TypeError: unhashable type: 'set' when loading LoRA adapter on Llama-3.1-8B #10859

Reminder

• I have read the above rules and searched the existing issues.

System Info

Hi,
Environment:

• Windows 11
• RTX 5080
• CUDA 12.8
• torch 2.12.0.dev20260408+cu128
• peft 0.18.1
• accelerate 1.11.0

Training works successfully.

LoRA adapter is created successfully:

• adapter_model.safetensors
• adapter_config.json

Training finishes successfully.

When loading the adapter in Chat, the following error occurs:

Traceback:
File "...peft_model.py", line 1415, in load_adapter
max_memory = get_balanced_memory(

File "...accelerate\utils\modeling.py", line 1019
if set(no_split_children.keys()) == set(no_split_module_classes):

TypeError: unhashable type: 'set'

The same error occurs in:

• LLaMA-Factory 0.9.6.dev0
• LLaMA-Factory 0.9.5

Base model:
Llama-3.1-8B

Regards

Reproduction

Steps to reproduce

1. Train a LoRA adapter on Llama-3.1-8B.

2. Training completes successfully and the following files are created:

   adapter_model.safetensors
   adapter_config.json

3. Open Chat in LLaMA-Factory.

4. Load the base model:
   C:\Models\Llama-3.1-8B

5. Select the trained LoRA adapter:
   C:\LLM\LLaMA-Factory\saves\Llama-3.1-8B\lora\train_2026-09-30-08-58-34

6. Click "Load Model".

7. The adapter loading fails with:

   TypeError: unhashable type: 'set'

Traceback:

File "...peft_model.py", line 1415, in load_adapter
max_memory = get_balanced_memory(

File "...accelerate\utils\modeling.py", line 1019, in get_balanced_memory
if set(no_split_children.keys()) == set(no_split_module_classes):
TypeError: unhashable type: 'set'

Others

Additional information

This issue is reproduced on a very recent NVIDIA RTX 5080 GPU
(Blackwell architecture, compute capability 12.0).

The following setup works correctly:

• CUDA 12.8
• torch 2.12.0.dev20260408+cu128
• LoRA training completes successfully
• adapter_model.safetensors is created successfully

The failure only occurs when loading the LoRA adapter in Chat.

What was tested

• Training works successfully.
• LoRA adapter is created successfully.
• adapter_model.safetensors exists.
• adapter_config.json exists.
• The issue happens only when loading the adapter.
• Reproduced with:
  • LLaMA-Factory 0.9.6.dev0
  • LLaMA-Factory 0.9.5