AdamW optimizer for bfloat16 in PyTorch
This is a version of the AdamW optimizer for use in torch that achieves the same results in ViT training tests as training with the weights in float32 with operations in float32 or bfloat16 (autocast). By keeping your weights in bfloat16, you can save approximately half the weights they would normally take up in memory. It uses stochastic rounding and a correction term to achieve this.
There is a small (~10-20%) performance hit depending on your hardware.
To install:
pip install adamw-bf16
To use:
from adamw_bf16 import AdamWBF16
model = model.to(dtype=torch.bfloat16)
optimizer = AdamWBF16(model.parameters(), ...)
# Train your model
This repository was created using code from the following two projects. It was found that insights from both could be combined to match the performance with the model weights stored in float32.
1 maintainer
Metadata
1 maintainer
Release files for adamw-bf16 0.0.3
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | |
|---|---|---|---|---|
| adamw_bf16-0.0.3.tar.gz | 16.4 kB | Without provenance | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| adamw_bf16-0.0.3-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 33.3 kB
Release files / adamw_bf16-0.0.3.tar.gz
| Download URL | adamw_bf16-0.0.3.tar.gz |
|---|---|
| Size | 16.4 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
826b046e23ebd34ab1fd0e8700ea4e3f419d0d66e6211b814cf120440796cc0a
|
|
BLAKE2b-256 checksum How to use checksums |
2d2f5eb18ee927565eb420583a7b5c95779e196b3e14b958e98a1ee44bfe2b18
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/5.0.0 CPython/3.10.12
|
Release files / adamw_bf16-0.0.3-py3-none-any.whl
| Download URL | adamw_bf16-0.0.3-py3-none-any.whl |
|---|---|
| Size | 16.8 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
6795187b69b92cc962d0b02471173b997e6fb15b9ffc36825e9de3a207bebeec
|
|
BLAKE2b-256 checksum How to use checksums |
7c4c195a2835e02f862b6fe4d7067142ce66cde47a45274840c4e1b4b81fcc18
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/5.0.0 CPython/3.10.12
|
Security for adamw-bf16 0.0.3
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.
