Merge branch 'upstream' into blip2-dynamic-lm

vllm/model_executor/models/blip.py

@@ -1,6 +1,5 @@
 """Minimal implementation of BlipVisionModel intended to be only used 
 within a vision language model."""
-from array import array
 from typing import Iterable, Optional, Tuple, Union
 
 import torch

vllm/model_executor/models/blip2.py

@@ -1,5 +1,4 @@
 import itertools
-from array import array
 from typing import (Iterable, List, Literal, Mapping, Optional, Tuple,
                     TypedDict, Union)
 

vllm/model_executor/models/llava_next.py

@@ -631,12 +631,7 @@ def sample(
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         # prepare weight iterators for components
-        (
-            vit_weights,
-            mlp_weights,
-            newline_weights,
-            llm_weights,
-        ) = itertools.tee(weights, 4)
+        weights_group = group_weights_with_prefix(weights)
 
         # load vision encoder
         self.vision_tower.load_weights(weights_group["vision_tower"])

vllm/model_executor/models/llava_next_video.py

@@ -448,8 +448,8 @@ def sample(
         return self.language_model.sample(logits, sampling_metadata)
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
-        # prepare weight iterators
-        vit_weights, mlp_weights, llm_weights = itertools.tee(weights, 3)
+        # prepare weight iterators for components
+        weights_group = group_weights_with_prefix(weights)
 
         # load vision encoder
         self.vision_tower.load_weights(weights_group["vision_tower"])