v0.1.01b

Full Changelog: v0.1.0b...v0.1.01b

release notes version 0.1.01b

features

oneclick setup

just pick a project folder, and everything should be auto set up: the venv, project.json, everything should be all set up.

solution oriented erroring

in setup, a common feedback we got was "error x is happening, what do I do?" these errors were pretty repetitive. So I implemented solution-based erroring, which reads errors (ie pytorch requires python version >3.10) and outputs a solution (ie rerun python installer). its kind of buggy and not great but the scaffolding is there.

example configurations

now instead of picking a model, picking a dataset, and wondering about the dozens of LoRA and QLoRA settings, just click an example and start training.

multiple backends

Axolotl AI and Unsloth AI are now compatible with Mama (untested). These popular (and very similar) backends are now exportable from the mama application.

known bugs

• see issues
  I would like to point out that despite most export options not working, the Colab export option is fully working and tested.

EDIT: i fixed a bug on linux which caused the pywebview window to not open at all. I patched this bug and rebuilt it (using pyinstaller main.spec) on my personal machine (EndeavourOS on x86_64 with Hyprland) as the file mama_linux_new.zip, which may work better for your system, it may not.

in any case, if this bug appears, simply run the app like this:

WEBKIT_DISABLE_DMABUF_RENDERER=1 /path/to/mama/executable

and know that this issue is fixed for future versions.

v0.1.0b

Full Changelog: v0.0.50b...v0.1.0b

first official release that's not beta

official is a strong word

there are still many bugs, but I have deemed the app working enough to begin its playtesting journey on a bigger scale.

known bugs

linux

it won't even launch unless maybe you're on ubuntu. Largely untested territory, but I havent gotten it to work on EndeavourOS yet, not even after pre-installing everything. Good luck.

windows

on windows, when the app runs a python script in the background (including the main training script! big sad), a bunch of terminal windows will open. The python script will immediately stop when those windows are closed, so you have to leave them open. They go away on their own once the task is finished.

the auto updater

she's a little bitchy, but basically to have the app update itself, you have to

macOS

• press download
• press install and restart
• (wait for a minute)
• press restart
• (the app will freeze and become unresponsive)
• force-quit the app

but the good news is that after that it will relaunch itself as the newer version, so it has the capability to patch itself in the future.

windows

• not implemented (it may crash, it may be fixed. I'm honestly not sure at this point, I've lost access to a competent windows computer.)

HF_TOKEN not set!

when you do some calls to the Huggingface API, especially in regards to downloading a model, or calculating a minimum FLOPs estimate to fine tune it, the app may say an error that asks you to set your huggingface token.

For right now, the solution is to open a terminal and permanently add your token via environment variable:

HF_TOKEN = token

or to just wait (the model will download and cache anyway, even if it says "error: unknown type"). the FLOPs estimate may not work, but that's ok.

resource management

when you train a model that's a little too tough for your computer, it will seize up and become unresponsive, and the only way around this is to hard-restart.

unofficial program

there's no official payments I make or anything like that, so apple and microsoft mark the program as untrusted, which is pretty annoying on first install.

systems that are tested and pretty much work

• macOS m1 8gb macbook air laptops
• windows 10 with an nvidia-capable gpu

systems that definitely don't work right now

• linux on hyprland
• arch linux and variants

things that are untested

• linux that is not EndeavorOS with hyprland (ideally Ubuntu with a common DE such as wayland based)
• ROCm
• Tensorflow
• the template import button (may get removed in the next major release)
• windows auto-updater (since v0.0.53b)
• windows 11 as a whole (my machine has windows 10)
• arm64

future features

• import your own training script support
• an automatic crash/bug report feature
• Axolotl AI integration
• cluster gpu support
• export to another computer remotely support
• connect to your google colab support
• app hibernation while training
• huggingface "friendly interface" login support

v0.0.53b

linux deployment patch 2 (general)

v0.0.52b

linux deployment patch

v0.0.51b

redeploy test linux + windows

v0.0.50b

splash screen

v0.0.49b

the one where I actually got auto-update to work

v0.0.48b

auto updater bug fix 5

v0.0.46b

updated auto-updater

v0.0.45b

auto-updater test