This repository aims to compare the available open-source GEMM / GEMV kernels using a mixed precision scheme int4 / fp16, with per-group quantization.
- [x] https://github.com/qwopqwop200/GPTQ-for-LLaMa
- [x] https://github.com/turboderp/exllama
- [x] https://github.com/PanQiWei/AutoGPTQ
- [ ] https://github.com/NVIDIA/FasterTransformer (only per-channel quantization, per-block not open-sourced, so not compared but sounds promising as based on CUTLASS)
- [ ] AWQ implem https://github.com/mit-han-lab/llm-awq/tree/main/awq/kernels
- [ ] Probably missing others
On A100-SXM4-80GB & Intel Xeon Platinum 8275CL CPU + CUDA 11.7/11.8 (should be rerun in docker):
| m | n | k | implementation | act_order | Time (ms/op) | Max mem (MB) |
|---|---|---|---|---|---|---|
| 1 | 8192 | 8192 | baseline | True | 0.0937 | 177.6845 |
| 1 | 8192 | 8192 | gptqforllama | True | 0.2038 | 69.8450 |
| 1 | 8192 | 8192 | exllama | False | 0.0681 | 34.9143 |
| 1 | 8192 | 8192 | exllama | True | 0.0675 | 34.9471 |
| 1 | 8192 | 8192 | autogptq-triton | True | 0.3990 | 69.8450 |
| 1 | 8192 | 8192 | autogptq-cuda-old | False | 0.0831 | 71.9585 |
| 1 | 8192 | 8192 | autogptq-cuda | True | 0.1546 | 69.8778 |
On RTX 4090 + AMD Ryzen 9 7950X CPU + CUDA 11.8:
TODO
On A10G + AMD EPYC 7R32 CPU + CUDA 11.8 (docker):
| m | n | k | implementation | act_order | Time (ms/op) | Max mem (MB) |
|---|---|---|---|---|---|---|
| 1 | 8192 | 8192 | baseline | True | 0.2891 | 177.6845 |
| 1 | 8192 | 8192 | gptqforllama | True | 0.1746 | 69.8450 |
| 1 | 8192 | 8192 | autogptq-triton | True | 0.2963 | 69.8450 |
| 1 | 8192 | 8192 | autogptq-cuda-old | False | 0.0979 | 71.9585 |
| 1 | 8192 | 8192 | autogptq-cuda | True | 0.1483 | 69.8778 |
| 1 | 8192 | 8192 | exllama | False | 0.0842 | 34.9143 |
| 1 | 8192 | 8192 | exllama | True | 0.0839 | 34.9471 |
A=m * k, B=k * n, compute C= A*B^T
It can be a good idea to first lock the GPU frequency, see https://github.com/NVIDIA/cutlass/issues/430#issuecomment-1069535238
Run exllama in exllama env:
CUDA_VISIBLE_DEVICES=0 python run_benchmark.py --m 1 --n 8192 --k 8192 --group_size 128 --exllama-path ../exllama --act-order yes
Run gptqforllama in gptqforllama env:
CUDA_VISIBLE_DEVICES=0 python run_benchmark.py --m 1 --n 8192 --k 8192 --group_size 128 --gptqforllama-path ../GPTQ-for-LLaMa --act-order yes
Run AutoGPTQ (specify --autogptq-implem {triton, cuda-old, cuda}):
CUDA_VISIBLE_DEVICES=0 python run_benchmark.py --m 1 --n 8192 --k 8192 --group_size 128 --autogptq-path ../AutoGPTQ/ --autogptq-implem triton --act-order yes
Run PyTorch fp16 * fp16 baseline:
CUDA_VISIBLE_DEVICES=0 python run_benchmark.py --m 1 --n 8192 --k 8192 --group_size 128 --baseline
Follow https://stackoverflow.com/a/61737404 and
docker build -f Dockerfile --build-arg USER_ID=$(id -u) --build-arg GROUP_ID=$(id -g) -t container-q4f16 .
and
docker run --gpus device=0 -it --rm container-q4f16:latest /bin/bash run.sh