@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / 93f2c0aa08

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

[Models] Improve iteration over layers (#26425)

Signed-off-by: Lukas Geiger <lukas.geiger94@gmail.com>

Lukas Geiger committed Oct 8, 2025 at 21:48 UTC 93f2c0aa083173256796b607db2bcb19ec132696

8 files changed +23 -22

vllm/model_executor/models/apertus.py

+4 -1

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

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

28 from collections.abc import Iterable

29 +from itertools import islice

30 from typing import Any, Optional, Union

32 import torch

@@ -412,7 +413,9 @@ class ApertusModel(nn.Module):

413 residual = intermediate_tensors["residual"]

414

415 aux_hidden_states = []

415 - for idx, layer in enumerate(self.layers[self.start_layer : self.end_layer]):

416 + for idx, layer in enumerate(

417 + islice(self.layers, self.start_layer, self.end_layer)

418 + ):

419 if idx in self.aux_hidden_state_layers:

420 aux_hidden_states.append(hidden_states + residual)

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

vllm/model_executor/models/falcon_h1.py

+2 -2

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

3 """Inference-only FalconH1 model."""

5 from collections.abc import Iterable

6 +from itertools import islice

7 from typing import Optional

9 import torch

@@ -480,8 +481,7 @@ class FalconH1Model(nn.Module):

481 assert intermediate_tensors is not None

482 hidden_states = intermediate_tensors["hidden_states"]

483

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

484 - layer = self.layers[i]

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

485 hidden_states = layer(

486 positions=positions,

487 hidden_states=hidden_states,

vllm/model_executor/models/hunyuan_v1.py

+5 -6

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

27 import typing

28 from collections.abc import Callable, Iterable

29 +from itertools import islice

30 from typing import Any, Optional, Union

32 import regex as re

@@ -672,8 +673,9 @@ class HunYuanModel(nn.Module):

673

674 cla_factor = _get_cla_factor(self.config)

675 prev_kv_states = None

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

676 - layer = self.layers[i]

676 + for i, layer in enumerate(

677 + islice(self.layers, self.start_layer, self.end_layer)

678 + ):

679 hidden_states, residual, kv_states = layer(

680 positions,

681 hidden_states,

@@ -681,10 +683,7 @@ class HunYuanModel(nn.Module):

683 prev_kv_states,

684 )

685

684 - if (

685 - getattr(self.config, "use_cla", False)

686 - and (i - self.start_layer) % cla_factor == 0

687 - ):

686 + if getattr(self.config, "use_cla", False) and i % cla_factor == 0:

687 prev_kv_states = kv_states

688 else:

689 prev_kv_states = None

vllm/model_executor/models/lfm2_moe.py

+2 -1

@@ -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

7 import torch

@@ -492,7 +493,7 @@ class Lfm2MoeModel(nn.Module):

493 hidden_states = intermediate_tensors["hidden_states"]

494 residual = intermediate_tensors["residual"]

495

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

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

497 hidden_states, residual = layer(

498 positions=positions,

499 hidden_states=hidden_states,

vllm/model_executor/models/longcat_flash.py

+2 -2

@@ -35,6 +35,7 @@

36 import typing

37 from collections.abc import Callable, Iterable

38 +from itertools import islice

39 from typing import Optional, Union

41 import torch

@@ -519,8 +520,7 @@ class FlashModel(nn.Module):

520 hidden_states = intermediate_tensors["hidden_states"]

521 residual = intermediate_tensors["residual"]

522

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

523 - layer = self.layers[i]

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

524 hidden_states, residual = layer(

525 positions,

526 hidden_states,

vllm/model_executor/models/mamba.py

+2 -2

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

3 """PyTorch MAMBA model."""

5 from collections.abc import Iterable

6 +from itertools import islice

7 from typing import Optional

9 import torch

@@ -162,8 +163,7 @@ class MambaModel(nn.Module):

163 hidden_states = intermediate_tensors["hidden_states"]

164 residual = intermediate_tensors["residual"]

165

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

166 - layer = self.layers[i]

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

167 hidden_states, residual = layer(

168 positions=positions, hidden_states=hidden_states, residual=residual

169 )

vllm/model_executor/models/qwen3_vl.py

+3 -4

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

27 from collections.abc import Iterable, Mapping, Sequence

28 from functools import partial

29 +from itertools import islice

30 from typing import Any, Callable, Optional, Union

32 import numpy as np

@@ -1106,11 +1107,9 @@ class Qwen3LLMModel(Qwen3Model):

1107 assert intermediate_tensors is not None

1108 hidden_states = intermediate_tensors["hidden_states"]

1109 residual = intermediate_tensors["residual"]

1109 - for layer_idx, layer in enumerate(

1110 - self.layers[self.start_layer : self.end_layer]

1110 + for layer_idx, layer in islice(

1111 + enumerate(self.layers), self.start_layer, self.end_layer

1112 ):

1112 - layer_idx = layer_idx + self.start_layer

1113 -

1113 hidden_states, residual = layer(

1114 positions,

1115 hidden_states,

vllm/model_executor/models/qwen3_vl_moe.py

+3 -4

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

27 import typing

28 from collections.abc import Iterable

29 +from itertools import islice

30 from typing import Callable, Optional, Union

32 import torch

@@ -103,11 +104,9 @@ class Qwen3MoeLLMModel(Qwen3MoeModel):

104 assert intermediate_tensors is not None

105 hidden_states = intermediate_tensors["hidden_states"]

106 residual = intermediate_tensors["residual"]

106 - for layer_idx, layer in enumerate(

107 - self.layers[self.start_layer : self.end_layer]

107 + for layer_idx, layer in islice(

108 + enumerate(self.layers), self.start_layer, self.end_layer

109 ):

109 - layer_idx = layer_idx + self.start_layer

110 -

110 hidden_states, residual = layer(

111 positions,

112 hidden_states,