423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2026-08-17 13:35:44 +08:00
Code Issues Packages Projects Releases Wiki Activity

[model] handle unsloth model loading fallback during checkpoint resume (#7156) (#10551)

Browse Source
This commit is contained in:
Co-Cl2
2026-06-09 01:01:01 +08:00
committed by GitHub
parent 0b7aaf8f6a
commit 9ca4026efe
3 changed files with 23 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
src/llamafactory/model/loader.py
View File

@@ -34,7 +34,7 @@ from .adapter import init_adapter
from .model_utils.liger_kernel import apply_liger_kernel
from .model_utils.misc import register_autoclass
from .model_utils.mod import convert_pretrained_model_to_mod, load_mod_pretrained_model
from .model_utils.unsloth import load_unsloth_pretrained_model
from .model_utils.unsloth import load_unsloth_pretrained_model, load_unsloth_peft_model
from .model_utils.valuehead import load_valuehead_params
from .patcher import patch_config, patch_model, patch_processor, patch_tokenizer, patch_valuehead_model
@@ -142,14 +142,13 @@ def load_model(
apply_liger_kernel(config, model_args, is_trainable, require_logits=(finetuning_args.stage not in ["pt", "sft"]))
model = None
lazy_load = False
if model_args.use_unsloth:
if model_args.adapter_name_or_path is not None:
lazy_load = True
model = load_unsloth_peft_model(config, model_args, finetuning_args, is_trainable=is_trainable)
elif is_trainable:
model = load_unsloth_pretrained_model(config, model_args, finetuning_args)
if model is None and not lazy_load:
if model is None:
init_kwargs["config"] = config
init_kwargs["pretrained_model_name_or_path"] = model_args.model_name_or_path
init_kwargs["torch_dtype"] = "auto"
@@ -176,9 +175,8 @@ def load_model(
if model_args.mixture_of_depths == "convert":
model = convert_pretrained_model_to_mod(model, config, model_args)
if not lazy_load:
patch_model(model, tokenizer, model_args, is_trainable, add_valuehead)
register_autoclass(config, model, tokenizer)
patch_model(model, tokenizer, model_args, is_trainable, add_valuehead)
register_autoclass(config, model, tokenizer)
model = init_adapter(config, model, model_args, finetuning_args, is_trainable)