[Models] Improve iteration over layers (#19497)

vllm/model_executor/models/arcee.py

+2 -1

@@ -9,6 +9,7 @@

9 # activation.

10

11 from collections.abc import Iterable

12 +from itertools import islice

13 from typing import Any, Optional, Union

14

15 import torch

@@ -243,7 +244,7 @@ class ArceeModel(nn.Module):

244

245 aux_hidden_states: list[torch.Tensor] = []

246 for idx, layer in enumerate(

246 - self.layers[self.start_layer:self.end_layer]):

247 + islice(self.layers, self.start_layer, self.end_layer)):

248 if idx in self.aux_hidden_state_layers:

249 aux_hidden_states.append(

250 hidden_states +

vllm/model_executor/models/arctic.py

+2 -1

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

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

3 """Inference-only Snowflake Arctic model."""

4 from collections.abc import Iterable

5 +from itertools import islice

6 from typing import Optional, Union

7

8 import torch

@@ -403,7 +404,7 @@ class ArcticModel(nn.Module):

404 else:

405 assert intermediate_tensors is not None

406 hidden_states = intermediate_tensors["hidden_states"]

406 - for layer in self.layers[self.start_layer:self.end_layer]:

407 + for layer in islice(self.layers, self.start_layer, self.end_layer):

408 hidden_states = layer(positions, hidden_states)

409 if not get_pp_group().is_last_rank:

410 return IntermediateTensors({"hidden_states": hidden_states})

vllm/model_executor/models/baichuan.py

+2 -1

@@ -22,6 +22,7 @@

22 """Inference-only BaiChuan model compatible with HuggingFace weights."""

23 import math

24 from collections.abc import Iterable

25 +from itertools import islice

26 from typing import Optional, Union

27

28 import torch

@@ -309,7 +310,7 @@ class BaiChuanModel(nn.Module):

310 assert intermediate_tensors is not None

311 hidden_states = intermediate_tensors["hidden_states"]

312 residual = intermediate_tensors["residual"]

312 - for layer in self.layers[self.start_layer:self.end_layer]:

313 + for layer in islice(self.layers, self.start_layer, self.end_layer):

314 hidden_states, residual = layer(

315 positions,

316 hidden_states,

vllm/model_executor/models/bailing_moe.py

+2 -2

@@ -24,6 +24,7 @@

24 # limitations under the License.

25 """Inference-only BailingMoE model compatible with HuggingFace weights."""

26 from collections.abc import Iterable

27 +from itertools import islice

28 from typing import Optional, Union

29

30 import torch

@@ -359,8 +360,7 @@ class BailingMoeModel(nn.Module):

360 hidden_states = intermediate_tensors["hidden_states"]

361 residual = intermediate_tensors["residual"]

362

362 - for i in range(self.start_layer, self.end_layer):

363 - layer = self.layers[i]

363 + for layer in islice(self.layers, self.start_layer, self.end_layer):

364 hidden_states, residual = layer(

365 hidden_states,

366 position_ids,

vllm/model_executor/models/bamba.py

+1 -2

@@ -345,8 +345,7 @@ class BambaModel(nn.Module):

345

346 residual = None

347 num_attn = 0

348 - for i in range(len(self.layers)):

349 - layer = self.layers[i]

348 + for i, layer in enumerate(self.layers):

349 if isinstance(layer, BambaAttentionDecoderLayer):

350 num_attn += 1

351

vllm/model_executor/models/bloom.py

+2 -1

@@ -20,6 +20,7 @@

20 """Inference-only BLOOM model compatible with HuggingFace weights."""

21 import math

22 from collections.abc import Iterable

23 +from itertools import islice

24 from typing import Optional, Union

25

26 import torch

@@ -273,7 +274,7 @@ class BloomModel(nn.Module):

274 else:

275 assert intermediate_tensors is not None

276 hidden_states = intermediate_tensors["hidden_states"]

276 - for layer in self.h[self.start_layer:self.end_layer]:

277 + for layer in islice(self.h, self.start_layer, self.end_layer):

278 hidden_states = layer(position_ids, hidden_states)

279 if not get_pp_group().is_last_rank:

280 return IntermediateTensors({"hidden_states": hidden_states})

vllm/model_executor/models/chameleon.py

+2 -1

@@ -3,6 +3,7 @@

3

4 from collections.abc import Iterable, Mapping, Sequence

5 from functools import cached_property

6 +from itertools import islice

7 from typing import Annotated, Any, Literal, Optional, Union

8

9 import torch

@@ -914,7 +915,7 @@ class ChameleonModel(nn.Module):

915 assert intermediate_tensors is not None

916 hidden_states = intermediate_tensors["hidden_states"]

917 residual = intermediate_tensors["residual"]

917 - for layer in self.layers[self.start_layer:self.end_layer]:

918 + for layer in islice(self.layers, self.start_layer, self.end_layer):

919 hidden_states, residual = layer(

920 positions,

921 hidden_states,

vllm/model_executor/models/chatglm.py

+2 -1

@@ -5,6 +5,7 @@

5 """Inference-only ChatGLM model compatible with THUDM weights."""

6 import json

7 from collections.abc import Iterable

8 +from itertools import islice

9 from typing import Optional, Union

10

11 import torch

@@ -281,7 +282,7 @@ class GLMTransformer(nn.Module):

282 hidden_states: torch.Tensor,

283 position_ids: torch.Tensor,

284 ) -> Union[torch.Tensor, IntermediateTensors]:

284 - for layer in self.layers[self.start_layer:self.end_layer]:

285 + for layer in islice(self.layers, self.start_layer, self.end_layer):

286 hidden_states = layer(hidden_states=hidden_states,

287 position_ids=position_ids)

288

vllm/model_executor/models/commandr.py

+2 -1

@@ -23,6 +23,7 @@

23 # This file is based on the LLama model definition file in transformers

24 """PyTorch Cohere model."""

25 from collections.abc import Iterable

26 +from itertools import islice

27 from typing import Optional, Union

28

29 import torch

@@ -322,7 +323,7 @@ class CohereModel(nn.Module):

323 assert intermediate_tensors is not None

324 hidden_states = intermediate_tensors["hidden_states"]

325 residual = intermediate_tensors["residual"]

325 - for layer in self.layers[self.start_layer:self.end_layer]:

326 + for layer in islice(self.layers, self.start_layer, self.end_layer):

327 hidden_states, residual = layer(

328 positions,

329 hidden_states,

vllm/model_executor/models/dbrx.py

+2 -1

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

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

3

4 from collections.abc import Iterable

5 +from itertools import islice

6 from typing import Optional, Union

7

8 import torch

@@ -359,7 +360,7 @@ class DbrxModel(nn.Module):

360 else:

361 assert intermediate_tensors

362 hidden_states = intermediate_tensors["hidden_states"]

362 - for block in self.blocks[self.start_layer:self.end_layer]:

363 + for block in islice(self.blocks, self.start_layer, self.end_layer):

364 hidden_states = block(position_ids, hidden_states)

365 if not get_pp_group().is_last_rank:

366 return IntermediateTensors({"hidden_states": hidden_states})

vllm/model_executor/models/deepseek.py

+3 -2

@@ -24,6 +24,7 @@

24 # limitations under the License.

25 """Inference-only Deepseek model."""

26 from collections.abc import Iterable

27 +from itertools import islice

28 from typing import Any, Optional, Union

29

30 import torch

@@ -377,7 +378,7 @@ class DeepseekModel(nn.Module):

378 else:

379 hidden_states = intermediate_tensors["hidden_states"]

380 residual = intermediate_tensors["residual"]

380 - for layer in self.layers[self.start_layer:self.end_layer]:

381 + for layer in islice(self.layers, self.start_layer, self.end_layer):

382 hidden_states, residual = layer(positions, hidden_states, residual)

383 if not get_pp_group().is_last_rank:

384 return IntermediateTensors({

@@ -483,4 +484,4 @@ class DeepseekForCausalLM(nn.Module, SupportsPP):

484 def load_weights(self, weights: Iterable[tuple[str,

485 torch.Tensor]]) -> set[str]:

486 loader = AutoWeightsLoader(self)

486 - return loader.load_weights(weights)

\ No newline at end of file

487 + return loader.load_weights(weights)

vllm/model_executor/models/deepseek_v2.py

+2 -1

@@ -25,6 +25,7 @@

25 """Inference-only DeepseekV2/DeepseekV3 model."""

26 import typing

27 from collections.abc import Callable, Iterable

28 +from itertools import islice

29 from typing import Any, Optional, Union

30

31 import torch

@@ -712,7 +713,7 @@ class DeepseekV2Model(nn.Module):

713 hidden_states = intermediate_tensors["hidden_states"]

714 residual = intermediate_tensors["residual"]

715

715 - for layer in self.layers[self.start_layer:self.end_layer]:

716 + for layer in islice(self.layers, self.start_layer, self.end_layer):

717 hidden_states, residual = layer(positions, hidden_states, residual)

718

719 if not get_pp_group().is_last_rank:

vllm/model_executor/models/dots1.py

+2 -1

@@ -25,6 +25,7 @@

25 # limitations under the License.

26 """Inference-only dots1 model."""

27 from collections.abc import Iterable

28 +from itertools import islice

29 from typing import Any, Optional, Union

30

31 import torch

@@ -391,7 +392,7 @@ class Dots1Model(nn.Module):

392 assert intermediate_tensors is not None

393 hidden_states = intermediate_tensors["hidden_states"]

394 residual = intermediate_tensors["residual"]

394 - for layer in self.layers[self.start_layer:self.end_layer]:

395 + for layer in islice(self.layers, self.start_layer, self.end_layer):

396 hidden_states, residual = layer(

397 positions,

398 hidden_states,

vllm/model_executor/models/ernie45_moe.py

+2 -2

@@ -23,6 +23,7 @@

23 # limitations under the License.

24 """Inference-only ErineMoE model compatible with HuggingFace weights."""

25 from collections.abc import Iterable

26 +from itertools import islice

27 from typing import Any, Optional, Union

28

29 import torch

@@ -419,8 +420,7 @@ class Ernie4_5_MoeModel(nn.Module):

420 hidden_states = intermediate_tensors["hidden_states"]

421 residual = intermediate_tensors["residual"]

422

422 - for i in range(self.start_layer, self.end_layer):

423 - layer = self.layers[i]

423 + for layer in islice(self.layers, self.start_layer, self.end_layer):

424 hidden_states, residual = layer(positions, hidden_states, residual)

425

426 if not get_pp_group().is_last_rank:

vllm/model_executor/models/ernie45_vl_moe.py

+2 -2

@@ -23,6 +23,7 @@

23 # limitations under the License.

24 """Inference-only Erine VL model compatible with HuggingFace weights."""

25 from collections.abc import Iterable

26 +from itertools import islice

27 from typing import Any, Optional, Union

28

29 import torch

@@ -508,8 +509,7 @@ class Ernie4_5_VLMoeModel(nn.Module):

509 hidden_states = intermediate_tensors["hidden_states"]

510 residual = intermediate_tensors["residual"]

511

511 - for i in range(self.start_layer, self.end_layer):

512 - layer = self.layers[i]

512 + for layer in islice(self.layers, self.start_layer, self.end_layer):

513 hidden_states, residual = layer(positions, hidden_states, residual,

514 visual_token_mask, **kwargs)

515

vllm/model_executor/models/exaone.py

+2 -1

@@ -26,6 +26,7 @@

26 """Inference-only Exaone model compatible with HuggingFace weights."""

27

28 from collections.abc import Iterable

29 +from itertools import islice

30 from typing import Any, Optional, Union

31

32 import torch

@@ -371,7 +372,7 @@ class ExaoneModel(nn.Module):

372 hidden_states = intermediate_tensors["hidden_states"]

373 residual = intermediate_tensors["residual"]

374

374 - for layer in self.h[self.start_layer:self.end_layer]:

375 + for layer in islice(self.h, self.start_layer, self.end_layer):

376 hidden_states, residual = layer(

377 positions,

378 hidden_states,

vllm/model_executor/models/exaone4.py

+2 -1

@@ -22,6 +22,7 @@

22 """Inference-only Exaone model compatible with HuggingFace weights."""

23

24 from collections.abc import Iterable

25 +from itertools import islice

26 from typing import Any, Optional, Union

27

28 import torch

@@ -354,7 +355,7 @@ class Exaone4Model(nn.Module):

355 hidden_states = intermediate_tensors["hidden_states"]

356 residual = intermediate_tensors["residual"]

357

357 - for layer in self.layers[self.start_layer:self.end_layer]:

358 + for layer in islice(self.layers, self.start_layer, self.end_layer):

359 hidden_states, residual = layer(

360 positions,

361 hidden_states,

vllm/model_executor/models/falcon.py

+2 -1

@@ -22,6 +22,7 @@

22

23 import math

24 from collections.abc import Iterable

25 +from itertools import islice

26 from typing import Optional, Union

27

28 import torch

@@ -389,7 +390,7 @@ class FalconModel(nn.Module):

390 hidden_states = self.get_input_embeddings(input_ids)

391 else:

392 hidden_states = intermediate_tensors["hidden_states"]

392 - for layer in self.h[self.start_layer:self.end_layer]:

393 + for layer in islice(self.h, self.start_layer, self.end_layer):

394 hidden_states = layer(positions, hidden_states)

395 if not get_pp_group().is_last_rank:

396 return IntermediateTensors({"hidden_states": hidden_states})

vllm/model_executor/models/gemma.py

+2 -1

@@ -18,6 +18,7 @@

18 """Inference-only Gemma model compatible with HuggingFace weights."""

19 from collections.abc import Iterable

20 from functools import cache

21 +from itertools import islice

22 from typing import Optional, Union

23

24 import torch

@@ -308,7 +309,7 @@ class GemmaModel(nn.Module):

309 else:

310 hidden_states = intermediate_tensors["hidden_states"]

311 residual = intermediate_tensors["residual"]

311 - for layer in self.layers[self.start_layer:self.end_layer]:

312 + for layer in islice(self.layers, self.start_layer, self.end_layer):

313 hidden_states, residual = layer(

314 positions,

315 hidden_states,

vllm/model_executor/models/gemma2.py

+2 -1

@@ -17,6 +17,7 @@

17 # See the License for the specific language governing permissions and

18 # limitations under the License.

19 from collections.abc import Iterable

20 +from itertools import islice

21 from typing import Optional, Union

22

23 import torch

@@ -292,7 +293,7 @@ class Gemma2Model(nn.Module):

293 assert intermediate_tensors is not None

294 hidden_states = intermediate_tensors["hidden_states"]

295 residual = intermediate_tensors["residual"]

295 - for layer in self.layers[self.start_layer:self.end_layer]:

296 + for layer in islice(self.layers, self.start_layer, self.end_layer):

297 hidden_states, residual = layer(

298 positions,

299 hidden_states,

vllm/model_executor/models/gemma3.py

+2 -1

@@ -16,6 +16,7 @@

16 # See the License for the specific language governing permissions and

17 # limitations under the License.

18 from collections.abc import Iterable

19 +from itertools import islice

20 from typing import Optional, Union

21

22 import torch

@@ -398,7 +399,7 @@ class Gemma3Model(nn.Module):

399 assert intermediate_tensors is not None

400 hidden_states = intermediate_tensors["hidden_states"]

401 residual = intermediate_tensors["residual"]

401 - for layer in self.layers[self.start_layer:self.end_layer]:

402 + for layer in islice(self.layers, self.start_layer, self.end_layer):

403 hidden_states, residual = layer(

404 positions,

405 hidden_states,

vllm/model_executor/models/glm4_moe.py

+2 -2

@@ -24,6 +24,7 @@

24 """Inference-only GLM-4.5 model compatible with HuggingFace weights."""

25 import typing

26 from collections.abc import Callable, Iterable

27 +from itertools import islice

28 from typing import Any, Optional, Union

29

30 import torch

@@ -440,8 +441,7 @@ class Glm4MoeModel(nn.Module):

441 hidden_states = intermediate_tensors["hidden_states"]

442 residual = intermediate_tensors["residual"]

443

443 - for i in range(self.start_layer, self.end_layer):

444 - layer = self.layers[i]

444 + for layer in islice(self.layers, self.start_layer, self.end_layer):

445 hidden_states, residual = layer(positions, hidden_states, residual)

446

447 if not get_pp_group().is_last_rank:

vllm/model_executor/models/gpt2.py

+2 -1

@@ -20,6 +20,7 @@

20 # limitations under the License.

21 """Inference-only GPT-2 model compatible with HuggingFace weights."""

22 from collections.abc import Iterable

23 +from itertools import islice

24 from typing import Optional, Union

25

26 import torch

@@ -228,7 +229,7 @@ class GPT2Model(nn.Module):

229 assert intermediate_tensors is not None

230 hidden_states = intermediate_tensors["hidden_states"]

231

231 - for layer in self.h[self.start_layer:self.end_layer]:

232 + for layer in islice(self.h, self.start_layer, self.end_layer):

233 hidden_states = layer(hidden_states)

234

235 if not get_pp_group().is_last_rank:

vllm/model_executor/models/gpt_bigcode.py

+2 -1

@@ -21,6 +21,7 @@

21 # limitations under the License.

22 """Inference-only GPTBigCode model compatible with HuggingFace weights."""

23 from collections.abc import Iterable

24 +from itertools import islice

25 from typing import Optional, Union

26

27 import torch

@@ -246,7 +247,7 @@ class GPTBigCodeModel(nn.Module):

247 else:

248 hidden_states = intermediate_tensors["hidden_states"]

249

249 - for layer in self.h[self.start_layer:self.end_layer]:

250 + for layer in islice(self.h, self.start_layer, self.end_layer):

251 hidden_states = layer(hidden_states)

252

253 if not get_pp_group().is_last_rank:

vllm/model_executor/models/gpt_j.py

+3 -2

@@ -19,6 +19,7 @@

19 # limitations under the License.

20 """Inference-only GPT-J model compatible with HuggingFace weights."""

21 from collections.abc import Iterable

22 +from itertools import islice

23 from typing import Optional, Union

24

25 import torch

@@ -223,7 +224,7 @@ class GPTJModel(nn.Module):

224 hidden_states = self.get_input_embeddings(input_ids)

225 else:

226 hidden_states = intermediate_tensors["hidden_states"]

226 - for layer in self.h[self.start_layer:self.end_layer]:

227 + for layer in islice(self.h, self.start_layer, self.end_layer):

228 hidden_states = layer(position_ids, hidden_states)

229 if not get_pp_group().is_last_rank:

230 return IntermediateTensors({"hidden_states": hidden_states})

@@ -336,4 +337,4 @@ class GPTJForCausalLM(nn.Module, SupportsPP):

337 def load_weights(self, weights: Iterable[tuple[str,

338 torch.Tensor]]) -> set[str]:

339 loader = AutoWeightsLoader(self)

339 - return loader.load_weights(weights)

\ No newline at end of file

340 + return loader.load_weights(weights)

vllm/model_executor/models/gpt_neox.py

+2 -1

@@ -19,6 +19,7 @@

19 # limitations under the License.

20 """Inference-only GPT-NeoX model compatible with HuggingFace weights."""

21 from collections.abc import Iterable

22 +from itertools import islice

23 from typing import Optional, Union

24

25 import torch

@@ -235,7 +236,7 @@ class GPTNeoXModel(nn.Module):

236 hidden_states = self.get_input_embeddings(input_ids)

237 else:

238 hidden_states = intermediate_tensors["hidden_states"]

238 - for layer in self.layers[self.start_layer:self.end_layer]:

239 + for layer in islice(self.layers, self.start_layer, self.end_layer):

240 hidden_states = layer(position_ids, hidden_states)

241 if not get_pp_group().is_last_rank:

242 return IntermediateTensors({"hidden_states": hidden_states})

vllm/model_executor/models/granite.py

+2 -1

@@ -24,6 +24,7 @@

24 # limitations under the License.

25 """Inference-only IBM Granite model compatible with HuggingFace weights."""

26 from collections.abc import Iterable

27 +from itertools import islice

28 from typing import Any, Optional, Union

29

30 import torch

@@ -316,7 +317,7 @@ class GraniteModel(nn.Module):

317 hidden_states = intermediate_tensors["hidden_states"]

318 residual = intermediate_tensors["residual"]

319

319 - for layer in self.layers[self.start_layer:self.end_layer]:

320 + for layer in islice(self.layers, self.start_layer, self.end_layer):

321 hidden_states = layer(positions, hidden_states)

322

323 if not get_pp_group().is_last_rank:

vllm/model_executor/models/granitemoe.py

+2 -1

@@ -24,6 +24,7 @@

24 # limitations under the License.

25 """Inference-only GraniteMoe model."""

26 from collections.abc import Iterable

27 +from itertools import islice

28 from typing import Any, Optional

29

30 import torch

@@ -303,7 +304,7 @@ class GraniteMoeModel(nn.Module):

304 assert intermediate_tensors is not None

305 hidden_states = intermediate_tensors["hidden_states"]

306 residual = intermediate_tensors["residual"]

306 - for layer in self.layers[self.start_layer:self.end_layer]:

307 + for layer in islice(self.layers, self.start_layer, self.end_layer):

308 hidden_states = layer(positions, hidden_states)

309 if not get_pp_group().is_last_rank:

310 return IntermediateTensors({

vllm/model_executor/models/granitemoehybrid.py

+1 -2

@@ -397,8 +397,7 @@ class GraniteMoeHybridModel(nn.Module):

397 residual = intermediate_tensors["residual"]

398

399 num_attn = 0

400 - for i in range(len(self.layers)):

401 - layer = self.layers[i]

400 + for i, layer in enumerate(self.layers):

401 if isinstance(layer, GraniteMoeHybridAttentionDecoderLayer):

402 num_attn += 1

403

vllm/model_executor/models/granitemoeshared.py

+2 -2

@@ -6,6 +6,7 @@ The architecture is the same as granitemoe but with the addition of shared

6 experts.

7 """

8 from collections.abc import Iterable

9 +from itertools import islice

10 from typing import Optional

11

12 import torch

@@ -200,8 +201,7 @@ class GraniteMoeSharedModel(nn.Module):

201 assert intermediate_tensors is not None

202 hidden_states = intermediate_tensors["hidden_states"]

203 residual = intermediate_tensors["residual"]

203 - for i in range(self.start_layer, self.end_layer):

204 - layer = self.layers[i]

204 + for layer in islice(self.layers, self.start_layer, self.end_layer):

205 hidden_states = layer(positions, hidden_states)

206 if not get_pp_group().is_last_rank:

207 return IntermediateTensors({

vllm/model_executor/models/grok1.py

+2 -2

@@ -23,6 +23,7 @@

23 # limitations under the License.

24 """Inference-only Grok1 model."""

25 from collections.abc import Iterable

26 +from itertools import islice

27 from typing import Optional, Union

28

29 import torch

@@ -347,8 +348,7 @@ class Grok1Model(nn.Module):

348 hidden_states = intermediate_tensors["hidden_states"]

349 residual = intermediate_tensors["residual"]

350

350 - for i in range(self.start_layer, self.end_layer):

351 - layer = self.layers[i]

351 + for layer in islice(self.layers, self.start_layer, self.end_layer):

352 hidden_states, residual = layer(positions, hidden_states, residual)

353

354 if not get_pp_group().is_last_rank:

vllm/model_executor/models/internlm2.py

+2 -1

@@ -3,6 +3,7 @@

3

4 from collections.abc import Iterable

5 from functools import partial

6 +from itertools import islice

7 from typing import Any, Optional, Union

8

9 import torch

@@ -297,7 +298,7 @@ class InternLM2Model(nn.Module):

298 assert intermediate_tensors is not None

299 hidden_states = intermediate_tensors["hidden_states"]

300 residual = intermediate_tensors["residual"]

300 - for layer in self.layers[self.start_layer:self.end_layer]:

301 + for layer in islice(self.layers, self.start_layer, self.end_layer):

302 hidden_states, residual = layer(positions, hidden_states, residual)

303 if not get_pp_group().is_last_rank:

304 return IntermediateTensors({

vllm/model_executor/models/internlm2_ve.py

+2 -1

@@ -1,6 +1,7 @@

1 # SPDX-License-Identifier: Apache-2.0

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

3

4 +from itertools import islice

5 from typing import Optional, Union

6

7 import torch

@@ -123,7 +124,7 @@ class InternLM2VEModel(InternLM2Model):

124 assert intermediate_tensors is not None

125 hidden_states = intermediate_tensors["hidden_states"]

126 residual = intermediate_tensors["residual"]

126 - for layer in self.layers[self.start_layer:self.end_layer]:

127 + for layer in islice(self.layers, self.start_layer, self.end_layer):

128 hidden_states, residual = layer(

129 positions,

130 hidden_states,

vllm/model_executor/models/jais.py

+2 -1

@@ -23,6 +23,7 @@

23

24 import math

25 from collections.abc import Iterable

26 +from itertools import islice

27 from typing import Optional, Union

28

29 import torch

@@ -276,7 +277,7 @@ class JAISModel(nn.Module):

277 assert intermediate_tensors is not None

278 hidden_states = intermediate_tensors["hidden_states"]

279

279 - for layer in self.h[self.start_layer:self.end_layer]:

280 + for layer in islice(self.h, self.start_layer, self.end_layer):

281 hidden_states = layer(hidden_states)

282

283 if not get_pp_group().is_last_rank:

vllm/model_executor/models/jamba.py

+2 -1

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

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

3 """Inference-only Jamba model."""

4 from collections.abc import Iterable

5 +from itertools import islice

6 from typing import Optional

7

8 import torch

@@ -350,7 +351,7 @@ class JambaModel(nn.Module):

351

352 kv_cache_index = 0

353 mamba_cache_index = 0

353 - for layer in self.layers[self.start_layer:self.end_layer]:

354 + for layer in islice(self.layers, self.start_layer, self.end_layer):

355 layer_mamba_cache_params = None

356 if isinstance(layer, JambaAttentionDecoderLayer):

357 kv_cache_index += 1

vllm/model_executor/models/lfm2.py

+3 -2

@@ -1,6 +1,7 @@

1 # SPDX-License-Identifier: Apache-2.0

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

3 from collections.abc import Iterable

4 +from itertools import islice

5 from typing import Any, Optional

6

7 import torch

@@ -374,7 +375,7 @@ class Lfm2Model(nn.Module):

375 hidden_states = intermediate_tensors["hidden_states"]

376 residual = intermediate_tensors["residual"]

377

377 - for layer in self.layers[self.start_layer:self.end_layer]:

378 + for layer in islice(self.layers, self.start_layer, self.end_layer):

379 hidden_states, residual = layer(

380 positions=positions,

381 hidden_states=hidden_states,

@@ -554,4 +555,4 @@ class Lfm2ForCausalLM(nn.Module, HasInnerState, SupportsLoRA, SupportsPP,

555 skip_prefixes=(["lm_head."]

556 if self.config.tie_word_embeddings else None),

557 )

557 - return loader.load_weights(weights)

\ No newline at end of file

558 + return loader.load_weights(weights)

vllm/model_executor/models/llama.py

+2 -1

@@ -24,6 +24,7 @@

24 # limitations under the License.

25 """Inference-only LLaMA model compatible with HuggingFace weights."""

26 from collections.abc import Iterable

27 +from itertools import islice

28 from typing import Any, Optional, Union

29

30 import torch

@@ -383,7 +384,7 @@ class LlamaModel(nn.Module):

384

385 aux_hidden_states = []

386 for idx, layer in enumerate(

386 - self.layers[self.start_layer:self.end_layer]):

387 + islice(self.layers, self.start_layer, self.end_layer)):

388 if idx in self.aux_hidden_state_layers:

389 aux_hidden_states.append(hidden_states + residual)

390 hidden_states, residual = layer(positions, hidden_states, residual)

vllm/model_executor/models/mamba2.py

+1 -3

@@ -164,9 +164,7 @@ class Mamba2Model(nn.Module):

164 # v1 get mamba2_metadata from forward_context

165 mamba2_metadata = None

166

167 - for i in range(len(self.layers)):

168 - layer = self.layers[i]

169 -

167 + for i, layer in enumerate(self.layers):

168 hidden_states, residual = layer(

169 positions=positions,

170 hidden_states=hidden_states,

vllm/model_executor/models/mimo.py

+2 -1

@@ -26,6 +26,7 @@

26 # limitations under the License.

27 """Inference-only MiMo model compatible with HuggingFace weights."""

28 from collections.abc import Iterable

29 +from itertools import islice

30 from typing import Optional, Union

31

32 import torch

@@ -74,7 +75,7 @@ class MiMoModel(Qwen2Model):

75 assert intermediate_tensors is not None

76 hidden_states = intermediate_tensors["hidden_states"]

77 residual = intermediate_tensors["residual"]

77 - for layer in self.layers[self.start_layer:self.end_layer]:

78 + for layer in islice(self.layers, self.start_layer, self.end_layer):

79 hidden_states, residual = layer(

80 positions,

81 hidden_states,

vllm/model_executor/models/minicpm.py

+2 -1

@@ -25,6 +25,7 @@

25 """Inference-only MiniCPM model compatible with HuggingFace weights."""

26 import math

27 from collections.abc import Iterable

28 +from itertools import islice

29 from typing import Any, Optional, Union

30

31 import torch

@@ -414,7 +415,7 @@ class MiniCPMModel(nn.Module):

415 hidden_states = intermediate_tensors["hidden_states"]

416 residual = intermediate_tensors["residual"]

417

417 - for layer in self.layers[self.start_layer:self.end_layer]:

418 + for layer in islice(self.layers, self.start_layer, self.end_layer):

419 hidden_states, residual = layer(

420 positions,

421 hidden_states,

vllm/model_executor/models/minimax_text_01.py

+2 -2

@@ -3,6 +3,7 @@

3 """Inference-only MiniMaxText01 model."""

4 import math

5 from collections.abc import Iterable

6 +from itertools import islice

7 from typing import TYPE_CHECKING, Optional, Union

8

9 if TYPE_CHECKING:

@@ -1019,8 +1020,7 @@ class MiniMaxText01Model(nn.Module):

1020

1021 minimax_cache_index = 0

1022

1022 - for i in range(self.start_layer, self.end_layer):

1023 - layer = self.layers[i]

1023 + for layer in islice(self.layers, self.start_layer, self.end_layer):

1024 _caches = None

1025 if not envs.VLLM_USE_V1 and isinstance(

1026 layer.self_attn, MiniMaxText01LinearAttention):

vllm/model_executor/models/mixtral.py

+2 -1

@@ -24,6 +24,7 @@

24 # limitations under the License.

25 """Inference-only Mixtral model."""

26 from collections.abc import Iterable

27 +from itertools import islice

28 from typing import Optional, Union

29

30 import torch

@@ -307,7 +308,7 @@ class MixtralModel(nn.Module):

308 assert intermediate_tensors is not None

309 hidden_states = intermediate_tensors["hidden_states"]

310 residual = intermediate_tensors["residual"]

310 - for layer in self.layers[self.start_layer:self.end_layer]:

311 + for layer in islice(self.layers, self.start_layer, self.end_layer):

312 hidden_states, residual = layer(positions, hidden_states, residual)

313 if not get_pp_group().is_last_rank:

314 return IntermediateTensors({

vllm/model_executor/models/mixtral_quant.py

+2 -1

@@ -24,6 +24,7 @@

24 # limitations under the License.

25 """Inference-only Mixtral model."""

26 from collections.abc import Iterable

27 +from itertools import islice

28 from typing import Optional, Union

29

30 import numpy as np

@@ -346,7 +347,7 @@ class MixtralModel(nn.Module):

347 assert intermediate_tensors is not None

348 hidden_states = intermediate_tensors["hidden_states"]

349 residual = intermediate_tensors["residual"]

349 - for layer in self.layers[self.start_layer:self.end_layer]:

350 + for layer in islice(self.layers, self.start_layer, self.end_layer):

351 hidden_states, residual = layer(positions, hidden_states, residual)

352 if not get_pp_group().is_last_rank:

353 return IntermediateTensors({

vllm/model_executor/models/molmo.py

+2 -1

@@ -5,6 +5,7 @@ import math

5 from collections.abc import Iterable, Mapping, Sequence

6 from dataclasses import dataclass

7 from functools import cached_property, partial

8 +from itertools import islice

9 from typing import Annotated, Optional, Union

10

11 import numpy as np

@@ -842,7 +843,7 @@ class MolmoModel(nn.Module, SupportsQuant):

843 residual = intermediate_tensors["residual"]

844

845 # Apply blocks one-by-one.

845 - for layer in self.layers[self.start_layer:self.end_layer]:

846 + for layer in islice(self.layers, self.start_layer, self.end_layer):

847 hidden_states, residual = layer(

848 positions,

849 hidden_states,

vllm/model_executor/models/mpt.py

+2 -1

@@ -4,6 +4,7 @@

4 # Adapted from https://huggingface.co/mosaicml/mpt-7b/tree/main

5 import math

6 from collections.abc import Iterable

7 +from itertools import islice

8 from typing import Optional, Union

9

10 import torch

@@ -260,7 +261,7 @@ class MPTModel(nn.Module):

261 assert intermediate_tensors is not None

262 hidden_states = intermediate_tensors["hidden_states"]

263

263 - for block in self.blocks[self.start_layer:self.end_layer]:

264 + for block in islice(self.blocks, self.start_layer, self.end_layer):

265 hidden_states = block(position_ids, hidden_states)

266 if not get_pp_group().is_last_rank:

267 return IntermediateTensors({"hidden_states": hidden_states})

vllm/model_executor/models/nemotron.py

+2 -1

@@ -24,6 +24,7 @@

24 # limitations under the License.

25 """Inference-only Nemotron model compatible with HuggingFace weights."""

26 from collections.abc import Iterable

27 +from itertools import islice

28 from typing import Any, Optional, Union

29

30 import torch

@@ -353,7 +354,7 @@ class NemotronModel(nn.Module):

354 hidden_states = intermediate_tensors["hidden_states"]

355 residual = intermediate_tensors["residual"]

356

356 - for layer in self.layers[self.start_layer:self.end_layer]:

357 + for layer in islice(self.layers, self.start_layer, self.end_layer):

358 hidden_states, residual = layer(positions, hidden_states, residual)

359

360 if not get_pp_group().is_last_rank:

vllm/model_executor/models/nemotron_h.py

+1 -2

@@ -399,8 +399,7 @@ class NemotronHModel(nn.Module):

399

400 residual = None

401 num_non_mamba_layers = 0

402 - for i in range(len(self.layers)):

403 - layer = self.layers[i]

402 + for i, layer in enumerate(self.layers):

403 layer_mamba_cache_params = None

404 if isinstance(layer,

405 NemotronHMambaDecoderLayer) and mamba_cache_params:

vllm/model_executor/models/nemotron_nas.py

+2 -2

@@ -24,6 +24,7 @@

24 # limitations under the License.

25 """Inference-only deci model compatible with HuggingFace weights."""

26 from collections.abc import Iterable

27 +from itertools import islice

28 from typing import Any, Optional, Union

29

30 import torch

@@ -287,8 +288,7 @@ class DeciModel(nn.Module):

288 residual = intermediate_tensors["residual"]

289

290 kv_cache_index = 0

290 - for i in range(self.start_layer, self.end_layer):

291 - layer = self.layers[i]

291 + for layer in islice(self.layers, self.start_layer, self.end_layer):

292 if not layer._is_no_op_attention:

293 hidden_states, residual = layer(positions, hidden_states,

294 residual)

vllm/model_executor/models/olmo.py

+2 -1

@@ -24,6 +24,7 @@

24 # limitations under the License.

25 """Inference-only OLMo model compatible with HuggingFace weights."""

26 from collections.abc import Iterable

27 +from itertools import islice

28 from typing import Optional, Union

29

30 import torch

@@ -280,7 +281,7 @@ class OlmoModel(nn.Module):

281 hidden_states = intermediate_tensors["hidden_states"]

282

283 # Apply blocks one-by-one.

283 - for layer in self.layers[self.start_layer:self.end_layer]:

284 + for layer in islice(self.layers, self.start_layer, self.end_layer):

285 # shape: (batch_size, seq_len, d_model)

286 hidden_states = layer(positions, hidden_states)

287

vllm/model_executor/models/olmo2.py

+2 -1

@@ -26,6 +26,7 @@

26

27 from collections.abc import Iterable

28 from functools import partial

29 +from itertools import islice

30 from typing import Optional, Union

31

32 import torch

@@ -305,7 +306,7 @@ class Olmo2Model(nn.Module):

306 assert isinstance(hidden_states, torch.Tensor)

307

308 # Apply blocks one-by-one.

308 - for layer in self.layers[self.start_layer:self.end_layer]:

309 + for layer in islice(self.layers, self.start_layer, self.end_layer):

310 # shape: (batch_size, seq_len, d_model)

311 hidden_states = layer(positions, hidden_states)

312

vllm/model_executor/models/olmoe.py

+2 -1

@@ -15,6 +15,7 @@

15 """Inference-only OLMoE model compatible with HuggingFace weights."""

16 from collections.abc import Iterable

17 from functools import partial

18 +from itertools import islice

19 from typing import Any, Optional, Union

20

21 import torch

@@ -314,7 +315,7 @@ class OlmoeModel(nn.Module):

315 hidden_states = intermediate_tensors["hidden_states"]

316 residual = intermediate_tensors["residual"]

317

317 - for layer in self.layers[self.start_layer:self.end_layer]:

318 + for layer in islice(self.layers, self.start_layer, self.end_layer):

319 hidden_states, residual = layer(

320 positions,

321 hidden_states,

vllm/model_executor/models/opt.py

+2 -1

@@ -20,6 +20,7 @@

20 # limitations under the License.

21 """Inference-only OPT model compatible with HuggingFace weights."""

22 from collections.abc import Iterable

23 +from itertools import islice

24 from typing import Optional, Union

25

26 import torch

@@ -269,7 +270,7 @@ class OPTDecoder(nn.Module):

270 assert intermediate_tensors is not None

271 hidden_states = intermediate_tensors["hidden_states"]

272

272 - for layer in self.layers[self.start_layer:self.end_layer]:

273 + for layer in islice(self.layers, self.start_layer, self.end_layer):

274 hidden_states = layer(hidden_states)

275

276 if not get_pp_group().is_last_rank:

vllm/model_executor/models/orion.py

+2 -1

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

7 # LICENSE: https://huggingface.co/OrionStarAI/Orion-14B-Base/blob/main/LICENSE

8 """Inference-only Orion-14B model compatible with HuggingFace weights."""

9 from collections.abc import Iterable

10 +from itertools import islice

11 from typing import Any, Optional, Union

12

13 import torch

@@ -252,7 +253,7 @@ class OrionModel(nn.Module):

253 else:

254 assert intermediate_tensors is not None

255 hidden_states = intermediate_tensors["hidden_states"]

255 - for layer in self.layers[self.start_layer:self.end_layer]:

256 + for layer in islice(self.layers, self.start_layer, self.end_layer):

257 hidden_states = layer(positions, hidden_states)

258 if not get_pp_group().is_last_rank:

259 return IntermediateTensors({

vllm/model_executor/models/persimmon.py

+2 -1

@@ -23,6 +23,7 @@

23 # limitations under the License.

24 """Inference-only persimmon model compatible with HuggingFace weights."""

25 from collections.abc import Iterable

26 +from itertools import islice

27 from typing import Optional, Union

28

29 import torch

@@ -255,7 +256,7 @@ class PersimmonModel(nn.Module):

256 else:

257 assert intermediate_tensors is not None

258 hidden_states = intermediate_tensors["hidden_states"]

258 - for layer in self.layers[self.start_layer:self.end_layer]:

259 + for layer in islice(self.layers, self.start_layer, self.end_layer):

260 hidden_states = layer(positions, hidden_states)

261 if not get_pp_group().is_last_rank:

262 return IntermediateTensors({"hidden_states": hidden_states})

vllm/model_executor/models/phi.py

+2 -1

@@ -38,6 +38,7 @@

38 # OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

39 """Inference-only Phi-1.5 model compatible with HuggingFace weights."""

40 from collections.abc import Iterable

41 +from itertools import islice

42 from typing import Optional, Union

43

44 import torch

@@ -240,7 +241,7 @@ class PhiModel(nn.Module):

241 else:

242 assert intermediate_tensors is not None

243 hidden_states = intermediate_tensors["hidden_states"]

243 - for layer in self.layers[self.start_layer:self.end_layer]:

244 + for layer in islice(self.layers, self.start_layer, self.end_layer):

245 hidden_states = layer(positions, hidden_states)

246

247 if not get_pp_group().is_last_rank:

vllm/model_executor/models/phimoe.py

+2 -1

@@ -24,6 +24,7 @@

24 # limitations under the License.

25 """Inference-only PhiMoE model."""

26 from collections.abc import Iterable

27 +from itertools import islice

28 from typing import Optional, Union

29

30 import torch

@@ -500,7 +501,7 @@ class PhiMoEModel(nn.Module):

501 hidden_states = intermediate_tensors["hidden_states"]

502 residual = intermediate_tensors["residual"]

503

503 - for layer in self.layers[self.start_layer:self.end_layer]:

504 + for layer in islice(self.layers, self.start_layer, self.end_layer):

505 hidden_states, residual = layer(

506 positions,

507 hidden_states,

vllm/model_executor/models/plamo2.py

+2 -1

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

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

3 """Inference-only PLaMo2 model."""

4 from collections.abc import Iterable

5 +from itertools import islice

6 from typing import Optional

7

8 import torch

@@ -614,7 +615,7 @@ class Plamo2Decoder(torch.nn.Module):

615 mamba2_metadata: Mamba2Metadata,

616 ) -> torch.Tensor:

617 mamba_cache_index = 0

617 - for layer in self.layers[self.start_layer:self.end_layer]:

618 + for layer in islice(self.layers, self.start_layer, self.end_layer):

619 layer_mamba_cache_params = None

620 if layer.is_mamba:

621 layer_mamba_cache_params = mamba_cache_params.at_layer_idx(

vllm/model_executor/models/qwen.py

+2 -1

@@ -8,6 +8,7 @@

8 """Inference-only QWen model compatible with HuggingFace weights."""

9 import json

10 from collections.abc import Iterable

11 +from itertools import islice

12 from typing import Any, Optional, Union

13

14 import torch

@@ -234,7 +235,7 @@ class QWenModel(nn.Module):

235 hidden_states = intermediate_tensors["hidden_states"]

236 residual = intermediate_tensors["residual"]

237

237 - for layer in self.h[self.start_layer:self.end_layer]:

238 + for layer in islice(self.h, self.start_layer, self.end_layer):

239 hidden_states, residual = layer(

240 positions,

241 hidden_states,

vllm/model_executor/models/qwen2.py

+2 -1

@@ -25,6 +25,7 @@

25 # limitations under the License.

26 """Inference-only Qwen2 model compatible with HuggingFace weights."""

27 from collections.abc import Iterable

28 +from itertools import islice

29 from typing import Any, Optional, Union

30

31 import torch

@@ -358,7 +359,7 @@ class Qwen2Model(nn.Module):

359

360 aux_hidden_states = []

361 for idx, layer in enumerate(

361 - self.layers[self.start_layer:self.end_layer]):

362 + islice(self.layers, self.start_layer, self.end_layer)):

363 if idx in self.aux_hidden_state_layers:

364 aux_hidden_states.append(hidden_states + residual)

365 hidden_states, residual = layer(positions, hidden_states, residual)

vllm/model_executor/models/qwen2_moe.py

+2 -1

@@ -25,6 +25,7 @@

25 # limitations under the License.

26 """Inference-only Qwen2MoE model compatible with HuggingFace weights."""

27 from collections.abc import Iterable

28 +from itertools import islice

29 from typing import Any, Optional, Union

30

31 import torch

@@ -381,7 +382,7 @@ class Qwen2MoeModel(nn.Module):

382 assert intermediate_tensors is not None

383 hidden_states = intermediate_tensors["hidden_states"]

384 residual = intermediate_tensors["residual"]

384 - for layer in self.layers[self.start_layer:self.end_layer]:

385 + for layer in islice(self.layers, self.start_layer, self.end_layer):

386 hidden_states, residual = layer(positions, hidden_states, residual)

387 if not get_pp_group().is_last_rank:

388 return IntermediateTensors({

vllm/model_executor/models/qwen3_moe.py

+2 -2

@@ -24,6 +24,7 @@

24 """Inference-only Qwen3MoE model compatible with HuggingFace weights."""

25 import typing

26 from collections.abc import Callable, Iterable

27 +from itertools import islice

28 from typing import Any, Optional, Union

29

30 import torch

@@ -420,8 +421,7 @@ class Qwen3MoeModel(nn.Module):

421 assert intermediate_tensors is not None

422 hidden_states = intermediate_tensors["hidden_states"]

423 residual = intermediate_tensors["residual"]

423 - for i in range(self.start_layer, self.end_layer):

424 - layer = self.layers[i]

424 + for layer in islice(self.layers, self.start_layer, self.end_layer):

425 hidden_states, residual = layer(positions, hidden_states, residual)

426 if not get_pp_group().is_last_rank:

427 return IntermediateTensors({

vllm/model_executor/models/seed_oss.py

+2 -1

@@ -23,6 +23,7 @@

23 # limitations under the License.

24 """Inference-only SeedOss model compatible with HuggingFace weights."""

25 from collections.abc import Iterable

26 +from itertools import islice

27 from typing import Optional, Union

28

29 import torch

@@ -340,7 +341,7 @@ class SeedOssModel(nn.Module):

341 assert intermediate_tensors is not None

342 hidden_states = intermediate_tensors["hidden_states"]

343 residual = intermediate_tensors["residual"]

343 - for layer in self.layers[self.start_layer:self.end_layer]:

344 + for layer in islice(self.layers, self.start_layer, self.end_layer):

345 hidden_states, residual = layer(

346 positions,

347 hidden_states,

vllm/model_executor/models/stablelm.py

+2 -1

@@ -22,6 +22,7 @@

22 """Inference-only StabeLM (https://github.com/Stability-AI/StableLM)

23 model compatible with HuggingFace weights."""

24 from collections.abc import Iterable

25 +from itertools import islice

26 from typing import Optional, Union

27

28 import torch

@@ -247,7 +248,7 @@ class StableLMEpochModel(nn.Module):

248 else:

249 assert intermediate_tensors is not None

250 hidden_states = intermediate_tensors["hidden_states"]

250 - for layer in self.layers[self.start_layer:self.end_layer]:

251 + for layer in islice(self.layers, self.start_layer, self.end_layer):

252 hidden_states, residual = layer(positions, hidden_states)

253 if not get_pp_group().is_last_rank:

254 return IntermediateTensors({"hidden_states": hidden_states})

vllm/model_executor/models/starcoder2.py

+2 -1

@@ -21,6 +21,7 @@

21 # limitations under the License.

22 """ PyTorch Starcoder2 model."""

23 from collections.abc import Iterable

24 +from itertools import islice

25 from typing import Optional, Union

26

27 import torch

@@ -250,7 +251,7 @@ class Starcoder2Model(nn.Module):

251 else:

252 assert intermediate_tensors is not None

253 hidden_states = intermediate_tensors["hidden_states"]

253 - for layer in self.layers[self.start_layer:self.end_layer]:

254 + for layer in islice(self.layers, self.start_layer, self.end_layer):

255 hidden_states = layer(positions, hidden_states)

256 if not get_pp_group().is_last_rank:

257 return IntermediateTensors({"hidden_states": hidden_states})

vllm/model_executor/models/step3_text.py

+2 -2

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

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

3 """Inference-only Jurassic model."""

4 from collections.abc import Iterable

5 +from itertools import islice

6 from typing import Any, Optional

7

8 import torch

@@ -346,8 +347,7 @@ class Step3TextModel(nn.Module):

347 hidden_states = intermediate_tensors["hidden_states"]

348 residual = intermediate_tensors["residual"]

349

349 - for i in range(self.start_layer, self.end_layer):

350 - layer = self.layers[i]

350 + for layer in islice(self.layers, self.start_layer, self.end_layer):

351 hidden_states, residual = layer(positions, hidden_states, residual)

352

353 if not get_pp_group().is_last_rank: