2eb673a088 · tduncan/Nvidia-Optimization-Stack-VLLM

.buildkite/performance-benchmarks/scripts/convert-results-json-to-markdown.py

+1 -1

@@ -393,7 +393,7 @@ if __name__ == "__main__":

393 with open(results_folder / md_file, "w") as f:

394 results = read_markdown(

395 "../.buildkite/performance-benchmarks/"

396 - + "performance-benchmarks-descriptions.md"

396 + "performance-benchmarks-descriptions.md"

397 )

398 results = results.format(

399 latency_tests_markdown_table=latency_md_table,

csrc/quantization/machete/generate.py

+3 -3

@@ -288,8 +288,8 @@ def generate_sch_sig(schedule_config: ScheduleConfig) -> str:

288 )

289 cluster_shape = (

290 f"{schedule_config.cluster_shape_mnk[0]}"

291 - + f"x{schedule_config.cluster_shape_mnk[1]}"

292 - + f"x{schedule_config.cluster_shape_mnk[2]}"

291 + f"x{schedule_config.cluster_shape_mnk[1]}"

292 + f"x{schedule_config.cluster_shape_mnk[2]}"

293 )

294 kernel_schedule = VLLMKernelScheduleTag[schedule_config.kernel_schedule].split(

295 "::"

@@ -301,7 +301,7 @@ def generate_sch_sig(schedule_config: ScheduleConfig) -> str:

301

302 return (

303 f"{tile_shape}_{cluster_shape}_{kernel_schedule}"

304 - + f"_{epilogue_schedule}_{tile_scheduler}"

304 + f"_{epilogue_schedule}_{tile_scheduler}"

305 )

306

307

examples/offline_inference/automatic_prefix_caching.py

+1 -1

@@ -26,7 +26,7 @@ from vllm import LLM, SamplingParams

26 # A prompt containing a large markdown table. The table is randomly generated by GPT-4.

27 LONG_PROMPT = (

28 "You are a helpful assistant in recognizes the content of tables in markdown format. Here is a table as follows.\n# Table\n"

29 - + """

29 + """

31 |-----|---------------|-----|---------------|---------------|------------------------|----------------|------------------------------|

32 | 1 | John Doe | 29 | Engineer | USA | john.doe@example.com | 555-1234 | 123 Elm St, Springfield, IL |

examples/others/lmcache/disagg_prefill_lmcache_v1/disagg_proxy_server.py

+4 -4

@@ -69,10 +69,10 @@ class StatsCalculator:

69 np_arr = np.array(self._stats)

70 output_str = (

71 f"\nNum requests: {len(self._stats)}"

72 - + "\nPrefill node TTFT stats:"

73 - + f"\n - Average (ms): {np.mean(np_arr)}"

74 - + f"\n - Median (ms): {np.median(np_arr)}"

75 - + f"\n - 99th Percentile (ms): {np.percentile(np_arr, 99)}\n"

72 + "\nPrefill node TTFT stats:"

73 + f"\n - Average (ms): {np.mean(np_arr)}"

74 + f"\n - Median (ms): {np.median(np_arr)}"

75 + f"\n - 99th Percentile (ms): {np.percentile(np_arr, 99)}\n"

76 )

77 print(

78 "===============================",

pyproject.toml

+2

@@ -72,6 +72,8 @@ select = [

72 "UP",

73 # flake8-bugbear

74 "B",

75 + # flake8-implicit-str-concat

76 + "ISC",

77 # flake8-simplify

78 "SIM",

79 # isort

tests/entrypoints/openai/tool_parsers/test_llama4_pythonic_tool_parser.py

+1 -1

@@ -193,7 +193,7 @@ TEST_CASES = [

193 pytest.param(

194 False,

195 "<|python_start|>[get_weather(city='LA', metric='C'), "

196 - + "register_user(name='Doe', age=9)]",

196 + "register_user(name='Doe', age=9)]",

197 [

198 SIMPLE_FUNCTION_CALL,

199 FunctionCall(name="register_user", arguments='{"name": "Doe", "age": 9}'),

tests/tool_parsers/test_deepseekv31_tool_parser.py

+8 -8

@@ -24,9 +24,9 @@ def parser(deepseekv31_tokenizer):

24 def test_extract_tool_calls_with_tool(parser):

25 model_output = (

26 "normal text"

27 - + "<｜tool▁calls▁begin｜>"

28 - + '<｜tool▁call▁begin｜>foo<｜tool▁sep｜>{"x":1}<｜tool▁call▁end｜>'

29 - + "<｜tool▁calls▁end｜>"

27 + "<｜tool▁calls▁begin｜>"

28 + '<｜tool▁call▁begin｜>foo<｜tool▁sep｜>{"x":1}<｜tool▁call▁end｜>'

29 + "<｜tool▁calls▁end｜>"

30 )

31 result = parser.extract_tool_calls(model_output, None)

32 assert result.tools_called

@@ -39,11 +39,11 @@ def test_extract_tool_calls_with_tool(parser):

39 def test_extract_tool_calls_with_multiple_tools(parser):

40 model_output = (

41 "some prefix text"

42 - + "<｜tool▁calls▁begin｜>"

43 - + '<｜tool▁call▁begin｜>foo<｜tool▁sep｜>{"x":1}<｜tool▁call▁end｜>'

44 - + '<｜tool▁call▁begin｜>bar<｜tool▁sep｜>{"y":2}<｜tool▁call▁end｜>'

45 - + "<｜tool▁calls▁end｜>"

46 - + " some suffix text"

42 + "<｜tool▁calls▁begin｜>"

43 + '<｜tool▁call▁begin｜>foo<｜tool▁sep｜>{"x":1}<｜tool▁call▁end｜>'

44 + '<｜tool▁call▁begin｜>bar<｜tool▁sep｜>{"y":2}<｜tool▁call▁end｜>'

45 + "<｜tool▁calls▁end｜>"

46 + " some suffix text"

47 )

48

49 result = parser.extract_tool_calls(model_output, None)

tests/utils.py

+1 -1

@@ -1302,7 +1302,7 @@ def prep_prompts(batch_size: int, ln_range: tuple[int, int] = (800, 1100)):

1302 indices.append(idx)

1303 prompt = (

1304 "```python\n# We set a number of variables, "

1305 - + f"x{idx} will be important later\n"

1305 + f"x{idx} will be important later\n"

1306 )

1307 ln = random.randint(*ln_range)

1308 for k in range(30, ln):

vllm/benchmarks/datasets.py

+1 -2

@@ -134,8 +134,7 @@ class BenchmarkDataset(ABC):

134 content.append(mm_content)

135 else:

136 raise TypeError(

137 - "Could not process multimodal content of type: "

138 - + f"{type(mm_content)}"

137 + f"Could not process multimodal content of type: {type(mm_content)}"

138 )

139 return [{"role": "user", "content": content}]

140

vllm/compilation/wrapper.py

+2 -2

@@ -208,8 +208,8 @@ class TorchCompileWithNoGuardsWrapper:

208 if not hasattr(self._compiled_callable, "aot_compile"):

209 raise RuntimeError(

210 "aot_compile is not supported by the current configuration. "

211 - + "Please make sure torch.compile is enabled with the latest "

212 - + f"version of PyTorch (current using torch: {torch.__version__})"

211 + "Please make sure torch.compile is enabled with the latest "

212 + f"version of PyTorch (current using torch: {torch.__version__})"

213 )

214 return self._compiled_callable.aot_compile((args, kwargs))

215

vllm/entrypoints/openai/translations/speech_to_text.py

+2 -2

@@ -406,8 +406,8 @@ class OpenAISpeechToText(OpenAIServing):

406

407 if request.response_format not in ["text", "json", "verbose_json"]:

408 return self.create_error_response(

409 - ("Currently only support response_format")

410 - + ("`text`, `json` or `verbose_json`")

409 + "Currently only support response_format: "

410 + "`text`, `json` or `verbose_json`"

411 )

412

413 if (

vllm/model_executor/layers/quantization/kernels/scaled_mm/aiter.py

+4 -4

@@ -32,8 +32,8 @@ class AiterInt8ScaledMMLinearKernel(CutlassInt8ScaledMMLinearKernel):

32 return (

33 False,

34 "requires setting `VLLM_ROCM_USE_AITER=1` "

35 - + "and `VLLM_ROCM_USE_AITER_LINEAR=1`. "

36 - + "`VLLM_ROCM_USE_AITER_LINEAR` default is True.",

35 + "and `VLLM_ROCM_USE_AITER_LINEAR=1`. "

36 + "`VLLM_ROCM_USE_AITER_LINEAR` default is True.",

37 )

38 return True, None

39

@@ -97,9 +97,9 @@ class AiterInt8ScaledMMLinearKernel(CutlassInt8ScaledMMLinearKernel):

97 per_token_scale_a and per_channel_scale_b

98 ), (

99 "Currently only support per-tensor-per-tensor GEMM "

100 - + " and per-token-per-channel GEMM through AITER"

100 + " and per-token-per-channel GEMM through AITER"

101 " w8a8 scaled gemm. `AiterInt8ScaledMMLinearKernel` "

102 - + "does not support AITER block scaled GEMM."

102 + "does not support AITER block scaled GEMM."

103 )

104

105 # gemm_a8w8_CK(a, b, scale_a, scale_b, bias) expects

vllm/reasoning/olmo3_reasoning_parser.py

+1 -1

@@ -234,7 +234,7 @@ class Olmo3ReasoningParser(ReasoningParser):

234 # reasoning template.

235 reasoning_expr = (

236 rf"^(?:{self.think_start})?(?P<reasoning>.*?)"

237 - + rf"{self.think_end}(?P<content>.*)$"

237 + rf"{self.think_end}(?P<content>.*)$"

238 )

239 self.reasoning_regex = re.compile(reasoning_expr, re.DOTALL)

240

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

+1 -1

@@ -216,7 +216,7 @@ class FlashAttnMLAMetadataBuilder(MLACommonMetadataBuilder[FlashAttnMLAMetadata]

216 # Ensure the persistent buffer is large enough

217 assert n <= self.scheduler_metadata.shape[0], (

218 f"Scheduler metadata size {n} exceeds buffer size "

219 - + f"{self.scheduler_metadata.shape[0]}"

219 + f"{self.scheduler_metadata.shape[0]}"

220 )

221 self.scheduler_metadata[:n] = scheduler_metadata

222 # NOTE(woosuk): We should zero out the rest of the scheduler

vllm/v1/core/single_type_kv_cache_manager.py

+1 -1

@@ -646,7 +646,7 @@ class ChunkedLocalAttentionManager(SingleTypeKVCacheManager):

646 """

647 assert isinstance(kv_cache_spec, ChunkedLocalAttentionSpec), (

648 "ChunkedLocalAttentionManager can only be used for "

649 - + "chunked local attention groups"

649 + "chunked local attention groups"

650 )

651 assert use_eagle is False, (

652 "Hybrid KV cache is not supported for " + "eagle + chunked local attention."

vllm/v1/engine/async_llm.py

+1 -1

@@ -618,7 +618,7 @@ class AsyncLLM(EngineClient):

618 except Exception as e2:

619 s = (

620 f"{e.__class__.__name__}: "

621 - + "error during printing an exception of class"

621 + "error during printing an exception of class"

622 + e2.__class__.__name__

623 )

624 logger.info("Request %s failed due to %s.", request_id, s)

Add flake8-implicit-str-concat rules to Ruff (#33191)