4f5b059f14 · tduncan/Nvidia-Optimization-Stack-VLLM

vllm/model_executor/models/arctic.py

-1

@@ -370,7 +370,6 @@ class ArcticModel(nn.Module):

370 cache_config = vllm_config.cache_config

371 quant_config = vllm_config.quant_config

372

373 - self.padding_idx = config.pad_token_id

373 self.vocab_size = config.vocab_size

374 self.embed_tokens = VocabParallelEmbedding(

375 self.vocab_size,

vllm/model_executor/models/baichuan.py

-1

@@ -267,7 +267,6 @@ class BaiChuanModel(nn.Module):

267 quant_config = vllm_config.quant_config

268

269 self.config = config

270 - self.padding_idx = config.pad_token_id

270 self.vocab_size = config.vocab_size

271

272 self.embed_tokens = VocabParallelEmbedding(

vllm/model_executor/models/bart.py

-1

@@ -725,7 +725,6 @@ class BartModel(nn.Module):

725

726 self.config = config

727

728 - self.padding_idx = config.pad_token_id

728 lora_vocab = (lora_config.lora_extra_vocab_size *

729 (lora_config.max_loras or 1)) if lora_config else 0

730 self.vocab_size = config.vocab_size + lora_vocab

vllm/model_executor/models/chameleon.py

-1

@@ -851,7 +851,6 @@ class ChameleonModel(nn.Module):

851 quant_config = vllm_config.quant_config

852

853 self.config = config

854 - self.padding_idx = config.pad_token_id

854 self.vocab_size = config.vocab_size

855 self.embed_tokens = VocabParallelEmbedding(

856 self.vocab_size,

vllm/model_executor/models/deepseek.py

-1

@@ -339,7 +339,6 @@ class DeepseekModel(nn.Module):

339 cache_config = vllm_config.cache_config

340 quant_config = vllm_config.quant_config

341

342 - self.padding_idx = config.pad_token_id

342 self.vocab_size = config.vocab_size

343

344 self.embed_tokens = VocabParallelEmbedding(

vllm/model_executor/models/deepseek_v2.py

-1

@@ -570,7 +570,6 @@ class DeepseekV2Model(nn.Module):

570 cache_config = vllm_config.cache_config

571 quant_config = vllm_config.quant_config

572

573 - self.padding_idx = config.pad_token_id

573 self.vocab_size = config.vocab_size

574

575 if get_pp_group().is_first_rank:

vllm/model_executor/models/exaone.py

-1

@@ -313,7 +313,6 @@ class ExaoneModel(nn.Module):

313 lora_config = vllm_config.lora_config

314

315 self.config = config

316 - self.padding_idx = config.pad_token_id

316 lora_vocab = ((lora_config.lora_extra_vocab_size *

317 (lora_config.max_loras or 1)) if lora_config else 0)

318 self.vocab_size = config.vocab_size + lora_vocab

vllm/model_executor/models/florence2.py

-1

@@ -592,7 +592,6 @@ class Florence2LanguageModel(nn.Module):

592

593 self.config = config

594

595 - self.padding_idx = config.pad_token_id

595 self.vocab_size = config.vocab_size

596

597 self.shared = BartScaledWordEmbedding(self.vocab_size, config.d_model)

vllm/model_executor/models/fuyu.py

-1

@@ -255,7 +255,6 @@ class FuyuForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):

255 self.config = config

256 self.multimodal_config = multimodal_config

257

258 - self.padding_idx = config.pad_token_id

258 self.vocab_size = config.text_config.vocab_size

259 self.image_token_id = _IMAGE_TOKEN_ID

260 self.image_feature_size = config.patch_size**2 * config.num_channels

vllm/model_executor/models/granite.py

-1

@@ -260,7 +260,6 @@ class GraniteModel(nn.Module):

260 lora_config = vllm_config.lora_config

261

262 self.config = config

263 - self.padding_idx = config.pad_token_id

263 lora_vocab = (lora_config.lora_extra_vocab_size *

264 (lora_config.max_loras or 1)) if lora_config else 0

265 self.vocab_size = config.vocab_size + lora_vocab

vllm/model_executor/models/granitemoe.py

-1

@@ -252,7 +252,6 @@ class GraniteMoeModel(nn.Module):

252 quant_config = vllm_config.quant_config

253 lora_config = vllm_config.lora_config

254

255 - self.padding_idx = config.pad_token_id

255 lora_vocab = (lora_config.lora_extra_vocab_size *

256 (lora_config.max_loras or 1)) if lora_config else 0

257 self.vocab_size = config.vocab_size + lora_vocab

vllm/model_executor/models/idefics3.py

-1

@@ -404,7 +404,6 @@ class Idefics3Model(nn.Module):

404 quant_config = vllm_config.quant_config

405

406 self.config = config

407 - self.padding_idx = self.config.text_config.pad_token_id

407 self.vocab_size = self.config.text_config.vocab_size

408 self.vision_model = Idefics3VisionTransformer(

409 config.vision_config,

vllm/model_executor/models/internlm2.py

-1

@@ -261,7 +261,6 @@ class InternLM2Model(nn.Module):

261 quant_config = vllm_config.quant_config

262

263 self.config = config

264 - self.padding_idx = config.pad_token_id

264 self.vocab_size = config.vocab_size

265 self.tok_embeddings = VocabParallelEmbedding(

266 config.vocab_size,

vllm/model_executor/models/jamba.py

-1

@@ -271,7 +271,6 @@ class JambaModel(nn.Module):

271 lora_config = vllm_config.lora_config

272

273 self.config = config

274 - self.padding_idx = config.pad_token_id

274 lora_vocab = ((lora_config.lora_extra_vocab_size *

275 (lora_config.max_loras or 1)) if lora_config else 0)

276 self.vocab_size = config.vocab_size + lora_vocab

vllm/model_executor/models/llama.py

-1

@@ -302,7 +302,6 @@ class LlamaModel(nn.Module):

302

303 self.config = config

304 self.quant_config = quant_config

305 - self.padding_idx = config.pad_token_id

305 lora_vocab = (lora_config.lora_extra_vocab_size *

306 (lora_config.max_loras or 1)) if lora_config else 0

307 self.vocab_size = config.vocab_size + lora_vocab

vllm/model_executor/models/mamba.py

-1

@@ -90,7 +90,6 @@ class MambaModel(nn.Module):

90 is_lora_enabled = bool(lora_config)

91

92 self.config = config

93 - self.padding_idx = config.pad_token_id

93 lora_vocab = ((lora_config.lora_extra_vocab_size *

94 (lora_config.max_loras or 1)) if lora_config else 0)

95 self.vocab_size = config.vocab_size + lora_vocab

vllm/model_executor/models/minicpm.py

-1

@@ -365,7 +365,6 @@ class MiniCPMModel(nn.Module):

365 self.config = config

366 self.cache_config = cache_config

367 self.quant_config = quant_config

368 - self.padding_idx = config.pad_token_id

368 lora_vocab = (lora_config.lora_extra_vocab_size *

369 (lora_config.max_loras or 1)) if lora_config else 0

370 self.vocab_size = config.vocab_size + lora_vocab

vllm/model_executor/models/mixtral.py

-1

@@ -254,7 +254,6 @@ class MixtralModel(nn.Module):

254 quant_config = vllm_config.quant_config

255 lora_config = vllm_config.lora_config

256

257 - self.padding_idx = config.pad_token_id

257 lora_vocab = (lora_config.lora_extra_vocab_size *

258 (lora_config.max_loras or 1)) if lora_config else 0

259 self.vocab_size = config.vocab_size + lora_vocab

vllm/model_executor/models/mixtral_quant.py

-1

@@ -302,7 +302,6 @@ class MixtralModel(nn.Module):

302 cache_config = vllm_config.cache_config

303 quant_config = vllm_config.quant_config

304

305 - self.padding_idx = config.pad_token_id

305 self.vocab_size = config.vocab_size

306

307 self.embed_tokens = VocabParallelEmbedding(

vllm/model_executor/models/mllama.py

-1

@@ -1031,7 +1031,6 @@ class MllamaTextModel(nn.Module):

1031 cache_config = vllm_config.cache_config

1032 quant_config = vllm_config.quant_config

1033

1034 - self.padding_idx = config.pad_token_id

1034 self.vocab_size = config.vocab_size

1035 self.embed_tokens = VocabParallelEmbedding(config.vocab_size + 8,

1036 config.hidden_size)

vllm/model_executor/models/nemotron.py

-1

@@ -300,7 +300,6 @@ class NemotronModel(nn.Module):

300 lora_config = vllm_config.lora_config

301

302 self.config = config

303 - self.padding_idx = config.pad_token_id

303 lora_vocab = (lora_config.lora_extra_vocab_size *

304 (lora_config.max_loras or 1)) if lora_config else 0

305 self.vocab_size = config.vocab_size + lora_vocab

vllm/model_executor/models/olmoe.py

-1

@@ -252,7 +252,6 @@ class OlmoeModel(nn.Module):

252 cache_config = vllm_config.cache_config

253 quant_config = vllm_config.quant_config

254

255 - self.padding_idx = config.pad_token_id

255 self.vocab_size = config.vocab_size

256

257 self.embed_tokens = VocabParallelEmbedding(

vllm/model_executor/models/opt.py

-1

@@ -200,7 +200,6 @@ class OPTDecoder(nn.Module):

200 ):

201 super().__init__()

202 self.config = config

203 - self.padding_idx = config.pad_token_id

203 self.max_target_positions = config.max_position_embeddings

204 self.vocab_size = config.vocab_size

205

vllm/model_executor/models/orion.py

-1

@@ -217,7 +217,6 @@ class OrionModel(nn.Module):

217 quant_config = vllm_config.quant_config

218

219 self.config = config

220 - self.padding_idx = config.pad_token_id

220 self.vocab_size = config.vocab_size

221 self.embed_tokens = VocabParallelEmbedding(

222 config.vocab_size,

vllm/model_executor/models/phimoe.py

-1

@@ -441,7 +441,6 @@ class PhiMoEModel(nn.Module):

441 quant_config = vllm_config.quant_config

442 lora_config = vllm_config.lora_config

443

444 - self.padding_idx = config.pad_token_id

444 lora_vocab = ((lora_config.lora_extra_vocab_size *

445 (lora_config.max_loras or 1)) if lora_config else 0)

446 self.vocab_size = config.vocab_size + lora_vocab

vllm/model_executor/models/qwen2.py

-1

@@ -284,7 +284,6 @@ class Qwen2Model(nn.Module):

284

285 self.config = config

286 self.quant_config = quant_config

287 - self.padding_idx = config.pad_token_id

287 self.vocab_size = config.vocab_size

288

289 if get_pp_group().is_first_rank or (config.tie_word_embeddings

vllm/model_executor/models/qwen2_moe.py

-1

@@ -325,7 +325,6 @@ class Qwen2MoeModel(nn.Module):

325 cache_config = vllm_config.cache_config

326 quant_config = vllm_config.quant_config

327

328 - self.padding_idx = config.pad_token_id

328 self.vocab_size = config.vocab_size

329

330 self.embed_tokens = VocabParallelEmbedding(

vllm/model_executor/models/solar.py

-1

@@ -269,7 +269,6 @@ class SolarModel(nn.Module):

269 lora_config = vllm_config.lora_config

270

271 self.config = config

272 - self.padding_idx = config.pad_token_id

272 lora_vocab = ((lora_config.lora_extra_vocab_size *

273 (lora_config.max_loras or 1)) if lora_config else 0)

274 self.vocab_size = config.vocab_size + lora_vocab

vllm/model_executor/models/starcoder2.py

-2

@@ -212,10 +212,8 @@ class Starcoder2Model(nn.Module):

212 quant_config = vllm_config.quant_config

213

214 self.config = config

215 - self.padding_idx = config.pad_token_id

215 self.vocab_size = config.vocab_size

216

218 - # TODO: consider padding_idx (currently removed)

217 self.embed_tokens = VocabParallelEmbedding(

218 config.vocab_size,

219 config.hidden_size,

vllm/model_executor/models/whisper.py

+1 -5

@@ -49,10 +49,7 @@ class WhisperAudioInputs(TypedDict):

49

50 class WhisperPositionalEmbedding(nn.Embedding):

51

52 - def __init__(self,

53 - num_positions: int,

54 - embedding_dim: int,

55 - padding_idx: Optional[int] = None):

52 + def __init__(self, num_positions: int, embedding_dim: int):

53 super().__init__(num_positions, embedding_dim)

54

55 def forward(self, position_ids):

@@ -359,7 +356,6 @@ class WhisperEncoder(nn.Module):

356 config = vllm_config.model_config.hf_config

357 embed_dim = config.d_model

358 self.num_mel_bins = config.num_mel_bins

362 - self.padding_idx = config.pad_token_id

359 self.max_source_positions = config.max_source_positions

360 self.embed_scale = (math.sqrt(embed_dim)

361 if config.scale_embedding else 1.0)

Clean up unused padding_idx variables across many model definitions (#13240)