* . *
  • About
  • Advertise
  • Privacy & Policy
  • Contact
Friday, October 24, 2025
Earth-News
  • Home
  • Business
  • Entertainment
    ‘Chainsaw Man — The Movie: Reze Arc’ Review: Hit Manga Gets an Ultra-Violent, Surprisingly Emotional Big-Screen Adaptation – Yahoo

    Chainsaw Man – The Movie: Reze Arc Review: A Brutal and Unexpectedly Emotional Big-Screen Adaptation

    Reba McEntire Details Personal Relationship With Late Stepson Brandon Blackstock – KNDU

    Reba McEntire Shares Emotional Tribute to Her Late Stepson Brandon Blackstock

    Sacramento city leaders approve adding 2 entertainment zones in midtown – CBS News

    Sacramento City Leaders Approve Two Thrilling New Entertainment Zones in Midtown

    AMC brings first new Dolby Experience to Gwinnett since 2017 – Wyoming News Now

    AMC Launches First New Dolby Experience in Gwinnett Since 2017

    Hetzel Design: blending architecture and entertainment – Blooloop

    Hetzel Design: Where Architecture and Entertainment Unite in Perfect Harmony

    Country music legend rushed to hospital year after heart surgery. Here’s what we know – PennLive.com

    Country Music Legend Rushed to Hospital One Year After Heart Surgery – What’s Happening Now?

  • General
  • Health
  • News

    Cracking the Code: Why China’s Economic Challenges Aren’t Shaking Markets, Unlike America’s” – Bloomberg

    Trump’s Narrow Window to Spread the Truth About Harris

    Trump’s Narrow Window to Spread the Truth About Harris

    Israel-Gaza war live updates: Hamas leader Ismail Haniyeh assassinated in Iran, group says

    Israel-Gaza war live updates: Hamas leader Ismail Haniyeh assassinated in Iran, group says

    PAP Boss to Niger Delta Youths, Stay Away from the Protest

    PAP Boss to Niger Delta Youths, Stay Away from the Protest

    Court Restricts Protests In Lagos To Freedom, Peace Park

    Court Restricts Protests In Lagos To Freedom, Peace Park

    Fans React to Jazz Jennings’ Inspiring Weight Loss Journey

    Fans React to Jazz Jennings’ Inspiring Weight Loss Journey

    Trending Tags

    • Trump Inauguration
    • United Stated
    • White House
    • Market Stories
    • Election Results
  • Science
  • Sports
  • Technology
    Project underway to upgrade technology on 911 towers in Kanawha County – WCHS

    Kanawha County Launches Major Upgrade to 911 Tower Technology

    Next steps: Technology opens new options for greater mobility – Missouri Independent

    Next Steps: How Technology is Opening Exciting New Doors to Greater Mobility

    Rydberg Technologies Inc. Announces Launch of Rydberg Photonics in Berlin – The Quantum Insider

    Rydberg Technologies Launches Exciting New Photonics Division in Berlin

    A look into new technology at Columbia University that could help prevent a dangerous pregnancy complication – ABC7 New York

    A look into new technology at Columbia University that could help prevent a dangerous pregnancy complication – ABC7 New York

    Office Technology: Dealers’ Managed IT Revenue up Nearly 30% – The Cannata Report –

    Office Technology: Dealers’ Managed IT Revenue up Nearly 30% – The Cannata Report –

    3 E Network Technology Group Limited Closes $1.5 Million Convertible Promissory Note Offering – Quiver Quantitative

    3 E Network Technology Group Limited Closes $1.5 Million Convertible Promissory Note Offering – Quiver Quantitative

    Trending Tags

    • Nintendo Switch
    • CES 2017
    • Playstation 4 Pro
    • Mark Zuckerberg
No Result
View All Result
  • Home
  • Business
  • Entertainment
    ‘Chainsaw Man — The Movie: Reze Arc’ Review: Hit Manga Gets an Ultra-Violent, Surprisingly Emotional Big-Screen Adaptation – Yahoo

    Chainsaw Man – The Movie: Reze Arc Review: A Brutal and Unexpectedly Emotional Big-Screen Adaptation

    Reba McEntire Details Personal Relationship With Late Stepson Brandon Blackstock – KNDU

    Reba McEntire Shares Emotional Tribute to Her Late Stepson Brandon Blackstock

    Sacramento city leaders approve adding 2 entertainment zones in midtown – CBS News

    Sacramento City Leaders Approve Two Thrilling New Entertainment Zones in Midtown

    AMC brings first new Dolby Experience to Gwinnett since 2017 – Wyoming News Now

    AMC Launches First New Dolby Experience in Gwinnett Since 2017

    Hetzel Design: blending architecture and entertainment – Blooloop

    Hetzel Design: Where Architecture and Entertainment Unite in Perfect Harmony

    Country music legend rushed to hospital year after heart surgery. Here’s what we know – PennLive.com

    Country Music Legend Rushed to Hospital One Year After Heart Surgery – What’s Happening Now?

  • General
  • Health
  • News

    Cracking the Code: Why China’s Economic Challenges Aren’t Shaking Markets, Unlike America’s” – Bloomberg

    Trump’s Narrow Window to Spread the Truth About Harris

    Trump’s Narrow Window to Spread the Truth About Harris

    Israel-Gaza war live updates: Hamas leader Ismail Haniyeh assassinated in Iran, group says

    Israel-Gaza war live updates: Hamas leader Ismail Haniyeh assassinated in Iran, group says

    PAP Boss to Niger Delta Youths, Stay Away from the Protest

    PAP Boss to Niger Delta Youths, Stay Away from the Protest

    Court Restricts Protests In Lagos To Freedom, Peace Park

    Court Restricts Protests In Lagos To Freedom, Peace Park

    Fans React to Jazz Jennings’ Inspiring Weight Loss Journey

    Fans React to Jazz Jennings’ Inspiring Weight Loss Journey

    Trending Tags

    • Trump Inauguration
    • United Stated
    • White House
    • Market Stories
    • Election Results
  • Science
  • Sports
  • Technology
    Project underway to upgrade technology on 911 towers in Kanawha County – WCHS

    Kanawha County Launches Major Upgrade to 911 Tower Technology

    Next steps: Technology opens new options for greater mobility – Missouri Independent

    Next Steps: How Technology is Opening Exciting New Doors to Greater Mobility

    Rydberg Technologies Inc. Announces Launch of Rydberg Photonics in Berlin – The Quantum Insider

    Rydberg Technologies Launches Exciting New Photonics Division in Berlin

    A look into new technology at Columbia University that could help prevent a dangerous pregnancy complication – ABC7 New York

    A look into new technology at Columbia University that could help prevent a dangerous pregnancy complication – ABC7 New York

    Office Technology: Dealers’ Managed IT Revenue up Nearly 30% – The Cannata Report –

    Office Technology: Dealers’ Managed IT Revenue up Nearly 30% – The Cannata Report –

    3 E Network Technology Group Limited Closes $1.5 Million Convertible Promissory Note Offering – Quiver Quantitative

    3 E Network Technology Group Limited Closes $1.5 Million Convertible Promissory Note Offering – Quiver Quantitative

    Trending Tags

    • Nintendo Switch
    • CES 2017
    • Playstation 4 Pro
    • Mark Zuckerberg
No Result
View All Result
Earth-News
No Result
View All Result
Home Technology

Maxtext: A simple, performant and scalable Jax LLM

April 24, 2024
in Technology
Maxtext: A simple, performant and scalable Jax LLM
Share on FacebookShare on Twitter

Unit Tests

MaxText is a high performance, highly scalable, open-source LLM written in pure Python/Jax and targeting Google Cloud TPUs and GPUs for training and inference. MaxText achieves high MFUs and scales from single host to very large clusters while staying simple and “optimization-free” thanks to the power of Jax and the XLA compiler.

MaxText aims to be a launching off point for ambitious LLM projects both in research and production. We encourage users to start by experimenting with MaxText out of the box and then fork and modify MaxText to meet their needs.

We have used MaxText to demonstrate high-performance, well-converging training in int8 and scale training to ~51K chips.

Key supported features:

TPUs and GPUs (in preview)
Training and Inference (in preview)
Models: Llama2, Mistral and Gemma

Getting Started
Runtime Performance Results
Comparison To Alternatives
Development
Features and Diagnostics

For your first time running MaxText, we provide specific instructions.

MaxText supports training and inference of various open models. Follow user guides in the getting started folder to know more.

Some extra helpful guides:

Gemma: a family of open-weights Large Language Model (LLM) by Google DeepMind, based on Gemini research and technology. You can run decode and finetuning using these instructions.
Llama2: a family of open-weights Large Language Model (LLM) by Meta. You can run decode and finetuning using these instructions.

In addition to the getting started guides, there are always other MaxText capabilities that are being constantly being added! The full suite of end-to-end tests is in end_to_end. We run them with a nightly cadence. They can be a good source for understanding MaxText Alternatively you can see the continuous unit tests which are run almost continuously.

More details on reproducing these results can be found in MaxText/configs/README.md.

No. of params
Accelerator Type
TFLOP/chip/sec
Model flops utilization (MFU)

32B
v5p-128
3.28e+02
71.47%

64B
v5p-128
3.23e+02
70.31%

128B
v5p-256
3.15e+02
68.68%

128B
v5p-512
3.15e+02
68.53%

256B
v5p-1024
3.16e+02
68.82%

512B
v5p-1024
2.94e+02
63.99%

1024B
v5p-2048
2.49e+02
64.05%

1024B
v5p-4096
2.97e+02
64.80%

1160B
v5p-7680
2.95e+02
64.27%

1160B
v5p-12288
3.04e+02
66.23%

For 16B, 32B, 64B, and 128B models. See full run configs in MaxText/configs/v5e/ as 16b.sh, 32b.sh, 64b.sh, 128b.sh.

Hardware
16B TFLOP/sec/chip
16B MFU
32B TFLOP/sec/chip
32B MFU
64B TFLOP/sec/chip
64B MFU
128B TFLOP/sec/chip
128B MFU

1x v5e-256
120
61.10%
132
66.86%
118
59.90%
110
56.06%

2x v5e-256
117
59.37%
128
64.81%
112
56.66%
110
55.82%

4x v5e-256
117
59.14%
126
64.10%
110
55.85%
108
54.93%

8x v5e-256
115
58.27%
125
63.67%
108
54.96%
104
52.93%

16x v5e-256
111
56.56%
123
62.26%
105
53.29%
100
50.86%

32x v5e-256
108
54.65%
119
60.40%
99
50.18%
91
46.25%

MaxText is heavily inspired by MinGPT/NanoGPT, elegant standalone GPT implementations written in PyTorch and targeting Nvidia GPUs. MaxText is more complex, supporting more industry standard models and scaling to tens of thousands of chips. Ultimately MaxText has an MFU more than three times the 17% reported most recently with that codebase, is massively scalable and implements a key-value cache for efficient auto-regressive decoding.

MaxText is more similar to Nvidia/Megatron-LM, a very well tuned LLM implementation targeting Nvidia GPUs. The two implementations achieve comparable MFUs. The difference in the codebases highlights the different programming strategies. MaxText is pure Python, relying heavily on the XLA compiler to achieve high performance. By contrast, Megatron-LM is a mix of Python and CUDA, relying on well-optimized CUDA kernels to achieve high performance.

MaxText is also comparable to Pax. Like Pax, MaxText provides high-performance and scalable implementations of LLMs in Jax. Pax focuses on enabling powerful configuration parameters, enabling developers to change the model by editing config parameters. By contrast, MaxText is a simple, concrete implementation of various LLMs that encourages users to extend by forking and directly editing the source code.

When running a Single Program, Multiple Data (SPMD) job on accelerators, the overall process can hang if there is any error or any VM hangs/crashes for some reason. In this scenario, capturing stack traces will help to identify and troubleshoot the issues for the jobs running on TPU VMs.

The following configurations will help to debug a fault or when a program is stuck or hung somewhere by collecting stack traces. Change the parameter values accordingly in MaxText/configs/base.yml:

Set collect_stack_trace: True to enable collection of stack traces on faults or when the program is hung. This setting will periodically dump the traces for the program to help in debugging. To disable this, set collect_stack_trace: False.
Set stack_trace_to_cloud: False to display stack traces on console. stack_trace_to_cloud: True will create a temporary file in /tmp/debugging in the TPUs to store the stack traces. There is an agent running on TPU VMs that will periodically upload the traces from the temporary directory to cloud logging in the gcp project. You can view the traces in Logs Explorer on Cloud Logging using the following query:

logName=”projects//logs/tpu.googleapis.com%2Fruntime_monitor”
jsonPayload.verb=”stacktraceanalyzer”

stack_trace_interval_seconds signifies the duration in seconds between each stack trace collection event. Setting stack_trace_interval_seconds: 600 will collect the stack traces every 600 seconds (10 minutes).

Here is the related PyPI package: https://pypi.org/project/cloud-tpu-diagnostics.

Ahead of Time Compilation (AOT, tpu-only)

To compile your training run ahead of time, we provide a tool train_compile.py. This tool allows you to compile the main train_step in train.py for target hardware (e.g. a large number of v5e devices) without using the target hardware, and instead you may use only a CPU or a single VM from a different family. This compilation helps with two main goals:

It will flag any out of memory (OOM) information, such as when the per_device_batch_size is set too high, with an identical OOM stack trace as if it was compiled on the target hardware.

The ahead of time compilation can be saved and then loaded for fast startup and restart times on the target hardware.

The tool train_compile.py is tightly linked to train.py and uses the same configuration file configs/base.yml. Although you don’t need to run on a TPU, you do need to install jax[tpu] in addition to other dependencies, so we recommend running setup.sh to install these if you have not already done so.

Example AOT 1: Compile ahead of time basics

After installing the dependencies listed above, you are ready to compile ahead of time:

# Run the below on a single machine, e.g. a CPU
python3 MaxText/train_compile.py MaxText/configs/base.yml compile_topology=v5e-256 compile_topology_num_slices=2
global_parameter_scale=16 per_device_batch_size=4

This will compile a 16B parameter MaxText model on 2 v5e pods.

Example AOT 2: Save compiled function, then load and run it

Here is an example that saves then loads the compiled train_step, starting with the save:

Step 1: Run AOT and save compiled function

# Run the below on a single machine, e.g. a CPU
export LIBTPU_INIT_ARGS=”–xla_enable_async_all_gather=true”
python3 MaxText/train_compile.py MaxText/configs/base.yml compile_topology=v5e-256
compile_topology_num_slices=2
compiled_trainstep_file=my_compiled_train.pickle global_parameter_scale=16
per_device_batch_size=4 steps=10000 learning_rate=1e-3

Step 2: Run train.py and load the compiled function

To load the compiled train_step, you just need to pass compiled_trainstep_file=my_compiled_train.pickle into train.py:

# Run the below on each host of the target hardware, e.g. each host on 2 slices of v5e-256
export LIBTPU_INIT_ARGS=”–xla_enable_async_all_gather=true”
python3 MaxText/train.py MaxText/configs/base.yml run_name=example_load_compile
compiled_trainstep_file=my_compiled_train.pickle
global_parameter_scale=16 per_device_batch_size=4 steps=10000 learning_rate=1e-3
base_output_directory=gs://my-output-bucket dataset_path=gs://my-dataset-bucket

In the save step of example 2 above we included exporting the compiler flag LIBTPU_INIT_ARGS and learning_rate because those affect the compiled object my_compiled_train.pickle. The sizes of the model (e.g. global_parameter_scale, max_sequence_length and per_device_batch) are fixed when you initially compile via compile_train.py, you will see a size error if you try to run the saved compiled object with different sizes than you compiled with. However a subtle note is that the learning rate schedule is also fixed when you run compile_train – which is determined by both steps and learning_rate. The optimizer parameters such as adam_b1 are passed only as shaped objects to the compiler – thus their real values are determined when you run train.py, not during the compilation. If you do pass in different shapes (e.g. per_device_batch), you will get a clear error message reporting that the compiled signature has different expected shapes than what was input. If you attempt to run on different hardware than the compilation targets requested via compile_topology, you will get an error saying there is a failure to map the devices from the compiled to your real devices. Using different XLA flags or a LIBTPU than what was compiled will probably run silently with the environment you compiled in without error. However there is no guaranteed behavior in this case; you should run in the same environment you compiled in.

Automatically Upload Logs to Vertex Tensorboard

MaxText supports automatic upload of logs collected in a directory to a Tensorboard instance in Vertex AI. Follow user guide to know more.

>>> Read full article>>>
Copyright for syndicated content belongs to the linked Source : Hacker News – https://github.com/google/maxtext

Tags: MaxtextSimpletechnology
Previous Post

BeeBase, a programmable relational database with graphical user interface

Next Post

Golang PGO builds using GitHub Actions

Ecology researchers work to update Columbia River fish advisories — but face new challenges in climate change – Longview Daily News

Ecology researchers work to update Columbia River fish advisories — but face new challenges in climate change – Longview Daily News

October 24, 2025

Unlock the Secrets to Growing a Giant Pumpkin Using Science

October 24, 2025
Rare dinosaur mummies help scientists recreate their prehistoric lives – ABC News – Breaking News, Latest News and Videos

Rare Dinosaur Mummies Unlock Secrets of Their Prehistoric Lives

October 24, 2025
Orange Cat Celebrates Halloween by Putting Candy Corn in Her Water Bowl – Yahoo

Orange Cat’s Hilarious Halloween Trick: Candy Corn in Her Water Bowl!

October 24, 2025
Project underway to upgrade technology on 911 towers in Kanawha County – WCHS

Kanawha County Launches Major Upgrade to 911 Tower Technology

October 24, 2025
Sports gambling indictment revives tough questions about integrity – Axios

Sports Gambling Indictment Sparks Renewed Debate Over Integrity

October 23, 2025
Former world chess champion Vladimir Kramnik investigated for bullying following death of grandmaster Daniel Naroditsky – CNN

Former World Chess Champion Vladimir Kramnik Under Investigation Amid Tragic Loss of Grandmaster Daniel Naroditsky

October 23, 2025
Trump approval on economy hits new low: Quinnipiac poll – The Hill

Trump’s Economic Approval Hits Historic Low in Shocking New Poll

October 23, 2025
‘Chainsaw Man — The Movie: Reze Arc’ Review: Hit Manga Gets an Ultra-Violent, Surprisingly Emotional Big-Screen Adaptation – Yahoo

Chainsaw Man – The Movie: Reze Arc Review: A Brutal and Unexpectedly Emotional Big-Screen Adaptation

October 23, 2025
WHO launches new country guidance for health emergency coordination – World Health Organization (WHO)

WHO Launches Bold New Global Guidelines to Revolutionize Health Emergency Coordination

October 23, 2025

Categories

Archives

October 2025
M T W T F S S
 12345
6789101112
13141516171819
20212223242526
2728293031  
« Sep    
Earth-News.info

The Earth News is an independent English-language daily published Website from all around the World News

Browse by Category

  • Business (20,132)
  • Ecology (883)
  • Economy (904)
  • Entertainment (21,775)
  • General (17,768)
  • Health (9,945)
  • Lifestyle (917)
  • News (22,149)
  • People (905)
  • Politics (914)
  • Science (16,115)
  • Sports (21,404)
  • Technology (15,884)
  • World (887)

Recent News

Ecology researchers work to update Columbia River fish advisories — but face new challenges in climate change – Longview Daily News

Ecology researchers work to update Columbia River fish advisories — but face new challenges in climate change – Longview Daily News

October 24, 2025

Unlock the Secrets to Growing a Giant Pumpkin Using Science

October 24, 2025
  • About
  • Advertise
  • Privacy & Policy
  • Contact

© 2023 earth-news.info

No Result
View All Result

© 2023 earth-news.info

No Result
View All Result

© 2023 earth-news.info

Go to mobile version