21 lines
1,005 B
Markdown
21 lines
1,005 B
Markdown
# Running LLM Finetuning using BigDL-LLM on Intel GPU
|
|
|
|
This folder contains examples of running different training mode with BigDL-LLM on Intel GPU:
|
|
|
|
- [LoRA](LoRA): examples of running LoRA finetuning
|
|
- [QLoRA](QLoRA): examples of running QLoRA finetuning
|
|
- [QA-LoRA](QA-LoRA): examples of running QA-LoRA finetuning
|
|
- [ReLora](ReLora): examples of running ReLora finetuning
|
|
- [DPO](DPO): examples of running DPO finetuning
|
|
- [common](common): common templates and utility classes in finetuning examples
|
|
- [HF-PEFT](HF-PEFT): run finetuning on Intel GPU using Hugging Face PEFT code without modification
|
|
|
|
|
|
## Troubleshooting
|
|
- If you fail to finetune on multi cards because of following error message:
|
|
```bash
|
|
RuntimeError: oneCCL: comm_selector.cpp:57 create_comm_impl: EXCEPTION: ze_data was not initialized
|
|
```
|
|
Please try `sudo apt install level-zero-dev` to fix it.
|
|
|
|
- Please raise the system open file limit using `ulimit -n 1048576`. Otherwise, there may exist error `Too many open files`.
|