Skip to content

4-bit llada #54

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

4-bit llada #54

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
13 changes: 13 additions & 0 deletions README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -63,6 +63,19 @@ Please refer to Appendix B.5. of our [paper](https://arxiv.org/abs/2502.09992) f

We provide the code for evaluation using the open-source library lm-evaluation-harness. To begin, please install `lm_eval==0.4.5` and refer to `evaluation/eval.sh` for the specific commands.

## Quantization

We provide the code for quantization using [GPTQModel](https://github.com/ModelCloud/GPTQModel) in the `quantization` folder.
- [LLaDA-8B-Base-GPTQ-4bit](https://huggingface.co/FunAGI/LLaDA-8B-Base-gptqmodel-4bit)
- [LLaDA-8B-Instruct-GPTQ-4bit](https://huggingface.co/FunAGI/LLaDA-8B-Instruct-gptqmodel-4bit)

Based on the published evaluation code, we have evaluated the quantized **base model**. The results are as follows:

| Dataset | GPTQ-4bit | FP16 |
|----------------|------------|--------|
| MMLU | 65.20 | 65.90 |
| CMMLU | 69.23 | 69.90 |
| ARC-Challenge | 45.48 | 47.90 |

## FAQ
Here, we address some common questions about LLaDA.
Expand Down
25 changes: 25 additions & 0 deletions quantization/eval.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,25 @@
MODEL_PATH=/home/chentianqi/deepseek/GPTQModel/models/LLaDA-8B-Instruc-default:w4g128

#Instruct model
CUDA_VISIBLE_DEVICES=0 python eval_llada.py --tasks gsm8k --num_fewshot 4 --model llada_dist \
--apply_chat_template \
--batch_size 1 --model_args "model_path=$MODEL_PATH,cfg=0.0,is_check_greedy=False,max_length=256,block_length=8,steps=256"

#Base model
accelerate launch eval_llada.py --tasks gpqa_main_n_shot --num_fewshot 5 --model llada_dist --batch_size 8 --model_args model_path=$MODEL_PATH,cfg=0.5,is_check_greedy=False,mc_num=128

accelerate launch eval_llada.py --tasks truthfulqa_mc2 --num_fewshot 0 --model llada_dist --batch_size 8 --model_args model_path=$MODEL_PATH,cfg=2.0,is_check_greedy=False,mc_num=128

accelerate launch eval_llada.py --tasks arc_challenge --num_fewshot 0 --model llada_dist --batch_size 8 --model_args model_path=$MODEL_PATH,cfg=0.5,is_check_greedy=False,mc_num=128

accelerate launch eval_llada.py --tasks hellaswag --num_fewshot 0 --model llada_dist --batch_size 8 --model_args model_path=$MODEL_PATH,cfg=0.5,is_check_greedy=False,mc_num=128

accelerate launch eval_llada.py --tasks winogrande --num_fewshot 5 --model llada_dist --batch_size 8 --model_args model_path=$MODEL_PATH,cfg=0.0,is_check_greedy=False,mc_num=128

accelerate launch eval_llada.py --tasks piqa --num_fewshot 0 --model llada_dist --batch_size 8 --model_args model_path=$MODEL_PATH,cfg=0.5,is_check_greedy=False,mc_num=128

accelerate launch eval_llada.py --tasks mmlu --num_fewshot 5 --model llada_dist --batch_size 1 --model_args model_path=$MODEL_PATH,cfg=0.0,is_check_greedy=False,mc_num=1

accelerate launch eval_llada.py --tasks cmmlu --num_fewshot 5 --model llada_dist --batch_size 1 --model_args model_path=$MODEL_PATH,cfg=0.0,is_check_greedy=False,mc_num=1

accelerate launch eval_llada.py --tasks ceval-valid --num_fewshot 5 --model llada_dist --batch_size 1 --model_args model_path=$MODEL_PATH,cfg=0.0,is_check_greedy=False,mc_num=1
Loading