smartclip
Adaptive gradient clipping for PyTorch, TensorFlow, and JAX.
SmartClip keeps training stable with adaptive, per-step clipping you can enable in one line of code.
See the full documentation for details of the algorithms, framework usage examples, and logging metrics.
Supported Algorithms
-
AutoClip — Seetharaman et al., 2020 (MLSP). Adaptive percentile-based clipping of gradient norms.
-
Adaptive Gradient Clipping (AGC, NFNets-style) — Brock, De, Smith, 2021. Threshold scales with parameter norm.
-
Z-Score clipping (EMA mean/std) — standard z-score thresholding using streaming mean/variance
zmaxcontrols how aggressive clipping is: threshold ismean + zmax * stdover recent norms. Higherzmaxclips less (more tolerant), lower clips more (more aggressive). Start atzmax=3.0; try2.0–2.5if you see instability from spikes, or3.5–4.0if training seems over‑clipped.
Install
pip install smartclip
Optional extras provide helpers for specific frameworks (install framework wheels first per vendor docs):
pip install "smartclip[torch]" # PyTorch + Lightning/Transformers helpers
pip install "smartclip[tf]" # TensorFlow/Keras helpers
pip install "smartclip[jax]" # JAX/Flax/Optax helpers
Quickstart
PyTorch
import torch
import smartclip as sc
model = MyModel().to("cpu")
opt = torch.optim.AdamW(model.parameters(), lr=3e-4)
with sc.clip_context(model, opt): # Defaults to AutoClip
for x, y in loader:
opt.zero_grad(set_to_none=True)
loss = model(x).loss_fn(y)
loss.backward()
opt.step() # clipped automatically
TensorFlow/Keras
import tensorflow as tf
import smartclip as sc
model = MyModel()
opt = tf.keras.optimizers.Adam(3e-4)
with sc.clip_context(model, opt, clipper=sc.ZScoreClip(zmax=3.0)): # Use the zscore algorithm
model.fit(ds, epochs=5)
JAX/Optax
import jax
import optax
from flax import linen as nn
import smartclip as sc
model = MyModel() # Flax Module
tx = optax.adam(3e-4)
with sc.clip_context(model, tx): # wraps tx.update
grads = jax.grad(loss_fn)(params, batch)
updates, opt_state = tx.update(grads, opt_state, params) # clipped automatically
params = optax.apply_updates(params, updates)
See documentation for full guides for TensorFlow, JAX, Lightning, Keras, and HF Trainer.
Contributing
We welcome issues and pull requests. See contribute.md for developer setup, testing, docs, and release workflows.
License
MIT
Metadata
Release files for smartclip 0.1.1
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 | Actions |
|---|---|---|---|---|
| smartclip-0.1.1.tar.gz | 32.3 kB | With provenance | Details for smartclip-0.1.1.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| smartclip-0.1.1-py3-none-any.whl | Python 3 | none | any | Details for smartclip-0.1.1-py3-none-any.whl |
Total release size: 65.6 kB
Release files / smartclip-0.1.1.tar.gz
| Download URL | smartclip-0.1.1.tar.gz |
|---|---|
| Size | 32.3 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
c0dff18670d92a33b16b95b857a911f6addfcf0c24895279c537c5be969d0dfc
|
|
BLAKE2b-256 checksum How to use checksums |
43c637a8bc1c786fb3d16d9af1708e5ca4ac9fd8113c95c9a7f3ddbb97f636b0
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/6.1.0 CPython/3.13.7
|
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.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Sep 13, 2025.
Transparency logRelease files / smartclip-0.1.1-py3-none-any.whl
| Download URL | smartclip-0.1.1-py3-none-any.whl |
|---|---|
| Size | 33.3 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
3002e3b47c946cf6686a80096b455176fc30572fc48f595c05d3df5ac778d580
|
|
BLAKE2b-256 checksum How to use checksums |
b1e722adae999f4d443ab78f52185737251ad38a45d5c4bcb09818419ced31e9
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/6.1.0 CPython/3.13.7
|
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.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Sep 13, 2025.
Transparency logSecurity for smartclip 0.1.1
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.
Full provenance
Both files in this release have been uploaded with provenance metadata.
Attestation for smartclip-0.1.1.tar.gz
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Sep 13, 2025.
Transparency logA 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.