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 3, 11:45 UTC
Revert checkpoint specific workaround in Transformers modelling backend (#42923)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Harry Mellor committed
May 18, 2026 at 17:31 UTC
9537542537728af9fac418ecf1604ad8e8d9ff93
1 file changed
-8
vllm/model_executor/models/transformers/base.py
-8
@@ -353,14 +353,6 @@ class Base(
353
for source, target in ccm.items():
354
orig_to_new_regex[re.compile(source)] = target
355
356
- # Gemma3 checkpoints saved with older Transformers versions include an
357
- # extra `vision_model` level that the current AutoModel no longer has.
358
- vision_tower = getattr(self.model, "vision_tower", None)
359
- if vision_tower is not None and not hasattr(vision_tower, "vision_model"):
360
- orig_to_new_regex[
361
- re.compile(r"^(?:model\.)?vision_tower\.vision_model\.(.+)")
362
- ] = r"model.vision_tower.\1"
363
-
356
# Handle unexpected weights which should be ignored
357
if self.model._keys_to_ignore_on_load_unexpected is not None:
358
for key in self.model._keys_to_ignore_on_load_unexpected: