3b352a2f92 · tduncan/Nvidia-Optimization-Stack-VLLM

benchmarks/kernels/benchmark_rmsnorm.py

+1 -1

@@ -139,7 +139,7 @@ def calculate_diff(batch_size, seq_len, hidden_size, use_residual=True):

139

140 print(f"Naive output={output_naive}")

141 print(f"FlashInfer output={output_flashinfer}")

142 - print(f"VLLM output={output_vllm}")

142 + print(f"vLLM output={output_vllm}")

143

144 if torch.allclose(output_naive, output_flashinfer, atol=1e-2,

145 rtol=1e-2) and torch.allclose(

docs/source/contributing/vulnerability_management.md

+1 -1

@@ -37,7 +37,7 @@ you may contact the following individuals:

37

38 ## Slack Discussion

39

40 -You may use the `#security` channel in the [VLLM Slack](https://slack.vllm.ai)

40 +You may use the `#security` channel in the [vLLM Slack](https://slack.vllm.ai)

41 to discuss security-related topics. However, please do not disclose any

42 vulnerabilities in this channel. If you need to report a vulnerability, please

43 use the GitHub security advisory system or contact a VMT member privately.

docs/source/design/v1/metrics.md

+1 -1

@@ -509,7 +509,7 @@ cache to complete other requests), we swap kv cache blocks out to CPU

509 memory. This is also known as "KV cache offloading" and is configured

510 with `--swap-space` and `--preemption-mode`.

511

512 -In v0, [VLLM has long supported beam

512 +In v0, [vLLM has long supported beam

513 search](gh-issue:6226). The

514 SequenceGroup encapsulated the idea of N Sequences which

515 all shared the same prompt kv blocks. This enabled KV cache block

examples/offline_inference/disaggregated_prefill_lmcache.py

+1 -1

@@ -5,7 +5,7 @@ with LMCache.

5 We will launch 2 vllm instances (GPU 0 for prefill and GPU 1 for decode),

6 and launch an additional LMCache server.

7 KV cache is transferred in the following manner:

8 -VLLM prefill node -> LMCache server -> VLLM decode node.

8 +vLLM prefill node -> LMCache server -> vLLM decode node.

9

10 Note that `pip install lmcache` is needed to run this example.

11 Learn more about LMCache in https://github.com/LMCache/LMCache.

tests/tpu/test_quantization_accuracy.py

+1 -1

@@ -25,7 +25,7 @@ ACCURACY_CONFIGS = [

25 GSM8KAccuracyTestConfig(

26 model_name="neuralmagic/Meta-Llama-3.1-8B-Instruct-quantized.w8a8",

27 excepted_value=0.76), # no bias

28 - # NOTE(rob): We cannot re-initialize VLLM in the same process for TPU,

28 + # NOTE(rob): We cannot re-initialize vLLM in the same process for TPU,

29 # so only one of these tests can run in a single call to pytest. As

30 # a follow up, move this into the LM-EVAL section of the CI.

31 # GSM8KAccuracyTestConfig(

vllm/attention/selector.py

+1 -1

@@ -51,7 +51,7 @@ def get_env_variable_attn_backend() -> Optional[_Backend]:

51 # (default behavior if this variable is None)

52 #

53 # THIS SELECTION TAKES PRECEDENCE OVER THE

54 -# VLLM ATTENTION BACKEND ENVIRONMENT VARIABLE

54 +# VLLM_ATTENTION_BACKEND ENVIRONMENT VARIABLE

55 forced_attn_backend: Optional[_Backend] = None

56

57

vllm/compilation/backends.py

+1 -1

@@ -278,7 +278,7 @@ class PiecewiseCompileInterpreter(torch.fx.Interpreter):

278

279

280 class VllmBackend:

281 - """The compilation backend for `torch.compile` with VLLM.

281 + """The compilation backend for `torch.compile` with vLLM.

282 It is used for compilation level of `CompilationLevel.PIECEWISE`,

283 where we customize the compilation.

284

vllm/compilation/compiler_interface.py

+1 -1

@@ -31,7 +31,7 @@ class CompilerInterface:

31

32 def compute_hash(self, vllm_config: VllmConfig) -> str:

33 """

34 - Gather all the relevant information from the VLLM config,

34 + Gather all the relevant information from the vLLM config,

35 to compute a hash so that we can cache the compiled model.

36

37 See :meth:`VllmConfig.compute_hash` to check what information

vllm/config.py

+4 -4

@@ -3572,11 +3572,11 @@ _current_vllm_config: Optional[VllmConfig] = None

3572 @contextmanager

3573 def set_current_vllm_config(vllm_config: VllmConfig, check_compile=False):

3574 """

3575 - Temporarily set the current VLLM config.

3575 + Temporarily set the current vLLM config.

3576 Used during model initialization.

3577 - We save the current VLLM config in a global variable,

3577 + We save the current vLLM config in a global variable,

3578 so that all modules can access it, e.g. custom ops

3579 - can access the VLLM config to determine how to dispatch.

3579 + can access the vLLM config to determine how to dispatch.

3580 """

3581 global _current_vllm_config

3582 old_vllm_config = _current_vllm_config

@@ -3611,7 +3611,7 @@ def get_current_vllm_config() -> VllmConfig:

3611 # in ci, usually when we test custom ops/modules directly,

3612 # we don't set the vllm config. In that case, we set a default

3613 # config.

3614 - logger.warning("Current VLLM config is not set.")

3614 + logger.warning("Current vLLM config is not set.")

3615 from vllm.config import VllmConfig

3616 return VllmConfig()

3617 return _current_vllm_config

vllm/entrypoints/openai/protocol.py

+1 -1

@@ -237,7 +237,7 @@ class ChatCompletionRequest(OpenAIBaseModel):

237 tool_choice: Optional[Union[Literal["none"], Literal["auto"],

238 ChatCompletionNamedToolChoiceParam]] = "none"

239

240 - # NOTE this will be ignored by VLLM -- the model determines the behavior

240 + # NOTE this will be ignored by vLLM -- the model determines the behavior

241 parallel_tool_calls: Optional[bool] = False

242 user: Optional[str] = None

243

vllm/envs.py

+3 -3

@@ -164,7 +164,7 @@ environment_variables: dict[str, Callable[[], Any]] = {

164 "VERBOSE":

165 lambda: bool(int(os.getenv('VERBOSE', '0'))),

166

167 - # Root directory for VLLM configuration files

167 + # Root directory for vLLM configuration files

168 # Defaults to `~/.config/vllm` unless `XDG_CONFIG_HOME` is set

169 # Note that this not only affects how vllm finds its configuration files

170 # during runtime, but also affects how vllm installs its configuration

@@ -178,7 +178,7 @@ environment_variables: dict[str, Callable[[], Any]] = {

178

179 # ================== Runtime Env Vars ==================

180

181 - # Root directory for VLLM cache files

181 + # Root directory for vLLM cache files

182 # Defaults to `~/.cache/vllm` unless `XDG_CACHE_HOME` is set

183 "VLLM_CACHE_ROOT":

184 lambda: os.path.expanduser(

@@ -260,7 +260,7 @@ environment_variables: dict[str, Callable[[], Any]] = {

260 "VLLM_ENGINE_ITERATION_TIMEOUT_S":

261 lambda: int(os.environ.get("VLLM_ENGINE_ITERATION_TIMEOUT_S", "60")),

262

263 - # API key for VLLM API server

263 + # API key for vLLM API server

264 "VLLM_API_KEY":

265 lambda: os.environ.get("VLLM_API_KEY", None),

266

vllm/model_executor/models/phi4mm.py

+1 -1

@@ -1414,7 +1414,7 @@ def cat_with_pad(tensors, dim, padding_value=0):

1414 @INPUT_REGISTRY.register_input_processor(input_processor_for_phi4mm)

1415 class Phi4MMForCausalLM(nn.Module, SupportsLoRA, SupportsMultiModal):

1416 """

1417 - Implements the Phi-4-multimodal-instruct model in VLLM.

1417 + Implements the Phi-4-multimodal-instruct model in vLLM.

1418 """

1419 packed_modules_mapping = {

1420 "qkv_proj": [

vllm/platforms/cuda.py

+1 -1

@@ -119,7 +119,7 @@ class CudaPlatformBase(Platform):

119 if envs.VLLM_USE_V1:

120 raise NotImplementedError(

121 "Multi-step scheduling is not supported (and not "

122 - "needed) on VLLM V1. Please launch without "

122 + "needed) on vLLM V1. Please launch without "

123 "--num-scheduler-steps.")

124 else:

125 parallel_config.worker_cls = \

vllm/platforms/rocm.py

+2 -2

@@ -173,7 +173,7 @@ class RocmPlatform(Platform):

173 if envs.VLLM_USE_V1:

174 raise NotImplementedError(

175 "Multi-step scheduling is not supported (and not "

176 - "needed) on VLLM V1. Please launch without "

176 + "needed) on vLLM V1. Please launch without "

177 "--num-scheduler-steps.")

178 else:

179 parallel_config.worker_cls = \

@@ -181,7 +181,7 @@ class RocmPlatform(Platform):

181 elif vllm_config.speculative_config:

182 if envs.VLLM_USE_V1:

183 raise NotImplementedError(

184 - "Speculative decoding is not yet supported on VLLM V1."

184 + "Speculative decoding is not yet supported on vLLM V1."

185 )

186 else:

187 parallel_config.worker_cls = \

vllm/transformers_utils/tokenizers/mistral.py

+1 -1

@@ -249,7 +249,7 @@ class MistralTokenizer(TokenizerBase):

249 revision=revision)

250 return tokenizer_file

251

252 - # the following attributes are set to fit VLLM's design and are used

252 + # the following attributes are set to fit vLLM's design and are used

253 # by the guided structured output backends.

254 @property

255 def all_special_tokens_extended(self) -> List[str]:

vllm/v1/engine/core_client.py

+1 -1

@@ -255,7 +255,7 @@ class MPClient(EngineCoreClient):

255 # TODO(rob): rather than killing the main process, we should

256 # figure out how to raise an AsyncEngineDeadError and

257 # handle at the API server level so we can return a better

258 - # error code to the clients calling VLLM.

258 + # error code to the clients calling vLLM.

259 def sigusr1_handler(signum, frame):

260 logger.fatal("Got fatal signal from worker processes, shutting "

261 "down. See stack trace above for root cause issue.")

vllm/v1/engine/output_processor.py

+1 -1

@@ -248,7 +248,7 @@ class OutputProcessor:

248

249 ****************** NOTE FOR DEVELOPERS ******************

250

251 - VLLM V1 minimizes the number of python loops over the full

251 + vLLM V1 minimizes the number of python loops over the full

252 batch to ensure system overheads are minimized. This is the

253 only function that should loop over EngineCoreOutputs.

254

vllm/v1/engine/processor.py

+2 -2

@@ -93,10 +93,10 @@ class Processor:

93 ) -> None:

94 # Best of not yet supported.

95 if params.best_of is not None and params.best_of > 1:

96 - raise ValueError("VLLM V1 does not yet support best_of.")

96 + raise ValueError("vLLM V1 does not yet support best_of.")

97 # Logits processors not supported.

98 if params.logits_processors:

99 - raise ValueError("VLLM V1 does not support per request "

99 + raise ValueError("vLLM V1 does not support per request "

100 "user provided logits processors.")

101

102 def _validate_params(

Correct capitalisation: `VLLM` -> `vLLM` (#14562)