This release is a pre-release and may not be stable for production use.
Haliax
Though you don’t seem to be much for listening, it’s best to be careful. If you managed to catch hold of even just a piece of my name, you’d have all manner of power over me.
— Patrick Rothfuss, The Name of the Wind
Haliax is a JAX library for building neural networks with named tensors, in the tradition of Alexander Rush's Tensor Considered Harmful. Named tensors improve the legibility and compositionality of tensor programs by using named axes instead of positional indices as typically used in NumPy, PyTorch, etc.
Despite the focus on legibility, Haliax is also fast, typically about as fast as "pure" JAX code. Haliax is also built to be scalable: it can support Fully-Sharded Data Parallelism (FSDP) and Tensor Parallelism with just a few lines of code. Haliax powers Levanter, our companion library for training large language models and other foundation models, with scale proven up to 70B parameters and up to TPU v4-2048.
Example: Attention
Here's a minimal attention module implementation in Haliax. For a more detailed introduction, please see the Haliax tutorial. (We use the excellent Equinox library for its module system and tree transformations.)
import equinox as eqx
import jax
import jax.numpy as jnp
import haliax as hax
import haliax.nn as hnn
Pos = hax.Axis("position", 1024) # sequence length
KPos = Pos.alias("key_position")
Head = hax.Axis("head", 8) # number of attention heads
Key = hax.Axis("key", 64) # key size
Embed = hax.Axis("embed", 512) # embedding size
# alternatively:
#Pos, KPos, Head, Key, Embed = hax.make_axes(pos=1024, key_pos=1024, head=8, key=64, embed=512)
def attention_scores(Key, KPos, query, key, mask):
# how similar is each query to each key
scores = hax.dot(query, key, axis=Key) / jnp.sqrt(Key.size)
if mask is not None:
scores -= 1E9 * (1.0 - mask)
# convert to probabilities
scores = haliax.nn.softmax(scores, KPos)
return scores
def attention(Key, KPos, query, key, value, mask):
scores = attention_scores(Key, KPos, query, key, mask)
answers = hax.dot(scores, value, axis=KPos)
return answers
# Causal Mask means that if pos >= key_pos, then pos can attend to key_pos
causal_mask = hax.arange(Pos).broadcast_axis(KPos) >= hax.arange(KPos)
class Attention(eqx.Module):
proj_q: hnn.Linear # [Embed] -> [Head, Key]
proj_k: hnn.Linear # [Embed] -> [Head, Key]
proj_v: hnn.Linear # [Embed] -> [Head, Key]
proj_answer: hnn.Linear # output projection from [Head, Key] -> [Embed]
@staticmethod
def init(Embed, Head, Key, *, key):
k_q, k_k, k_v, k_ans = jax.random.split(key, 4)
proj_q = hnn.Linear.init(In=Embed, Out=(Head, Key), key=k_q)
proj_k = hnn.Linear.init(In=Embed, Out=(Head, Key), key=k_k)
proj_v = hnn.Linear.init(In=Embed, Out=(Head, Key), key=k_v)
proj_answer = hnn.Linear.init(In=(Head, Key), Out=Embed, key=k_ans)
return Attention(proj_q, proj_k, proj_v, proj_answer)
def __call__(self, x, mask=None):
q = self.proj_q(x)
# Rename "position" to "key_position" for self attention
k = self.proj_k(x).rename({"position": "key_position"})
v = self.proj_v(x).rename({"position": "key_position"})
answers = attention(Key, KPos, q, k, v, causal_mask)
x = self.proj_answer(answers)
return x
Haliax was created by Stanford's Center for Research on Foundation Models (CRFM)'s research engineering team. You can find us in the #levanter channel on the unofficial Jax LLM Discord.
Documentation
Tutorials
These are some tutorials to get you started with Haliax. They are available as Colab notebooks:
- Introduction to Haliax with Transformers
- Distributed Training in Haliax (including FSDP)
- Tensor Parallelism in Haliax
- Mixed Precision with
jmp(This one is really a tutorial for jmp but it's how to use it with Haliax...)
API Reference
Haliax's API documentation is available at haliax.readthedocs.io.
Contributing
We welcome contributions! Please see CONTRIBUTING.md for more information. We also have a list of good first issues to help you get started. (If those don't appeal, don't hesitate to reach out to us on Discord!)
License
Haliax is licensed under the Apache License, Version 2.0. See LICENSE for the full license text.
Release files for marin-haliax 0.2.155.dev37928019690
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | |
|---|---|---|---|
| marin_haliax-0.2.155.dev37928019690.tar.gz | 781.0 kB | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| marin_haliax-0.2.155.dev37928019690-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 940.2 kB
Release files / marin_haliax-0.2.155.dev37928019690.tar.gz
| Download URL | marin_haliax-0.2.155.dev37928019690.tar.gz |
|---|---|
| Size | 781.0 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
3b5747e57da11bf396dd939743ed0625524707a13829cc656ae41484c14edb2b
|
|
BLAKE2b-256 checksum How to use checksums |
31c0582e20772144debbe245aaf3554c1419be593110bfad07a2660987e05029
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
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 Oct 9, 2026.
Transparency logRelease files / marin_haliax-0.2.155.dev37928019690-py3-none-any.whl
| Download URL | marin_haliax-0.2.155.dev37928019690-py3-none-any.whl |
|---|---|
| Size | 159.2 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
b786f95be33fa421587f742a74199866e1824e2ca851fb7061d040a96f8af084
|
|
BLAKE2b-256 checksum How to use checksums |
d520d057aab38ee890884913c0a809935ef5b29a6a75f49c6336734038100f52
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
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 Oct 9, 2026.
Transparency log
