ayo/ipex-llm
1
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

[NPU ] fix load logic of glm-edge models (#12698)

Browse source
This commit is contained in:
binbin Deng 2025-01-10 16:08:37 +08:00 committed by GitHub
parent 584c1c5373
commit da8bcb7db1
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 7 additions and 3 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

10
python/llm/src/ipex_llm/transformers/npu_model.py
View file

@ -182,13 +182,17 @@ class _BaseAutoModelClass:
if hasattr(model, "config") and model.config.model_type == "glm": if hasattr(model, "config") and model.config.model_type == "glm":
# convert to llama structure # convert to llama structure
from .npu_models.glm_edge import convert_config, load_weights, convert_state_dict from .npu_models.glm_edge import convert_config, load_weights, convert_state_dict
import json
original_path = model.config._name_or_path original_path = model.config._name_or_path
del model del model
with open(os.path.join(original_path, "config.json")) as f: original_config, _ = PretrainedConfig.get_config_dict(original_path)
original_config = json.load(f)
config = convert_config(original_config) config = convert_config(original_config)
if not os.path.isdir(original_path):
# all model files are already cached
from transformers.utils.hub import cached_file
resolved_file = cached_file(original_path, "config.json")
original_path = os.path.dirname(resolved_file)
original_state_dict = load_weights(original_path) original_state_dict = load_weights(original_path)
new_dict, _ = convert_state_dict(original_state_dict, config, new_dict, _ = convert_state_dict(original_state_dict, config,
original_config.get("partial_rotary_factor", 1.0), original_config.get("partial_rotary_factor", 1.0),