[Models] Improve iteration over layers (#19497)

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

vllm/model_executor/models/arcee.py

@@ -9,6 +9,7 @@
 # activation.
 
 from collections.abc import Iterable
+from itertools import islice
 from typing import Any, Optional, Union
 
 import torch
@@ -243,7 +244,7 @@ class ArceeModel(nn.Module):
 
         aux_hidden_states: list[torch.Tensor] = []
         for idx, layer in enumerate(
-                self.layers[self.start_layer:self.end_layer]):
+                islice(self.layers, self.start_layer, self.end_layer)):
             if idx in self.aux_hidden_state_layers:
                 aux_hidden_states.append(
                     hidden_states +

vllm/model_executor/models/arctic.py

@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Inference-only Snowflake Arctic model."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -403,7 +404,7 @@ class ArcticModel(nn.Module):
         else:
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states = layer(positions, hidden_states)
         if not get_pp_group().is_last_rank:
             return IntermediateTensors({"hidden_states": hidden_states})

vllm/model_executor/models/baichuan.py

@@ -22,6 +22,7 @@
 """Inference-only BaiChuan model compatible with HuggingFace weights."""
 import math
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -309,7 +310,7 @@ class BaiChuanModel(nn.Module):
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states, residual = layer(
                 positions,
                 hidden_states,

vllm/model_executor/models/bailing_moe.py

@@ -24,6 +24,7 @@
 # limitations under the License.
 """Inference-only BailingMoE model compatible with HuggingFace weights."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -359,8 +360,7 @@ class BailingMoeModel(nn.Module):
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
 
-        for i in range(self.start_layer, self.end_layer):
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
-            layer = self.layers[i]
             hidden_states, residual = layer(
                 hidden_states,
                 position_ids,

vllm/model_executor/models/bamba.py

@@ -345,8 +345,7 @@ class BambaModel(nn.Module):
 
         residual = None
         num_attn = 0
-        for i in range(len(self.layers)):
+        for i, layer in enumerate(self.layers):
-            layer = self.layers[i]
             if isinstance(layer, BambaAttentionDecoderLayer):
                 num_attn += 1
 

vllm/model_executor/models/bloom.py

@@ -20,6 +20,7 @@
 """Inference-only BLOOM model compatible with HuggingFace weights."""
 import math
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -273,7 +274,7 @@ class BloomModel(nn.Module):
         else:
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
-        for layer in self.h[self.start_layer:self.end_layer]:
+        for layer in islice(self.h, self.start_layer, self.end_layer):
             hidden_states = layer(position_ids, hidden_states)
         if not get_pp_group().is_last_rank:
             return IntermediateTensors({"hidden_states": hidden_states})

vllm/model_executor/models/chameleon.py

@@ -3,6 +3,7 @@
 
 from collections.abc import Iterable, Mapping, Sequence
 from functools import cached_property
+from itertools import islice
 from typing import Annotated, Any, Literal, Optional, Union
 
 import torch
@@ -914,7 +915,7 @@ class ChameleonModel(nn.Module):
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states, residual = layer(
                 positions,
                 hidden_states,

vllm/model_executor/models/chatglm.py

@@ -5,6 +5,7 @@
 """Inference-only ChatGLM model compatible with THUDM weights."""
 import json
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -281,7 +282,7 @@ class GLMTransformer(nn.Module):
         hidden_states: torch.Tensor,
         position_ids: torch.Tensor,
     ) -> Union[torch.Tensor, IntermediateTensors]:
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states = layer(hidden_states=hidden_states,
                                   position_ids=position_ids)
 

vllm/model_executor/models/commandr.py

@@ -23,6 +23,7 @@
 # This file is based on the LLama model definition file in transformers
 """PyTorch Cohere model."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -322,7 +323,7 @@ class CohereModel(nn.Module):
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states, residual = layer(
                 positions,
                 hidden_states,

vllm/model_executor/models/dbrx.py

@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -359,7 +360,7 @@ class DbrxModel(nn.Module):
         else:
             assert intermediate_tensors
             hidden_states = intermediate_tensors["hidden_states"]
-        for block in self.blocks[self.start_layer:self.end_layer]:
+        for block in islice(self.blocks, self.start_layer, self.end_layer):
             hidden_states = block(position_ids, hidden_states)
         if not get_pp_group().is_last_rank:
             return IntermediateTensors({"hidden_states": hidden_states})

vllm/model_executor/models/deepseek.py

@@ -24,6 +24,7 @@
 # limitations under the License.
 """Inference-only Deepseek model."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Any, Optional, Union
 
 import torch
@@ -377,7 +378,7 @@ class DeepseekModel(nn.Module):
         else:
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states, residual = layer(positions, hidden_states, residual)
         if not get_pp_group().is_last_rank:
             return IntermediateTensors({

vllm/model_executor/models/deepseek_v2.py

@@ -25,6 +25,7 @@
 """Inference-only DeepseekV2/DeepseekV3 model."""
 import typing
 from collections.abc import Callable, Iterable
+from itertools import islice
 from typing import Any, Optional, Union
 
 import torch
@@ -712,7 +713,7 @@ class DeepseekV2Model(nn.Module):
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
 
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states, residual = layer(positions, hidden_states, residual)
 
         if not get_pp_group().is_last_rank:

vllm/model_executor/models/dots1.py

@@ -25,6 +25,7 @@
 # limitations under the License.
 """Inference-only dots1 model."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Any, Optional, Union
 
 import torch
@@ -391,7 +392,7 @@ class Dots1Model(nn.Module):
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states, residual = layer(
                 positions,
                 hidden_states,

vllm/model_executor/models/ernie45_moe.py

@@ -23,6 +23,7 @@
 # limitations under the License.
 """Inference-only ErineMoE model compatible with HuggingFace weights."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Any, Optional, Union
 
 import torch
@@ -419,8 +420,7 @@ class Ernie4_5_MoeModel(nn.Module):
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
 
-        for i in range(self.start_layer, self.end_layer):
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
-            layer = self.layers[i]
             hidden_states, residual = layer(positions, hidden_states, residual)
 
         if not get_pp_group().is_last_rank:

vllm/model_executor/models/ernie45_vl_moe.py

@@ -23,6 +23,7 @@
 # limitations under the License.
 """Inference-only Erine VL model compatible with HuggingFace weights."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Any, Optional, Union
 
 import torch
@@ -508,8 +509,7 @@ class Ernie4_5_VLMoeModel(nn.Module):
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
 
-        for i in range(self.start_layer, self.end_layer):
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
-            layer = self.layers[i]
             hidden_states, residual = layer(positions, hidden_states, residual,
                                             visual_token_mask, **kwargs)
 

vllm/model_executor/models/exaone.py

@@ -26,6 +26,7 @@
 """Inference-only Exaone model compatible with HuggingFace weights."""
 
 from collections.abc import Iterable
+from itertools import islice
 from typing import Any, Optional, Union
 
 import torch
@@ -371,7 +372,7 @@ class ExaoneModel(nn.Module):
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
 
-        for layer in self.h[self.start_layer:self.end_layer]:
+        for layer in islice(self.h, self.start_layer, self.end_layer):
             hidden_states, residual = layer(
                 positions,
                 hidden_states,

vllm/model_executor/models/exaone4.py

@@ -22,6 +22,7 @@
 """Inference-only Exaone model compatible with HuggingFace weights."""
 
 from collections.abc import Iterable
+from itertools import islice
 from typing import Any, Optional, Union
 
 import torch
@@ -354,7 +355,7 @@ class Exaone4Model(nn.Module):
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
 
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states, residual = layer(
                 positions,
                 hidden_states,

vllm/model_executor/models/falcon.py

@@ -22,6 +22,7 @@
 
 import math
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -389,7 +390,7 @@ class FalconModel(nn.Module):
                 hidden_states = self.get_input_embeddings(input_ids)
         else:
             hidden_states = intermediate_tensors["hidden_states"]
-        for layer in self.h[self.start_layer:self.end_layer]:
+        for layer in islice(self.h, self.start_layer, self.end_layer):
             hidden_states = layer(positions, hidden_states)
         if not get_pp_group().is_last_rank:
             return IntermediateTensors({"hidden_states": hidden_states})

vllm/model_executor/models/gemma.py

@@ -18,6 +18,7 @@
 """Inference-only Gemma model compatible with HuggingFace weights."""
 from collections.abc import Iterable
 from functools import cache
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -308,7 +309,7 @@ class GemmaModel(nn.Module):
         else:
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states, residual = layer(
                 positions,
                 hidden_states,

vllm/model_executor/models/gemma2.py

@@ -17,6 +17,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -292,7 +293,7 @@ class Gemma2Model(nn.Module):
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states, residual = layer(
                 positions,
                 hidden_states,

vllm/model_executor/models/gemma3.py

@@ -16,6 +16,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -398,7 +399,7 @@ class Gemma3Model(nn.Module):
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states, residual = layer(
                 positions,
                 hidden_states,

vllm/model_executor/models/glm4_moe.py

@@ -24,6 +24,7 @@
 """Inference-only GLM-4.5 model compatible with HuggingFace weights."""
 import typing
 from collections.abc import Callable, Iterable
+from itertools import islice
 from typing import Any, Optional, Union
 
 import torch
@@ -440,8 +441,7 @@ class Glm4MoeModel(nn.Module):
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
 
-        for i in range(self.start_layer, self.end_layer):
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
-            layer = self.layers[i]
             hidden_states, residual = layer(positions, hidden_states, residual)
 
         if not get_pp_group().is_last_rank:

vllm/model_executor/models/gpt2.py

@@ -20,6 +20,7 @@
 # limitations under the License.
 """Inference-only GPT-2 model compatible with HuggingFace weights."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -228,7 +229,7 @@ class GPT2Model(nn.Module):
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
 
-        for layer in self.h[self.start_layer:self.end_layer]:
+        for layer in islice(self.h, self.start_layer, self.end_layer):
             hidden_states = layer(hidden_states)
 
         if not get_pp_group().is_last_rank:

vllm/model_executor/models/gpt_bigcode.py

@@ -21,6 +21,7 @@
 # limitations under the License.
 """Inference-only GPTBigCode model compatible with HuggingFace weights."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -246,7 +247,7 @@ class GPTBigCodeModel(nn.Module):
         else:
             hidden_states = intermediate_tensors["hidden_states"]
 
-        for layer in self.h[self.start_layer:self.end_layer]:
+        for layer in islice(self.h, self.start_layer, self.end_layer):
             hidden_states = layer(hidden_states)
 
         if not get_pp_group().is_last_rank:

vllm/model_executor/models/gpt_j.py

@@ -19,6 +19,7 @@
 # limitations under the License.
 """Inference-only GPT-J model compatible with HuggingFace weights."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -223,7 +224,7 @@ class GPTJModel(nn.Module):
                 hidden_states = self.get_input_embeddings(input_ids)
         else:
             hidden_states = intermediate_tensors["hidden_states"]
-        for layer in self.h[self.start_layer:self.end_layer]:
+        for layer in islice(self.h, self.start_layer, self.end_layer):
             hidden_states = layer(position_ids, hidden_states)
         if not get_pp_group().is_last_rank:
             return IntermediateTensors({"hidden_states": hidden_states})

vllm/model_executor/models/gpt_neox.py

@@ -19,6 +19,7 @@
 # limitations under the License.
 """Inference-only GPT-NeoX model compatible with HuggingFace weights."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -235,7 +236,7 @@ class GPTNeoXModel(nn.Module):
                 hidden_states = self.get_input_embeddings(input_ids)
         else:
             hidden_states = intermediate_tensors["hidden_states"]
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states = layer(position_ids, hidden_states)
         if not get_pp_group().is_last_rank:
             return IntermediateTensors({"hidden_states": hidden_states})

vllm/model_executor/models/granite.py

@@ -24,6 +24,7 @@
 # limitations under the License.
 """Inference-only IBM Granite model compatible with HuggingFace weights."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Any, Optional, Union
 
 import torch
@@ -316,7 +317,7 @@ class GraniteModel(nn.Module):
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
 
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states = layer(positions, hidden_states)
 
         if not get_pp_group().is_last_rank:

vllm/model_executor/models/granitemoe.py

@@ -24,6 +24,7 @@
 # limitations under the License.
 """Inference-only GraniteMoe model."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Any, Optional
 
 import torch
@@ -303,7 +304,7 @@ class GraniteMoeModel(nn.Module):
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states = layer(positions, hidden_states)
         if not get_pp_group().is_last_rank:
             return IntermediateTensors({

vllm/model_executor/models/granitemoehybrid.py

@@ -397,8 +397,7 @@ class GraniteMoeHybridModel(nn.Module):
             residual = intermediate_tensors["residual"]
 
         num_attn = 0
-        for i in range(len(self.layers)):
+        for i, layer in enumerate(self.layers):
-            layer = self.layers[i]
             if isinstance(layer, GraniteMoeHybridAttentionDecoderLayer):
                 num_attn += 1
 

vllm/model_executor/models/granitemoeshared.py

@@ -6,6 +6,7 @@ The architecture is the same as granitemoe but with the addition of shared
 experts.
 """
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional
 
 import torch
@@ -200,8 +201,7 @@ class GraniteMoeSharedModel(nn.Module):
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
-        for i in range(self.start_layer, self.end_layer):
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
-            layer = self.layers[i]
             hidden_states = layer(positions, hidden_states)
         if not get_pp_group().is_last_rank:
             return IntermediateTensors({

vllm/model_executor/models/grok1.py

@@ -23,6 +23,7 @@
 # limitations under the License.
 """Inference-only Grok1 model."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -347,8 +348,7 @@ class Grok1Model(nn.Module):
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
 
-        for i in range(self.start_layer, self.end_layer):
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
-            layer = self.layers[i]
             hidden_states, residual = layer(positions, hidden_states, residual)
 
         if not get_pp_group().is_last_rank:

vllm/model_executor/models/internlm2.py

@@ -3,6 +3,7 @@
 
 from collections.abc import Iterable
 from functools import partial
+from itertools import islice
 from typing import Any, Optional, Union
 
 import torch
@@ -297,7 +298,7 @@ class InternLM2Model(nn.Module):
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states, residual = layer(positions, hidden_states, residual)
         if not get_pp_group().is_last_rank:
             return IntermediateTensors({

vllm/model_executor/models/internlm2_ve.py

@@ -1,6 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -123,7 +124,7 @@ class InternLM2VEModel(InternLM2Model):
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states, residual = layer(
                 positions,
                 hidden_states,

vllm/model_executor/models/jais.py

@@ -23,6 +23,7 @@
 
 import math
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -276,7 +277,7 @@ class JAISModel(nn.Module):
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
 
-        for layer in self.h[self.start_layer:self.end_layer]:
+        for layer in islice(self.h, self.start_layer, self.end_layer):
             hidden_states = layer(hidden_states)
 
         if not get_pp_group().is_last_rank:

vllm/model_executor/models/jamba.py

@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Inference-only Jamba model."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional
 
 import torch
@@ -350,7 +351,7 @@ class JambaModel(nn.Module):
 
         kv_cache_index = 0
         mamba_cache_index = 0
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             layer_mamba_cache_params = None
             if isinstance(layer, JambaAttentionDecoderLayer):
                 kv_cache_index += 1

vllm/model_executor/models/lfm2.py

@@ -1,6 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 from collections.abc import Iterable
+from itertools import islice
 from typing import Any, Optional
 
 import torch
@@ -374,7 +375,7 @@ class Lfm2Model(nn.Module):
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
 
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states, residual = layer(
                 positions=positions,
                 hidden_states=hidden_states,

vllm/model_executor/models/llama.py

@@ -24,6 +24,7 @@
 # limitations under the License.
 """Inference-only LLaMA model compatible with HuggingFace weights."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Any, Optional, Union
 
 import torch
@@ -383,7 +384,7 @@ class LlamaModel(nn.Module):
 
         aux_hidden_states = []
         for idx, layer in enumerate(
-                self.layers[self.start_layer:self.end_layer]):
+                islice(self.layers, self.start_layer, self.end_layer)):
             if idx in self.aux_hidden_state_layers:
                 aux_hidden_states.append(hidden_states + residual)
             hidden_states, residual = layer(positions, hidden_states, residual)

vllm/model_executor/models/mamba2.py

@@ -164,9 +164,7 @@ class Mamba2Model(nn.Module):
             # v1 get mamba2_metadata from forward_context
             mamba2_metadata = None
 
-        for i in range(len(self.layers)):
+        for i, layer in enumerate(self.layers):
-            layer = self.layers[i]
-
             hidden_states, residual = layer(
                 positions=positions,
                 hidden_states=hidden_states,

vllm/model_executor/models/mimo.py

@@ -26,6 +26,7 @@
 # limitations under the License.
 """Inference-only MiMo model compatible with HuggingFace weights."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -74,7 +75,7 @@ class MiMoModel(Qwen2Model):
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states, residual = layer(
                 positions,
                 hidden_states,

vllm/model_executor/models/minicpm.py

@@ -25,6 +25,7 @@
 """Inference-only MiniCPM model compatible with HuggingFace weights."""
 import math
 from collections.abc import Iterable
+from itertools import islice
 from typing import Any, Optional, Union
 
 import torch
@@ -414,7 +415,7 @@ class MiniCPMModel(nn.Module):
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
 
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states, residual = layer(
                 positions,
                 hidden_states,

vllm/model_executor/models/minimax_text_01.py

@@ -3,6 +3,7 @@
 """Inference-only MiniMaxText01 model."""
 import math
 from collections.abc import Iterable
+from itertools import islice
 from typing import TYPE_CHECKING, Optional, Union
 
 if TYPE_CHECKING:
@@ -1019,8 +1020,7 @@ class MiniMaxText01Model(nn.Module):
 
         minimax_cache_index = 0
 
-        for i in range(self.start_layer, self.end_layer):
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
-            layer = self.layers[i]
             _caches = None
             if not envs.VLLM_USE_V1 and isinstance(
                     layer.self_attn, MiniMaxText01LinearAttention):

vllm/model_executor/models/mixtral.py

@@ -24,6 +24,7 @@
 # limitations under the License.
 """Inference-only Mixtral model."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -307,7 +308,7 @@ class MixtralModel(nn.Module):
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states, residual = layer(positions, hidden_states, residual)
         if not get_pp_group().is_last_rank:
             return IntermediateTensors({

vllm/model_executor/models/mixtral_quant.py

@@ -24,6 +24,7 @@
 # limitations under the License.
 """Inference-only Mixtral model."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import numpy as np
@@ -346,7 +347,7 @@ class MixtralModel(nn.Module):
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states, residual = layer(positions, hidden_states, residual)
         if not get_pp_group().is_last_rank:
             return IntermediateTensors({

vllm/model_executor/models/molmo.py

@@ -5,6 +5,7 @@ import math
 from collections.abc import Iterable, Mapping, Sequence
 from dataclasses import dataclass
 from functools import cached_property, partial
+from itertools import islice
 from typing import Annotated, Optional, Union
 
 import numpy as np
@@ -842,7 +843,7 @@ class MolmoModel(nn.Module, SupportsQuant):
             residual = intermediate_tensors["residual"]
 
         # Apply blocks one-by-one.
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states, residual = layer(
                 positions,
                 hidden_states,

vllm/model_executor/models/mpt.py

@@ -4,6 +4,7 @@
 # Adapted from https://huggingface.co/mosaicml/mpt-7b/tree/main
 import math
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -260,7 +261,7 @@ class MPTModel(nn.Module):
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
 
-        for block in self.blocks[self.start_layer:self.end_layer]:
+        for block in islice(self.blocks, self.start_layer, self.end_layer):
             hidden_states = block(position_ids, hidden_states)
         if not get_pp_group().is_last_rank:
             return IntermediateTensors({"hidden_states": hidden_states})

vllm/model_executor/models/nemotron.py

@@ -24,6 +24,7 @@
 # limitations under the License.
 """Inference-only Nemotron model compatible with HuggingFace weights."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Any, Optional, Union
 
 import torch
@@ -353,7 +354,7 @@ class NemotronModel(nn.Module):
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
 
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states, residual = layer(positions, hidden_states, residual)
 
         if not get_pp_group().is_last_rank:

vllm/model_executor/models/nemotron_h.py

@@ -399,8 +399,7 @@ class NemotronHModel(nn.Module):
 
         residual = None
         num_non_mamba_layers = 0
-        for i in range(len(self.layers)):
+        for i, layer in enumerate(self.layers):
-            layer = self.layers[i]
             layer_mamba_cache_params = None
             if isinstance(layer,
                           NemotronHMambaDecoderLayer) and mamba_cache_params:

vllm/model_executor/models/nemotron_nas.py

@@ -24,6 +24,7 @@
 # limitations under the License.
 """Inference-only deci model compatible with HuggingFace weights."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Any, Optional, Union
 
 import torch
@@ -287,8 +288,7 @@ class DeciModel(nn.Module):
             residual = intermediate_tensors["residual"]
 
         kv_cache_index = 0
-        for i in range(self.start_layer, self.end_layer):
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
-            layer = self.layers[i]
             if not layer._is_no_op_attention:
                 hidden_states, residual = layer(positions, hidden_states,
                                                 residual)

vllm/model_executor/models/olmo.py

@@ -24,6 +24,7 @@
 # limitations under the License.
 """Inference-only OLMo model compatible with HuggingFace weights."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -280,7 +281,7 @@ class OlmoModel(nn.Module):
             hidden_states = intermediate_tensors["hidden_states"]
 
         # Apply blocks one-by-one.
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             # shape: (batch_size, seq_len, d_model)
             hidden_states = layer(positions, hidden_states)
 

vllm/model_executor/models/olmo2.py

@@ -26,6 +26,7 @@
 
 from collections.abc import Iterable
 from functools import partial
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -305,7 +306,7 @@ class Olmo2Model(nn.Module):
             assert isinstance(hidden_states, torch.Tensor)
 
         # Apply blocks one-by-one.
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             # shape: (batch_size, seq_len, d_model)
             hidden_states = layer(positions, hidden_states)
 

vllm/model_executor/models/olmoe.py

@@ -15,6 +15,7 @@
 """Inference-only OLMoE model compatible with HuggingFace weights."""
 from collections.abc import Iterable
 from functools import partial
+from itertools import islice
 from typing import Any, Optional, Union
 
 import torch
@@ -314,7 +315,7 @@ class OlmoeModel(nn.Module):
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
 
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states, residual = layer(
                 positions,
                 hidden_states,

vllm/model_executor/models/opt.py

@@ -20,6 +20,7 @@
 # limitations under the License.
 """Inference-only OPT model compatible with HuggingFace weights."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -269,7 +270,7 @@ class OPTDecoder(nn.Module):
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
 
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states = layer(hidden_states)
 
         if not get_pp_group().is_last_rank:

vllm/model_executor/models/orion.py

@@ -7,6 +7,7 @@
 # LICENSE: https://huggingface.co/OrionStarAI/Orion-14B-Base/blob/main/LICENSE
 """Inference-only Orion-14B model compatible with HuggingFace weights."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Any, Optional, Union
 
 import torch
@@ -252,7 +253,7 @@ class OrionModel(nn.Module):
         else:
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states = layer(positions, hidden_states)
         if not get_pp_group().is_last_rank:
             return IntermediateTensors({

vllm/model_executor/models/persimmon.py

@@ -23,6 +23,7 @@
 # limitations under the License.
 """Inference-only persimmon model compatible with HuggingFace weights."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -255,7 +256,7 @@ class PersimmonModel(nn.Module):
         else:
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states = layer(positions, hidden_states)
         if not get_pp_group().is_last_rank:
             return IntermediateTensors({"hidden_states": hidden_states})

vllm/model_executor/models/phi.py

@@ -38,6 +38,7 @@
 # OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 """Inference-only Phi-1.5 model compatible with HuggingFace weights."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -240,7 +241,7 @@ class PhiModel(nn.Module):
         else:
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states = layer(positions, hidden_states)
 
         if not get_pp_group().is_last_rank:

vllm/model_executor/models/phimoe.py

@@ -24,6 +24,7 @@
 # limitations under the License.
 """Inference-only PhiMoE model."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -500,7 +501,7 @@ class PhiMoEModel(nn.Module):
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
 
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states, residual = layer(
                 positions,
                 hidden_states,

vllm/model_executor/models/plamo2.py

@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Inference-only PLaMo2 model."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional
 
 import torch
@@ -614,7 +615,7 @@ class Plamo2Decoder(torch.nn.Module):
         mamba2_metadata: Mamba2Metadata,
     ) -> torch.Tensor:
         mamba_cache_index = 0
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             layer_mamba_cache_params = None
             if layer.is_mamba:
                 layer_mamba_cache_params = mamba_cache_params.at_layer_idx(

vllm/model_executor/models/qwen.py

@@ -8,6 +8,7 @@
 """Inference-only QWen model compatible with HuggingFace weights."""
 import json
 from collections.abc import Iterable
+from itertools import islice
 from typing import Any, Optional, Union
 
 import torch
@@ -234,7 +235,7 @@ class QWenModel(nn.Module):
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
 
-        for layer in self.h[self.start_layer:self.end_layer]:
+        for layer in islice(self.h, self.start_layer, self.end_layer):
             hidden_states, residual = layer(
                 positions,
                 hidden_states,

vllm/model_executor/models/qwen2.py

@@ -25,6 +25,7 @@
 # limitations under the License.
 """Inference-only Qwen2 model compatible with HuggingFace weights."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Any, Optional, Union
 
 import torch
@@ -358,7 +359,7 @@ class Qwen2Model(nn.Module):
 
         aux_hidden_states = []
         for idx, layer in enumerate(
-                self.layers[self.start_layer:self.end_layer]):
+                islice(self.layers, self.start_layer, self.end_layer)):
             if idx in self.aux_hidden_state_layers:
                 aux_hidden_states.append(hidden_states + residual)
             hidden_states, residual = layer(positions, hidden_states, residual)

vllm/model_executor/models/qwen2_moe.py

@@ -25,6 +25,7 @@
 # limitations under the License.
 """Inference-only Qwen2MoE model compatible with HuggingFace weights."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Any, Optional, Union
 
 import torch
@@ -381,7 +382,7 @@ class Qwen2MoeModel(nn.Module):
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states, residual = layer(positions, hidden_states, residual)
         if not get_pp_group().is_last_rank:
             return IntermediateTensors({

vllm/model_executor/models/qwen3_moe.py

@@ -24,6 +24,7 @@
 """Inference-only Qwen3MoE model compatible with HuggingFace weights."""
 import typing
 from collections.abc import Callable, Iterable
+from itertools import islice
 from typing import Any, Optional, Union
 
 import torch
@@ -420,8 +421,7 @@ class Qwen3MoeModel(nn.Module):
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
-        for i in range(self.start_layer, self.end_layer):
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
-            layer = self.layers[i]
             hidden_states, residual = layer(positions, hidden_states, residual)
         if not get_pp_group().is_last_rank:
             return IntermediateTensors({

vllm/model_executor/models/seed_oss.py

@@ -23,6 +23,7 @@
 # limitations under the License.
 """Inference-only SeedOss model compatible with HuggingFace weights."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -340,7 +341,7 @@ class SeedOssModel(nn.Module):
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states, residual = layer(
                 positions,
                 hidden_states,

vllm/model_executor/models/stablelm.py

@@ -22,6 +22,7 @@
 """Inference-only StabeLM (https://github.com/Stability-AI/StableLM)
 model compatible with HuggingFace weights."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -247,7 +248,7 @@ class StableLMEpochModel(nn.Module):
         else:
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states, residual = layer(positions, hidden_states)
         if not get_pp_group().is_last_rank:
             return IntermediateTensors({"hidden_states": hidden_states})

vllm/model_executor/models/starcoder2.py

@@ -21,6 +21,7 @@
 # limitations under the License.
 """ PyTorch Starcoder2 model."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Optional, Union
 
 import torch
@@ -250,7 +251,7 @@ class Starcoder2Model(nn.Module):
         else:
             assert intermediate_tensors is not None
             hidden_states = intermediate_tensors["hidden_states"]
-        for layer in self.layers[self.start_layer:self.end_layer]:
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
             hidden_states = layer(positions, hidden_states)
         if not get_pp_group().is_last_rank:
             return IntermediateTensors({"hidden_states": hidden_states})

vllm/model_executor/models/step3_text.py

@@ -2,6 +2,7 @@
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 """Inference-only Jurassic model."""
 from collections.abc import Iterable
+from itertools import islice
 from typing import Any, Optional
 
 import torch
@@ -346,8 +347,7 @@ class Step3TextModel(nn.Module):
             hidden_states = intermediate_tensors["hidden_states"]
             residual = intermediate_tensors["residual"]
 
-        for i in range(self.start_layer, self.end_layer):
+        for layer in islice(self.layers, self.start_layer, self.end_layer):
-            layer = self.layers[i]
             hidden_states, residual = layer(positions, hidden_states, residual)
 
         if not get_pp_group().is_last_rank: