outlines

73 return os.path.join(home_dir, ".cache", "outlines")

vllm/v1/worker/tpu_worker.py

+1 -1

@@ -250,7 +250,7 @@ class TPUWorker:

250 scheduler_output: "SchedulerOutput",

251 ) -> Optional[ModelRunnerOutput]:

252 output = self.model_runner.execute_model(scheduler_output)

253 - # every worker's output is needed when kv_transfer_group is setup

253 + # every worker's output is needed when kv_transfer_group is set up

254 return output if self.is_driver_worker or has_kv_transfer_group(

255 ) else None

256

[Doc]: fix typos in Python comments (#24093)