I tried to save the model with the below code but failed. Planning to use unsloth/Llama-3.2-11B-Vision-Instruct as a base model to fine-tune a new model.

!pip install unsloth
model, tokenizer = FastVisionModel.from_pretrained(
    "unsloth/Llama-3.2-11B-Vision-Instruct",
    load_in_4bit = False, 
    use_gradient_checkpointing = "unsloth", 
)
... 
if True: 
    model.save_pretrained_merged(
        "unsloth_finetune", 
        tokenizer,
        save_method = "merged_16bit",)

with error

File ~/.local/lib/python3.10/site-packages/unsloth/save.py:2357, in unsloth_generic_save_pretrained_merged(self, save_directory, tokenizer, save_method, push_to_hub, token, is_main_process, state_dict, save_function, max_shard_size, safe_serialization, variant, save_peft_format, tags, temporary_location, maximum_memory_usage)
   [2355](~/.local/lib/python3.10/site-packages/unsloth/save.py:2355) arguments["model"] = self
   [2356](~/.local/lib/python3.10/site-packages/unsloth/save.py:2356) del arguments["self"]
-> [2357](~/.local/lib/python3.10/site-packages/unsloth/save.py:2357) unsloth_generic_save(**arguments)
   [2358](~/.local/lib/python3.10/site-packages/unsloth/save.py:2358) for _ in range(3):
   [2359](~/.local/lib/python3.10/site-packages/unsloth/save.py:2359)     gc.collect()

File ~/.local/lib/python3.10/site-packages/torch/utils/_contextlib.py:116, in context_decorator.<locals>.decorate_context(*args, **kwargs)
    [113](~/.local/lib/python3.10/site-packages/torch/utils/_contextlib.py:113) @functools.wraps(func)
...
--> [472](~/.local/lib/python3.10/site-packages/unsloth_zoo/saving_utils.py:472) try_save_directory = temp_file.name
    [474]~/.local/lib/python3.10/site-packages/unsloth_zoo/saving_utils.py:474) total, used, free = shutil.disk_usage(save_directory)
    [475](~/.local/lib/python3.10/site-packages/unsloth_zoo/saving_utils.py:475) free = int(free*0.95)

AttributeError: 'NoneType' object has no attribute 'name'

All the help is appreciated.

I tried to save the model with the below code but failed. Planning to use unsloth/Llama-3.2-11B-Vision-Instruct as a base model to fine-tune a new model.

!pip install unsloth
model, tokenizer = FastVisionModel.from_pretrained(
    "unsloth/Llama-3.2-11B-Vision-Instruct",
    load_in_4bit = False, 
    use_gradient_checkpointing = "unsloth", 
)
... 
if True: 
    model.save_pretrained_merged(
        "unsloth_finetune", 
        tokenizer,
        save_method = "merged_16bit",)

with error

File ~/.local/lib/python3.10/site-packages/unsloth/save.py:2357, in unsloth_generic_save_pretrained_merged(self, save_directory, tokenizer, save_method, push_to_hub, token, is_main_process, state_dict, save_function, max_shard_size, safe_serialization, variant, save_peft_format, tags, temporary_location, maximum_memory_usage)
   [2355](~/.local/lib/python3.10/site-packages/unsloth/save.py:2355) arguments["model"] = self
   [2356](~/.local/lib/python3.10/site-packages/unsloth/save.py:2356) del arguments["self"]
-> [2357](~/.local/lib/python3.10/site-packages/unsloth/save.py:2357) unsloth_generic_save(**arguments)
   [2358](~/.local/lib/python3.10/site-packages/unsloth/save.py:2358) for _ in range(3):
   [2359](~/.local/lib/python3.10/site-packages/unsloth/save.py:2359)     gc.collect()

File ~/.local/lib/python3.10/site-packages/torch/utils/_contextlib.py:116, in context_decorator.<locals>.decorate_context(*args, **kwargs)
    [113](~/.local/lib/python3.10/site-packages/torch/utils/_contextlib.py:113) @functools.wraps(func)
...
--> [472](~/.local/lib/python3.10/site-packages/unsloth_zoo/saving_utils.py:472) try_save_directory = temp_file.name
    [474]~/.local/lib/python3.10/site-packages/unsloth_zoo/saving_utils.py:474) total, used, free = shutil.disk_usage(save_directory)
    [475](~/.local/lib/python3.10/site-packages/unsloth_zoo/saving_utils.py:475) free = int(free*0.95)

AttributeError: 'NoneType' object has no attribute 'name'

All the help is appreciated.

• python
• model
• fine-tuning

Share Improve this question Follow edited Mar 24 at 6:07 John asked Mar 23 at 3:23 JohnJohn 83322 gold badges1111 silver badges2727 bronze badges

Add a comment  | 

1 Answer 1

Sorted by: Reset to default 0

# Eğer Jupyter/Colab kullanıyorsan bu satır kalabilir
!pip install unsloth

from unsloth import FastVisionModel

# Modeli yükle (4-bit modda, belleği daha az kullanır)
model, tokenizer = FastVisionModel.from_pretrained(
    "unsloth/Llama-3.2-11B-Vision-Instruct",
    load_in_4bit=True,  # RTX 3060 için zorunlu gibi
    use_gradient_checkpointing=True,  # "unsloth" yerine True
)

# Eğitilmiş modeli kaydet (16-bit birleşik olarak)
model.save_pretrained_merged(
    save_directory="unsloth_finetune",  # Klasör adı
    tokenizer=tokenizer,
    save_method="merged_16bit",  # Hata alırsan bunu "merged" yapabilirsin
)