@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 032d661d27

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

[Docs] Fix warnings in mkdocs build (continued) (#25042)

Signed-off-by: wwl2755 <wangwenlong2755@gmail.com>

Wenlong Wang committed Sep 20, 2025 at 04:45 UTC 032d661d27db873ddd71ffb609c79c581b3f27b9

7 files changed +24 -15

vllm/multimodal/__init__.py

+1 -1

@@ -15,7 +15,7 @@ is used by model runners to dispatch data processing according to the target

15 model.

17 Info:

18 - [mm_processing](../../../design/mm_processing.html)

18 + [mm_processing](../../../design/mm_processing.md)

19 """

21 __all__ = [

vllm/utils/__init__.py

+3 -3

@@ -3216,7 +3216,7 @@ def cprofile_context(save_file: Optional[str] = None):

3216

3217 Args:

3218 save_file: path to save the profile result. "1" or

3219 - None will result in printing to stdout.

3219 + None will result in printing to stdout.

3220 """

3221 import cProfile

3222

@@ -3273,7 +3273,7 @@ def check_use_alibi(model_config: ModelConfig) -> bool:

3273 and getattr(cfg.attn_config, "alibi", False)))))

3274

3275

3276 -def sha256(input) -> bytes:

3276 +def sha256(input: Any) -> bytes:

3277 """Hash any picklable Python object using SHA-256.

3278

3279 The input is serialized using pickle before hashing, which allows

@@ -3290,7 +3290,7 @@ def sha256(input) -> bytes:

3290 return hashlib.sha256(input_bytes).digest()

3291

3292

3293 -def sha256_cbor(input) -> bytes:

3293 +def sha256_cbor(input: Any) -> bytes:

3294 """

3295 Hash objects using CBOR serialization and SHA-256.

3296

vllm/v1/core/kv_cache_utils.py

+2 -2

@@ -1229,8 +1229,8 @@ def get_kv_cache_configs(vllm_config: VllmConfig,

1229 Args:

1230 vllm_config: The global VllmConfig

1231 kv_cache_specs: List of dict[layer_name, KVCacheSpec] for each worker.

1232 - available_memory: Memory available for KV cache in bytes for each

1233 - worker.

1232 + available_memory: Memory available for KV cache in bytes for each

1233 + worker.

1234

1235 Returns:

1236 The generated KVCacheConfigs for each worker.

vllm/v1/sample/rejection_sampler.py

+5 -5

@@ -351,17 +351,17 @@ def generate_uniform_probs(

351 without a seed.

352

353 Args:

354 - num_tokens : int

354 + num_tokens: int

355 Total number of tokens.

356 - num_draft_tokens : List[List[int]]

356 + num_draft_tokens: List[List[int]]

357 Number of draft tokens per request.

358 - generators : Optional[Dict[int, torch.Generator]]

358 + generators: Optional[Dict[int, torch.Generator]]

359 A dictionary mapping indices in the batch to

360 `torch.Generator` objects.

361 - device : torch.device

361 + device: torch.device

362 The device on which to allocate the tensor.

363 Returns:

364 - uniform_rand : torch.Tensor

364 + uniform_rand: torch.Tensor

365 A tensor of shape `(num_tokens, )` containing uniform

366 random values in the range [0, 1).

367 """

vllm/v1/worker/gpu_model_runner.py

+1 -1

@@ -1479,7 +1479,7 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):

1479

1480 Args:

1481 scheduler_output: The scheduler output containing scheduled encoder

1482 - inputs.

1482 + inputs.

1483

1484 Returns:

1485 A tuple of (mm_kwargs, req_ids_pos) where:

vllm/v1/worker/utils.py

+2 -1

@@ -205,7 +205,8 @@ def gather_mm_placeholders(

205 """

206 Reconstructs the embeddings from the placeholder tokens.

207

208 - This is the operation of [scatter_mm_placeholders][].

208 + This is the operation of [`scatter_mm_placeholders`]

209 + [vllm.v1.worker.utils.scatter_mm_placeholders].

210 """

211 if is_embed is None:

212 return placeholders

vllm/worker/model_runner.py

+10 -2

@@ -1810,7 +1810,8 @@ class ModelRunner(GPUModelRunnerBase[ModelInputForGPUWithSamplingMetadata]):

1810

1811 return [output]

1812

1813 - def need_recv_kv(self, model_input, kv_caches) -> bool:

1813 + def need_recv_kv(self, model_input: ModelInputForGPUWithSamplingMetadata,

1814 + kv_caches: List[torch.Tensor]) -> bool:

1815 """Check if we need to receive kv-cache from the other worker.

1816 We need to receive KV when

1817 1. current vLLM instance is KV cache consumer/decode vLLM instance

@@ -1825,6 +1826,9 @@ class ModelRunner(GPUModelRunnerBase[ModelInputForGPUWithSamplingMetadata]):

1826 if self.vllm_config.kv_transfer_config is None:

1827 return False

1828

1829 + if model_input.attn_metadata is None:

1830 + raise ValueError("model_input.attn_metadata cannot be None")

1831 +

1832 prefill_meta = model_input.attn_metadata.prefill_metadata

1833

1834 # check if the current run is profiling

@@ -1835,7 +1839,8 @@ class ModelRunner(GPUModelRunnerBase[ModelInputForGPUWithSamplingMetadata]):

1839 return self.vllm_config.kv_transfer_config.is_kv_consumer and (

1840 not is_profile_run) and is_prefill_run

1841

1838 - def need_send_kv(self, model_input, kv_caches) -> bool:

1842 + def need_send_kv(self, model_input: ModelInputForGPUWithSamplingMetadata,

1843 + kv_caches: List[torch.Tensor]) -> bool:

1844 """Check if we need to send kv-cache to the other worker.

1845 We need to send KV when

1846 1. current vLLM instance is KV cache producer/prefill vLLM instance

@@ -1850,6 +1855,9 @@ class ModelRunner(GPUModelRunnerBase[ModelInputForGPUWithSamplingMetadata]):

1855 if self.vllm_config.kv_transfer_config is None:

1856 return False

1857

1858 + if model_input.attn_metadata is None:

1859 + raise ValueError("model_input.attn_metadata cannot be None")

1860 +

1861 prefill_meta = model_input.attn_metadata.prefill_metadata

1862

1863 # check if the current run is profiling