@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / a766b30349

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 8, 03:45 UTC

[Renderer] Deprecate code paths for old input processing (#34775)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

Cyrus Leung committed Feb 18, 2026 at 16:35 UTC a766b303497bb743defee3f7d0b3477f13477418

6 files changed +70 -51

docs/design/plugin_system.md

@@ -155,3 +155,4 @@ The interface for the model/module may change during vLLM's development. If you

155 - `use_v1` parameter in `Platform.get_attn_backend_cls` is deprecated. It has been removed in v0.13.0.

156 - `_Backend` in `vllm.attention` is deprecated. It has been removed in v0.13.0. Please use `vllm.v1.attention.backends.registry.register_backend` to add new attention backend to `AttentionBackendEnum` instead.

157 - `seed_everything` platform interface is deprecated. It has been removed in v0.16.0. Please use `vllm.utils.torch_utils.set_random_seed` instead.

158 + - `prompt` in `Platform.validate_request` is deprecated and will be removed in v0.18.0.

vllm/entrypoints/llm.py

-24

@@ -519,7 +519,6 @@ class LLM:

519 ),

520 params=seq_params,

521 lora_requests=seq_lora_requests,

522 - tokenization_kwargs=tokenization_kwargs,

522 priorities=seq_priority,

523 )

524

@@ -1813,7 +1812,6 @@ class LLM:

1812 params=seq_params,

1813 use_tqdm=use_tqdm,

1814 lora_requests=seq_lora_requests,

1816 - tokenization_kwargs=tokenization_kwargs,

1815 priorities=seq_priority,

1816 )

1817

@@ -1872,7 +1870,6 @@ class LLM:

1870 params=seq_params,

1871 lora_requests=seq_lora_requests,

1872 use_tqdm=use_tqdm,

1875 - tokenization_kwargs=tokenization_kwargs,

1873 )

1874

1875 def _render_and_run_requests(

@@ -1881,7 +1878,6 @@ class LLM:

1878 params: Sequence[SamplingParams | PoolingParams],

1879 *,

1880 lora_requests: Sequence[LoRARequest | None] | None = None,

1884 - tokenization_kwargs: dict[str, Any] | None = None,

1881 priorities: Sequence[int] | None = None,

1882 use_tqdm: bool | Callable[..., tqdm] = True,

1883 ):

@@ -1899,7 +1895,6 @@ class LLM:

1895 prompts=prompts,

1896 params=params,

1897 lora_requests=lora_requests,

1902 - tokenization_kwargs=tokenization_kwargs,

1898 priorities=priorities,

1899 )

1900

@@ -1911,7 +1906,6 @@ class LLM:

1906 params: Sequence[SamplingParams | PoolingParams],

1907 *,

1908 lora_requests: Sequence[LoRARequest | None] | None = None,

1914 - tokenization_kwargs: dict[str, Any] | None = None,

1909 priorities: Sequence[int] | None = None,

1910 ) -> list[str]:

1911 added_request_ids: list[str] = []

@@ -1922,7 +1916,6 @@ class LLM:

1916 prompt,

1917 params[i],

1918 lora_request=None if lora_requests is None else lora_requests[i],

1925 - tokenization_kwargs=tokenization_kwargs,

1919 priority=0 if priorities is None else priorities[i],

1920 )

1921 added_request_ids.append(request_id)

@@ -1938,7 +1931,6 @@ class LLM:

1931 prompt: ProcessorInputs,

1932 params: SamplingParams | PoolingParams,

1933 lora_request: LoRARequest | None = None,

1941 - tokenization_kwargs: dict[str, Any] | None = None,

1934 priority: int = 0,

1935 ) -> str:

1936 if isinstance(params, SamplingParams):

@@ -1947,27 +1939,11 @@ class LLM:

1939

1940 request_id = str(next(self.request_counter))

1941

1950 - if params.truncate_prompt_tokens is not None:

1951 - params_type = type(params).__name__

1952 - warnings.warn(

1953 - f"The `truncate_prompt_tokens` parameter in `{params_type}` "

1954 - "is deprecated and will be removed in v0.16. "

1955 - "Please pass it via `tokenization_kwargs` instead.",

1956 - DeprecationWarning,

1957 - stacklevel=2,

1958 - )

1959 -

1960 - tokenization_kwargs = merge_kwargs(

1961 - tokenization_kwargs,

1962 - dict(truncate_prompt_tokens=params.truncate_prompt_tokens),

1963 - )

1964 -

1942 return self.llm_engine.add_request(

1943 request_id,

1944 prompt,

1945 params,

1946 lora_request=lora_request,

1970 - tokenization_kwargs=tokenization_kwargs,

1947 priority=priority,

1948 )

1949

vllm/platforms/interface.py

+2 -3

@@ -17,7 +17,7 @@ if TYPE_CHECKING:

17 from torch.distributed import PrefixStore, ProcessGroup

19 from vllm.config import VllmConfig

20 - from vllm.inputs import ProcessorInputs, PromptType

20 + from vllm.inputs import ProcessorInputs

21 from vllm.pooling_params import PoolingParams

22 from vllm.sampling_params import SamplingParams

23 from vllm.utils.argparse_utils import FlexibleArgumentParser

@@ -568,9 +568,8 @@ class Platform:

568 @classmethod

569 def validate_request(

570 cls,

571 - prompt: "PromptType | ProcessorInputs",

572 - params: "SamplingParams | PoolingParams",

571 processed_inputs: "ProcessorInputs",

572 + params: "SamplingParams | PoolingParams",

573 ) -> None:

574 """Raises if this request is unsupported on this platform"""

575

vllm/v1/engine/async_llm.py

+7 -16

@@ -27,7 +27,7 @@ from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalRegistry

27 from vllm.outputs import STREAM_FINISHED, PoolingRequestOutput, RequestOutput

28 from vllm.plugins.io_processors import get_io_processor

29 from vllm.pooling_params import PoolingParams

30 -from vllm.renderers import merge_kwargs, renderer_from_config

30 +from vllm.renderers import renderer_from_config

31 from vllm.renderers.inputs.preprocess import extract_prompt_components

32 from vllm.sampling_params import RequestOutputKind, SamplingParams

33 from vllm.tasks import SupportedTask

@@ -319,21 +319,6 @@ class AsyncLLM(EngineClient):

319 "prompt logprobs"

320 )

321

322 - if params.truncate_prompt_tokens is not None:

323 - params_type = type(params).__name__

324 - warnings.warn(

325 - f"The `truncate_prompt_tokens` parameter in `{params_type}` "

326 - "is deprecated and will be removed in v0.16. "

327 - "Please pass it via `tokenization_kwargs` instead.",

328 - DeprecationWarning,

329 - stacklevel=2,

330 - )

331 -

332 - tokenization_kwargs = merge_kwargs(

333 - tokenization_kwargs,

334 - dict(truncate_prompt_tokens=params.truncate_prompt_tokens),

335 - )

336 -

322 if isinstance(prompt, AsyncGenerator):

323 if reasoning_ended is not None:

324 raise NotImplementedError

@@ -353,6 +338,12 @@ class AsyncLLM(EngineClient):

338

339 # Convert Input --> Request.

340 if isinstance(prompt, EngineCoreRequest):

341 + logger.warning_once(

342 + "Passing EngineCoreRequest to AsyncLLM.generate() and .add_requests() "

343 + "is deprecated and will be removed in v0.18. You should instead pass "

344 + "the outputs of Renderer.render_cmpl() or Renderer.render_chat()."

345 + )

346 +

347 request = prompt

348 if request_id != request.request_id:

349 logger.warning_once(

vllm/v1/engine/input_processor.py

+53 -7

@@ -2,6 +2,7 @@

2 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project

4 import time

5 +import warnings

6 from collections.abc import Mapping

7 from typing import Any, Literal

@@ -28,6 +29,7 @@ from vllm.sampling_params import SamplingParams

29 from vllm.tasks import POOLING_TASKS, SupportedTask

30 from vllm.tokenizers import TokenizerLike

31 from vllm.utils import length_from_prompt_token_ids_or_embeds, random_uuid

32 +from vllm.utils.func_utils import supports_kw

33 from vllm.utils.jsontree import json_iter_leaves

34 from vllm.v1.engine import EngineCoreRequest

@@ -72,6 +74,33 @@ class InputProcessor:

74 mm_registry=mm_registry,

75 )

77 + from vllm.platforms import current_platform

78 +

79 + platform_validate_request = current_platform.validate_request

80 + if supports_kw(platform_validate_request, "prompt"):

81 + logger.warning_once(

82 + "The signature of Platform.validate_request has changed from "

83 + "`(cls, prompt, params, processed_inputs) -> None` to "

84 + "`(cls, processed_inputs, params) -> None`. The old signature "

85 + "will no longer be supported starting from v0.18."

86 + )

87 +

88 + orig_validate_request = platform_validate_request

89 +

90 + def compat_validate_request(

91 + processed_inputs: ProcessorInputs,

92 + params: SamplingParams | PoolingParams,

93 + ):

94 + return orig_validate_request(

95 + processed_inputs,

96 + params,

97 + processed_inputs, # type: ignore

98 + ) # type: ignore

99 +

100 + platform_validate_request = compat_validate_request

101 +

102 + self._platform_validate_request = platform_validate_request

103 +

104 @property

105 def tokenizer(self) -> TokenizerLike | None:

106 return self.renderer.tokenizer

@@ -87,6 +116,16 @@ class InputProcessor:

116 supported_tasks: tuple[SupportedTask, ...] | None,

117 ):

118 """Raise `ValueError` if SamplingParams or PoolingParams is not valid."""

119 + if params.truncate_prompt_tokens is not None:

120 + params_type = type(params).__name__

121 + warnings.warn(

122 + f"The `truncate_prompt_tokens` parameter in `{params_type}` "

123 + "is deprecated and will be removed in v0.17. "

124 + "Please pass it via `tokenization_kwargs` instead.",

125 + DeprecationWarning,

126 + stacklevel=2,

127 + )

128 +

129 if isinstance(params, SamplingParams):

130 params.verify(

131 self.model_config,

@@ -211,11 +250,24 @@ class InputProcessor:

250 )

251

252 if isinstance(prompt, dict) and "type" in prompt:

253 + if tokenization_kwargs:

254 + logger.warning_once(

255 + "Passing tokenization_kwargs to InputProcessor is deprecated "

256 + "and will be removed in v0.18. You should instead pass "

257 + "them to Renderer.render_cmpl() or Renderer.render_chat()."

258 + )

259 +

260 if arrival_time is None:

261 arrival_time = prompt.get("arrival_time", time.time()) # type: ignore[assignment]

262

263 processed_inputs: ProcessorInputs = prompt # type: ignore[assignment]

264 else:

265 + logger.warning_once(

266 + "Passing raw prompts to InputProcessor is deprecated "

267 + "and will be removed in v0.18. You should instead pass "

268 + "the outputs of Renderer.render_cmpl() or Renderer.render_chat()."

269 + )

270 +

271 if arrival_time is None:

272 arrival_time = time.time()

273

@@ -224,13 +276,7 @@ class InputProcessor:

276 tokenization_kwargs=tokenization_kwargs,

277 )

278

227 - from vllm.platforms import current_platform

228 -

229 - current_platform.validate_request(

230 - prompt=prompt,

231 - params=params,

232 - processed_inputs=processed_inputs,

233 - )

279 + self._platform_validate_request(processed_inputs, params)

280

281 encoder_inputs, decoder_inputs = split_enc_dec_inputs(processed_inputs)

282 self._validate_model_inputs(encoder_inputs, decoder_inputs)

vllm/v1/engine/llm_engine.py

+7 -1

@@ -234,10 +234,16 @@ class LLMEngine:

234

235 # Process raw inputs into the request.

236 if isinstance(prompt, EngineCoreRequest):

237 + logger.warning_once(

238 + "Passing EngineCoreRequest to LLMEngine.generate() and .add_requests() "

239 + "is deprecated and will be removed in v0.18. You should instead pass "

240 + "the outputs of Renderer.render_cmpl() or Renderer.render_chat()."

241 + )

242 +

243 request = prompt

244 if request_id != request.request_id:

245 logger.warning_once(

240 - "AsyncLLM.add_request() was passed a request_id parameter that "

246 + "LLMEngine.add_request() was passed a request_id parameter that "

247 "does not match the EngineCoreRequest.request_id attribute. The "

248 "latter will be used, and the former will be ignored."

249 )