tqdx 0.1.4

JAX compatible progress bars for scan and fori_loop.

tqdx

Adds tqdm progress bars to jax.lax.scan and jax.lax.fori_loop. Progress bars commonly used in Python, such as tqdm, are not compatible with JAX's jit-compiled functions due to restrictions on side effects like printing. tqdx addresses this limitation by using callbacks to update progress bars created on the host.

import tqdx

...
carry, ys = tqdx.scan(f, init, xs)

Processing: 100%|███████████████████████████████████████████| 50/50 [02:38<00:00,  3.20s/it]

Features

• Progress bars for JAX: See the progress of your computations when using jax.lax.scan and jax.lax.fori_loop.
• Works with jax.jit: Progress bars show up even inside jit-compiled code.
• Minimal syntax change: Just replace your calls to jax.lax.scan and jax.lax.fori_loop with tqdx.scan and tqdx.fori_loop.
• No extra dependencies: Only requires JAX and tqdm.

Usage

The following example demonstrates how to use tqdx with jax.lax.scan and jax.lax.fori_loop. You can arbitrarily nest these functions, and the progress bars will still work correctly.

import jax
import tqdx
from time import sleep

def step(carry, x):
    def body_fun(i, val):
        jax.debug.callback(lambda: sleep(0.5))
        return val + i
    jax.debug.callback(lambda: sleep(0.5))
    carry = tqdx.fori_loop(0, 10, body_fun, carry)
    return carry, x + 1

def f(xs):
    return tqdx.scan(step, 0, xs)


xs = jax.numpy.arange(10)
result, _ = jax.jit(f)(xs)

Installation

pip install tqdx