Skip to content

[Low-bit optim] Add Llama2-7B finetune benchmarks #746

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

Merged
msaroufim merged 5 commits into from
Sep 2, 2024
Merged

[Low-bit optim] Add Llama2-7B finetune benchmarks #746

Changes from 4 commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
f3ad199
add Llama3.1-8B finetune bench
gau-nernst Aug 25, 2024
1497b3d
Merge branch 'main' into update_optim_bench
gau-nernst Aug 27, 2024
06f0cc9
update doc
gau-nernst Aug 27, 2024
9d56a8c
Merge branch 'pytorch:main' into update_optim_bench
gau-nernst Sep 2, 2024
2de6df0
Update README.md
msaroufim Sep 2, 2024
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
18 changes: 14 additions & 4 deletions torchao/prototype/low_bit_optim/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,11 +30,9 @@ NOTE:

## Benchmarks

Benchmark script for fine-tuning a [timm](https://github.com/huggingface/pytorch-image-models) model on [resisc45](https://huggingface.co/datasets/timm/resisc45) dataset is available at [benchmarks/benchmark_low_bit_adam.py](../../../benchmarks/benchmark_low_bit_adam.py).
Fine-tune [timm](https://github.com/huggingface/pytorch-image-models)'s ViT-H (630M params) on [resisc45](https://huggingface.co/datasets/timm/resisc45) dataset. BF16 AMP, 1 epoch, batch size 8, cosine LR scheduler, 4070Ti SUPER, fixed random seed. Benchmark script is available at [benchmarks/benchmark_low_bit_adam.py](../../../benchmarks/benchmark_low_bit_adam.py).

Results for fine-tuning ViT-H (630M params) with BF16 AMP for 1 epoch, batch size 8, cosine LR scheduler, 4070Ti SUPER, fixed random seed:

Adam impl | max memory (GB) | imgs/s | accuracy
AdamW impl | Max memory (GB) | imgs/s | accuracy
----------------|-----------------|--------|----------
PyTorch (fused) | 12.23 | 41.8 | 94.38
bnb 8-bit | 8.32 | 43.6 | 94.18
Expand All @@ -46,6 +44,18 @@ lpmm 4-bit (*) | 7.74 | 26.6 | 94.25

(*) means rank-1 normalization is used for 2nd optimizer state. Refer to [paper](https://arxiv.org/abs/2309.01507) for more details.

Fine-tune [Llama2-7B](https://huggingface.co/meta-llama/Llama-2-7b) on [Alpaca](https://huggingface.co/datasets/tatsu-lab/alpaca) dataset. Full BF16, 1 epoch, A100, fixed random seed. Benchmark is done with [torchtune](https://github.com/pytorch/torchtune). See [#746](https://github.com/pytorch/ao/pull/746) for more details.

AdamW impl | Max memory (GB) | toks/s | `truthfulqa_mc2` acc | Compile time
-----------------|-----------------|--------|----------------------|-------------
Not fine-tuned | - | - | 38.95 | -
PyTorch (fused) | 52 | ~4500 | 42.12 | ~4 min
bnb 8-bit | 39 | ~4000 | 41.98 | ~4 min
ao 8-bit | 39 | ~4000 | 42.41 | ~12 min
ao 4-bit | 33 | ~3600 | 42.34 | ~4 min

NOTE: lpmm's 4-bit AdamW does not support BF16 weights.

## Optimizer CPU offload

This folder also implements optimizer CPU offload (i.e. ZeRO-Offload) for single GPU training. For multi-GPU training, you can use FSDP's built-in CPU offload.
Expand Down