Update to transformers v5 (#30566)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Signed-off-by: khluu <khluu000@gmail.com>
Signed-off-by: Kevin H. Luu <khluu000@gmail.com>
Signed-off-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
Co-authored-by: khluu <khluu000@gmail.com>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
Co-authored-by: jiang1.li <jiang1.li@intel.com>
This commit is contained in:
Harry Mellor
2026-04-15 16:29:15 -07:00
committed by GitHub
co-authored by khluu Cyrus Leung jiang1.li
parent 6dc9491406
commit 03f8d3a548
41 changed files with 445 additions and 115 deletions
@@ -265,12 +265,24 @@ class GGUFModelLoader(BaseModelLoader):
GGUF tensor name with suffix (e.g., 'mm.soft_emb_norm.weight')
or None if no mapping found
"""
# In transformers v5, multimodal models (e.g. Gemma3) wrap
# all sub-models under an outer 'model.' attribute, producing
# state_dict keys like 'model.language_model.layers.0...' and
# 'model.vision_tower.vision_model...'. Strip this outer
# prefix so the keys match what gguf-py expects.
if is_multimodal and hf_name.startswith("model."):
hf_name = hf_name[6:] # Remove outer 'model.'
# Strip 'language_model.' prefix for multimodal models - gguf-py
# tensor mappings expect parameter names without this prefix.
# Note: 'model.' prefix should be KEPT for text-only models as
# gguf-py expects it.
if hf_name.startswith("language_model."):
hf_name = hf_name[15:] # Remove 'language_model.'
# Re-add 'model.' prefix because gguf-py text tensor maps
# expect 'model.layers...' format.
if is_multimodal:
hf_name = "model." + hf_name
# Parse parameter name and suffix
if hf_name.endswith((".weight", ".bias")):