Metadata-Version: 2.4
Name: jmp
Version: 0.0.6
Summary: JMP is a Mixed Precision library for JAX. Forked from DeepMind's JMP.
Author-email: DeepMind <jmp-dev-os@google.com>, Clement POIRET <poiret.clement@outlook.fr>
License: 
                                         Apache License
                                   Version 2.0, January 2004
                                http://www.apache.org/licenses/
        
           TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
        
           1. Definitions.
        
              "License" shall mean the terms and conditions for use, reproduction,
              and distribution as defined by Sections 1 through 9 of this document.
        
              "Licensor" shall mean the copyright owner or entity authorized by
              the copyright owner that is granting the License.
        
              "Legal Entity" shall mean the union of the acting entity and all
              other entities that control, are controlled by, or are under common
              control with that entity. For the purposes of this definition,
              "control" means (i) the power, direct or indirect, to cause the
              direction or management of such entity, whether by contract or
              otherwise, or (ii) ownership of fifty percent (50%) or more of the
              outstanding shares, or (iii) beneficial ownership of such entity.
        
              "You" (or "Your") shall mean an individual or Legal Entity
              exercising permissions granted by this License.
        
              "Source" form shall mean the preferred form for making modifications,
              including but not limited to software source code, documentation
              source, and configuration files.
        
              "Object" form shall mean any form resulting from mechanical
              transformation or translation of a Source form, including but
              not limited to compiled object code, generated documentation,
              and conversions to other media types.
        
              "Work" shall mean the work of authorship, whether in Source or
              Object form, made available under the License, as indicated by a
              copyright notice that is included in or attached to the work
              (an example is provided in the Appendix below).
        
              "Derivative Works" shall mean any work, whether in Source or Object
              form, that is based on (or derived from) the Work and for which the
              editorial revisions, annotations, elaborations, or other modifications
              represent, as a whole, an original work of authorship. For the purposes
              of this License, Derivative Works shall not include works that remain
              separable from, or merely link (or bind by name) to the interfaces of,
              the Work and Derivative Works thereof.
        
              "Contribution" shall mean any work of authorship, including
              the original version of the Work and any modifications or additions
              to that Work or Derivative Works thereof, that is intentionally
              submitted to Licensor for inclusion in the Work by the copyright owner
              or by an individual or Legal Entity authorized to submit on behalf of
              the copyright owner. For the purposes of this definition, "submitted"
              means any form of electronic, verbal, or written communication sent
              to the Licensor or its representatives, including but not limited to
              communication on electronic mailing lists, source code control systems,
              and issue tracking systems that are managed by, or on behalf of, the
              Licensor for the purpose of discussing and improving the Work, but
              excluding communication that is conspicuously marked or otherwise
              designated in writing by the copyright owner as "Not a Contribution."
        
              "Contributor" shall mean Licensor and any individual or Legal Entity
              on behalf of whom a Contribution has been received by Licensor and
              subsequently incorporated within the Work.
        
           2. Grant of Copyright License. Subject to the terms and conditions of
              this License, each Contributor hereby grants to You a perpetual,
              worldwide, non-exclusive, no-charge, royalty-free, irrevocable
              copyright license to reproduce, prepare Derivative Works of,
              publicly display, publicly perform, sublicense, and distribute the
              Work and such Derivative Works in Source or Object form.
        
           3. Grant of Patent License. Subject to the terms and conditions of
              this License, each Contributor hereby grants to You a perpetual,
              worldwide, non-exclusive, no-charge, royalty-free, irrevocable
              (except as stated in this section) patent license to make, have made,
              use, offer to sell, sell, import, and otherwise transfer the Work,
              where such license applies only to those patent claims licensable
              by such Contributor that are necessarily infringed by their
              Contribution(s) alone or by combination of their Contribution(s)
              with the Work to which such Contribution(s) was submitted. If You
              institute patent litigation against any entity (including a
              cross-claim or counterclaim in a lawsuit) alleging that the Work
              or a Contribution incorporated within the Work constitutes direct
              or contributory patent infringement, then any patent licenses
              granted to You under this License for that Work shall terminate
              as of the date such litigation is filed.
        
           4. Redistribution. You may reproduce and distribute copies of the
              Work or Derivative Works thereof in any medium, with or without
              modifications, and in Source or Object form, provided that You
              meet the following conditions:
        
              (a) You must give any other recipients of the Work or
                  Derivative Works a copy of this License; and
        
              (b) You must cause any modified files to carry prominent notices
                  stating that You changed the files; and
        
              (c) You must retain, in the Source form of any Derivative Works
                  that You distribute, all copyright, patent, trademark, and
                  attribution notices from the Source form of the Work,
                  excluding those notices that do not pertain to any part of
                  the Derivative Works; and
        
              (d) If the Work includes a "NOTICE" text file as part of its
                  distribution, then any Derivative Works that You distribute must
                  include a readable copy of the attribution notices contained
                  within such NOTICE file, excluding those notices that do not
                  pertain to any part of the Derivative Works, in at least one
                  of the following places: within a NOTICE text file distributed
                  as part of the Derivative Works; within the Source form or
                  documentation, if provided along with the Derivative Works; or,
                  within a display generated by the Derivative Works, if and
                  wherever such third-party notices normally appear. The contents
                  of the NOTICE file are for informational purposes only and
                  do not modify the License. You may add Your own attribution
                  notices within Derivative Works that You distribute, alongside
                  or as an addendum to the NOTICE text from the Work, provided
                  that such additional attribution notices cannot be construed
                  as modifying the License.
        
              You may add Your own copyright statement to Your modifications and
              may provide additional or different license terms and conditions
              for use, reproduction, or distribution of Your modifications, or
              for any such Derivative Works as a whole, provided Your use,
              reproduction, and distribution of the Work otherwise complies with
              the conditions stated in this License.
        
           5. Submission of Contributions. Unless You explicitly state otherwise,
              any Contribution intentionally submitted for inclusion in the Work
              by You to the Licensor shall be under the terms and conditions of
              this License, without any additional terms or conditions.
              Notwithstanding the above, nothing herein shall supersede or modify
              the terms of any separate license agreement you may have executed
              with Licensor regarding such Contributions.
        
           6. Trademarks. This License does not grant permission to use the trade
              names, trademarks, service marks, or product names of the Licensor,
              except as required for reasonable and customary use in describing the
              origin of the Work and reproducing the content of the NOTICE file.
        
           7. Disclaimer of Warranty. Unless required by applicable law or
              agreed to in writing, Licensor provides the Work (and each
              Contributor provides its Contributions) on an "AS IS" BASIS,
              WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
              implied, including, without limitation, any warranties or conditions
              of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
              PARTICULAR PURPOSE. You are solely responsible for determining the
              appropriateness of using or redistributing the Work and assume any
              risks associated with Your exercise of permissions under this License.
        
           8. Limitation of Liability. In no event and under no legal theory,
              whether in tort (including negligence), contract, or otherwise,
              unless required by applicable law (such as deliberate and grossly
              negligent acts) or agreed to in writing, shall any Contributor be
              liable to You for damages, including any direct, indirect, special,
              incidental, or consequential damages of any character arising as a
              result of this License or out of the use or inability to use the
              Work (including but not limited to damages for loss of goodwill,
              work stoppage, computer failure or malfunction, or any and all
              other commercial damages or losses), even if such Contributor
              has been advised of the possibility of such damages.
        
           9. Accepting Warranty or Additional Liability. While redistributing
              the Work or Derivative Works thereof, You may choose to offer,
              and charge a fee for, acceptance of support, warranty, indemnity,
              or other liability obligations and/or rights consistent with this
              License. However, in accepting such obligations, You may act only
              on Your own behalf and on Your sole responsibility, not on behalf
              of any other Contributor, and only if You agree to indemnify,
              defend, and hold each Contributor harmless for any liability
              incurred by, or claims asserted against, such Contributor by reason
              of your accepting any such warranty or additional liability.
        
           END OF TERMS AND CONDITIONS
        
           APPENDIX: How to apply the Apache License to your work.
        
              To apply the Apache License to your work, attach the following
              boilerplate notice, with the fields enclosed by brackets "[]"
              replaced with your own identifying information. (Don't include
              the brackets!)  The text should be enclosed in the appropriate
              comment syntax for the file format. We also recommend that a
              file or class name and description of purpose be included on the
              same "printed page" as the copyright notice for easier
              identification within third-party archives.
        
           Copyright [yyyy] [name of copyright owner]
        
           Licensed under the Apache License, Version 2.0 (the "License");
           you may not use this file except in compliance with the License.
           You may obtain a copy of the License at
        
               http://www.apache.org/licenses/LICENSE-2.0
        
           Unless required by applicable law or agreed to in writing, software
           distributed under the License is distributed on an "AS IS" BASIS,
           WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
           See the License for the specific language governing permissions and
           limitations under the License.
License-File: LICENSE
Requires-Python: >=3.10
Requires-Dist: jax>=0.4.17
Requires-Dist: numpy>=1.26
Description-Content-Type: text/markdown

# Mixed precision training in [JAX]

![Test status](https://github.com/deepmind/jmp/workflows/pytest/badge.svg)
![PyPI version](https://img.shields.io/pypi/v/jmp)

[**Installation**](#installation)
| [**Examples**](#examples)
| [**Policies**](#policies)
| [**Loss scaling**](#loss-scaling)
| [**Citing JMP**](#citing-jmp)
| [**References**](#references)

Mixed precision training [[0]] is a technique that mixes the use of full and
half precision floating point numbers during training to reduce the memory
bandwidth requirements and improve the computational efficiency of a given
model.

This library implements support for mixed precision training in [JAX] by providing
two key abstractions (mixed precision "policies" and loss scaling). Neural
network libraries (such as [Haiku]) can integrate with `jmp` and provide
"Automatic Mixed Precision (AMP)" support (automating or simplifying applying
policies to modules).

All code examples below assume the following:

```python
import jax
import jax.numpy as jnp
import jmp

half = jnp.float16  # On TPU this should be jnp.bfloat16.
full = jnp.float32
```

## Installation

JMP is written in pure Python, but depends on C++ code via JAX and NumPy.

Because JAX installation is different depending on your CUDA version, JMP does
not list JAX as a dependency in `requirements.txt`.

First, follow [these instructions](https://github.com/google/jax#installation)
to install JAX with the relevant accelerator support.

Then, install JMP using pip:

```bash
$ pip install git+https://github.com/clementpoiret/jmp
```

## Examples

You can find a
[fully worked JMP example in Haiku](https://github.com/deepmind/dm-haiku/tree/master/examples/imagenet)
which shows how to use mixed f32/f16 precision to halve training time on GPU and
mixed f32/bf16 to reduce training time on TPU by a third.

## Policies

A mixed precision policy encapsulates the configuration in a mixed precision
experiment.

```python
# Our policy specifies that we will store parameters in full precision but will
# compute and return output in half precision.
my_policy = jmp.Policy(compute_dtype=half,
                       param_dtype=full,
                       output_dtype=half)
```

The policy object can be used to cast pytrees:

```python
def layer(params, x):
  params, x = my_policy.cast_to_compute((params, x))
  w, b = params
  y = x @ w + b
  return my_policy.cast_to_output(y)

params = {"w": jnp.ones([], dtype=my_policy.param_dtype)}
y = layer(params, x)
assert y.dtype == half
```

You can replace the output type of a given policy:

```python
my_policy = my_policy.with_output_dtype(full)
```

You can also define a policy via a string, which may be useful for specifying a
policy as a command-line argument or as a hyperparameter to your experiment:

```python
my_policy = jmp.get_policy("params=float32,compute=float16,output=float32")
float16 = jmp.get_policy("float16")  # Everything in f16.
half = jmp.get_policy("half")        # Everything in half (f16 or bf16).
```

## Loss scaling

When training with reduced precision, consider whether gradients will need to be
shifted into the representable range of the format that you are using. This is
particularly important when training with `float16` and less important for
`bfloat16`. See the NVIDIA mixed precision user guide [[1]] for more details.

The easiest way to shift gradients is with loss scaling, which scales your loss
and gradients by `S` and `1/S` respectively.

```python
def my_loss_fn(params, loss_scale: jmp.LossScale, ...):
  loss = ...
  # You should apply regularization etc before scaling.
  loss = loss_scale.scale(loss)
  return loss

def train_step(params, loss_scale: jmp.LossScale, ...):
  grads = jax.grad(my_loss_fn)(...)
  grads = loss_scale.unscale(grads)
  # You should put gradient clipping etc after unscaling.
  params = apply_optimizer(params, grads)
  return params

loss_scale = jmp.StaticLossScale(2 ** 15)
for _ in range(num_steps):
  params = train_step(params, loss_scale, ...)
```

The appropriate value for `S` depends on your model, loss, batch size and
potentially other factors. You can determine this with trial and error. As a
rule of thumb you want the largest value of `S` that does not introduce overflow
during backprop. NVIDIA [[1]] recommend computing statistics about the gradients
of your model (in full precision) and picking `S` such that its product with the
maximum norm of your gradients is below `65,504`.

We provide a dynamic loss scale, which adjusts the loss scale periodically
during training to find the largest value for `S` that produces finite
gradients. This is more convenient and robust compared with picking a static
loss scale, but has a small performance impact (between 1 and 5%).

```python
def my_loss_fn(params, loss_scale: jmp.LossScale, ...):
  loss = ...
  # You should apply regularization etc before scaling.
  loss = loss_scale.scale(loss)
  return loss

def train_step(params, loss_scale: jmp.LossScale, ...):
  grads = jax.grad(my_loss_fn)(...)
  grads = loss_scale.unscale(grads)
  # You should put gradient clipping etc after unscaling.

  # You definitely want to skip non-finite updates with the dynamic loss scale,
  # but you might also want to consider skipping them when using a static loss
  # scale if you experience NaN's when training.
  skip_nonfinite_updates = isinstance(loss_scale, jmp.DynamicLossScale)

  if skip_nonfinite_updates:
    grads_finite = jmp.all_finite(grads)
    # Adjust our loss scale depending on whether gradients were finite. The
    # loss scale will be periodically increased if gradients remain finite and
    # will be decreased if not.
    loss_scale = loss_scale.adjust(grads_finite)
    # Only apply our optimizer if grads are finite, if any element of any
    # gradient is non-finite the whole update is discarded.
    params = jmp.select_tree(grads_finite, apply_optimizer(params, grads), params)
  else:
    # With static or no loss scaling just apply our optimizer.
    params = apply_optimizer(params, grads)

  # Since our loss scale is dynamic we need to return the new value from
  # each step. All loss scales are `PyTree`s.
  return params, loss_scale

loss_scale = jmp.DynamicLossScale(jnp.float32(2**15))
for _ in range(num_steps):
  params, loss_scale = train_step(params, loss_scale, ...)
```

In general using a static loss scale should offer the best speed, but we have
optimized dynamic loss scaling to make it competitive. We recommend you start
with dynamic loss scaling and move to static loss scaling if performance is an
issue.

We finally offer a no-op loss scale which you can use as a drop in replacement.
It does nothing (apart from implement the `jmp.LossScale` API):

```python
loss_scale = jmp.NoOpLossScale()
assert loss is loss_scale.scale(loss)
assert grads is loss_scale.unscale(grads)
assert loss_scale is loss_scale.adjust(grads_finite)
assert loss_scale.loss_scale == 1
```

## Citing JMP

This repository is part of the [DeepMind JAX Ecosystem](https://deepmind.com/blog/article/using-jax-to-accelerate-our-research),
to cite JMP please use the [DeepMind JAX Ecosystem citation](https://github.com/deepmind/jax/blob/main/deepmind2020jax.txt).

## References

[[0]] Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich
Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh
Venkatesh, Hao Wu: "Mixed Precision Training", 2017; arXiv:1710.03740
https://arxiv.org/abs/1710.03740.

[[1]] "Training With Mixed Precision :: NVIDIA Deep Learning Performance
Documentation". Docs.Nvidia.Com, 2020,
https://docs.nvidia.com/deeplearning/performance/mixed-precision-training/.

[0]: https://arxiv.org/abs/1710.03740
[1]: https://docs.nvidia.com/deeplearning/performance/mixed-precision-training/index.html
[Haiku]: https://github.com/deepmind/dm-haiku
[JAX]: https://github.com/google/jax