@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 96860af655

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 10, 16:30 UTC

[Model] rename use_pad_token to use_sep_token (#31784)

Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>

wang.yuqi committed Jan 6, 2026 at 22:16 UTC 96860af655688163d0228e128a06dd6c462be271

5 files changed +33 -25

examples/pooling/score/convert_model_to_seq_cls.py

+7 -7

@@ -108,7 +108,7 @@ method_map = {

108

109

110 def converting(

111 - model_name, classifier_from_tokens, path, method, use_pad_token=False, device="cpu"

111 + model_name, classifier_from_tokens, path, method, use_sep_token=False, device="cpu"

112 ):

113 """

114 Main conversion function to transform a CausalLM model to SequenceClassification.

@@ -118,7 +118,7 @@ def converting(

118 classifier_from_tokens: List of tokens used for classification

119 path: Output path to save the converted model

120 method: Conversion method ('from_2_way_softmax' or 'no_post_processing')

121 - use_pad_token: Whether to use padding token in the sequence classification model

121 + use_sep_token: Whether to use separating token in the sequence classification model

122 device: Device to load the model on ('cpu' or 'cuda')

123 """

124 assert method in method_map, f"Unknown method: {method}"

@@ -149,10 +149,10 @@ def converting(

149 causal_lm, seq_cls_model, tokenizer, classifier_from_tokens, device

150 )

151

152 - # Configure padding token settings

153 - # Note: Reranker models typically don't use padding tokens by default

154 - seq_cls_model.config.use_pad_token = use_pad_token

155 - seq_cls_model.config.pad_token_id = tokenizer.pad_token_id

152 + # Configure separating token settings

153 + # Note: `llm as reranker` defaults to not using separating token.

154 + seq_cls_model.config.use_sep_token = use_sep_token

155 + seq_cls_model.config.sep_token_id = tokenizer.sep_token_id

156

157 # Save the converted model and tokenizer

158 seq_cls_model.save_pretrained(path)

@@ -203,6 +203,6 @@ if __name__ == "__main__":

203 model_name=args.model_name,

204 classifier_from_tokens=json.loads(args.classifier_from_tokens),

205 method=args.method,

206 - use_pad_token=args.use_pad_token,

206 + use_sep_token=args.use_sep_token,

207 path=args.path,

208 )

tests/entrypoints/pooling/score/test_utils.py

+8 -8

@@ -51,9 +51,9 @@ def llm_reranker_model_config():

51 CROSS_ENCODER_MODEL_ID,

52 runner="pooling",

53 )

54 - # use_pad_token is a property that reads from hf_config,

54 + # use_sep_token is a property that reads from hf_config,

55 # so we set it there to override the default (True)

56 - config.hf_config.use_pad_token = False

56 + config.hf_config.use_sep_token = False

57 return config

@@ -230,7 +230,7 @@ class TestGetScorePrompt:

230 cross_encoder_tokenizer, full_prompt, engine_prompt

231 )

232

233 - def test_fallback_with_pad_token(

233 + def test_fallback_with_sep_token(

234 self,

235 cross_encoder_model_config,

236 cross_encoder_tokenizer,

@@ -238,7 +238,7 @@ class TestGetScorePrompt:

238 mock_model_no_score_template,

239 ):

240 """Test fallback path when ChatTemplateResolutionError

241 - and use_pad_token=True."""

241 + and use_sep_token=True."""

242 with (

243 patch(

244 "vllm.model_executor.model_loader.get_model_cls",

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

250 ),

251 ):

252 full_prompt, engine_prompt = get_score_prompt(

253 - cross_encoder_model_config, # use_pad_token=True

253 + cross_encoder_model_config, # use_sep_token=True

254 cross_encoder_tokenizer,

255 tokenization_kwargs,

256 "query",

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

281 add_special_tokens=False,

282 )

283

284 - def test_fallback_without_pad_token(

284 + def test_fallback_without_sep_token(

285 self,

286 llm_reranker_model_config,

287 cross_encoder_tokenizer,

@@ -289,7 +289,7 @@ class TestGetScorePrompt:

289 mock_model_no_score_template,

290 ):

291 """Test fallback path when ChatTemplateResolutionError

292 - and use_pad_token=False."""

292 + and use_sep_token=False."""

293 with (

294 patch(

295 "vllm.model_executor.model_loader.get_model_cls",

@@ -301,7 +301,7 @@ class TestGetScorePrompt:

301 ),

302 ):

303 full_prompt, engine_prompt = get_score_prompt(

304 - llm_reranker_model_config, # use_pad_token=False

304 + llm_reranker_model_config, # use_sep_token=False

305 cross_encoder_tokenizer,

306 tokenization_kwargs,

307 "query",

vllm/config/model.py

+12 -4

@@ -1434,10 +1434,18 @@ class ModelConfig:

1434 return getattr(self.hf_config, "matryoshka_dimensions", None)

1435

1436 @property

1437 - def use_pad_token(self) -> bool:

1438 - # cross_encoder models defaults to using pad_token.

1439 - # `llm as reranker` models defaults to not using pad_token.

1440 - return getattr(self.hf_config, "use_pad_token", True)

1437 + def use_sep_token(self) -> bool:

1438 + # cross_encoder models defaults to using separating token.

1439 + # `llm as reranker` defaults to not using separating token.

1440 +

1441 + use_pad_token = getattr(self.hf_config, "use_pad_token", None)

1442 + if use_pad_token is not None:

1443 + logger.warning_once(

1444 + "use_pad_token has been deprecated; please use use_sep_token instead."

1445 + )

1446 + return use_pad_token

1447 +

1448 + return getattr(self.hf_config, "use_sep_token", True)

1449

1450 @property

1451 def head_dtype(self) -> torch.dtype:

vllm/entrypoints/score_utils.py

+3 -3

@@ -199,14 +199,14 @@ def get_score_prompt(

199 full_prompt = _apply_model_score_template(model_config, prompt_1, prompt_2)

200 prompt_inputs = tokenizer(full_prompt, **tokenization_kwargs)

201 else:

202 - if model_config.use_pad_token:

203 - # cross_encoder models defaults to using pad_token.

202 + if model_config.use_sep_token:

203 + # cross_encoder models defaults to using separating token.

204 prompt_inputs = tokenizer(

205 text=prompt_1, text_pair=prompt_2, **tokenization_kwargs

206 )

207 full_prompt = tokenizer.decode(prompt_inputs["input_ids"])

208 else:

209 - # `llm as reranker` models defaults to not using pad_token.

209 + # `llm as reranker` defaults to not using separating token.

210 full_prompt = prompt_1 + prompt_2

211 prompt_inputs = tokenizer(text=full_prompt, **tokenization_kwargs)

212 return full_prompt, prompt_inputs

vllm/model_executor/models/adapters.py

+3 -3

@@ -382,9 +382,9 @@ class SequenceClassificationConfig(VerifyAndUpdateConfig):

382 else:

383 text_config.num_labels = len(tokens)

384

385 - # `llm as reranker` defaults to not using pad_token

386 - use_pad_token = getattr(text_config, "use_pad_token", False)

387 - text_config.use_pad_token = use_pad_token

385 + # `llm as reranker` defaults to not using separating token.

386 + use_sep_token = getattr(text_config, "use_sep_token", False)

387 + text_config.use_sep_token = use_sep_token

388

389

390 def load_weights_using_from_2_way_softmax(