[ Request/Plea ] Let the Open Source Community Support You

Description of errors

How about you consider documenting structural changes?

Anyone wanting a custom finetune model is, as you have designed, stuck to 1.0.1 using heavily rewritten converters or paying you to convert things by presumably a closed-source tooling you won't ever open up. That's a great way to ensure only the bare minimum of adoption, and only by companies you already have deals with.

AMD; Charging companies to get special one off models converted and dependent on you for updates is NOT smart business, getting wider adoption IS. You sell chips not models, and it's utterly Wrong in both a business software development sense. Wrong as in the clear incorrect choice for literally any business.

You guys seem somehow new to open source. We have conventions such as what a Major vs minor version means -- you hide breaking changes under x.x.n bugfix level bumps. You actively make other developer's lives harder, and the only thing we can do is abandon your updates.

I watch. I see what you're all coding individually. I cherry pick. It turns out you didn't need to change any of the closed parts to fix any of the bug reports.

AMD, you can do better for your community. And that will sell your NPU chips. Don't break yourselves trying to defend something that IS replaceable.

Terrible idea overall. Might want to discuss what "open source" means, how the ecosystem actually works, and actually open things up. Until then, it's been nice. I'll maintain my 1.0.1 models, but otherwise everyone good luck.

Attach any links, screenshots, or additional evidence you think will be helpful.

https://github.com/Atomic-Germ/FastFlowLM_v1 on the 1.0.1-add1 branch Is significantly more capable than the official 1.0.3, without breaking 1.0.1 models.
https://huggingface.co/Atomic-Germ has dozens of models.
https://github.com/Atomic-Germ/q4nx-build can convert from tensors to q4nx in one shot.
https://github.com/Atomic-Germ/flm-test proves the model thinks, sees, and uses tools
https://github.com/Atomic-Germ/flm-add makes using them easy.

One person put that ecosystem together. Imagine what the whole community could do if you actually opened the damn code. The AMD team itself isn't enough, and FastFlow's growth and adoption is so bad most people don't even know what the NPU is for.

Try working out an explanation to shareholders as to why AMD thought it was a good idea, or even a sensible idea, to hamstring its own hardware and give people less reason to buy your chips. Does it seem convincing? No it doesn't of course it doesn't, because it isn't.

If I come off harshly, that's more passion than anger. Judge me and my point separately please. The title is the message.

Start by reading the issue discussion and comparing the linked FastFlowLM 1.0.1-add1 branch with official 1.0.3, then inspect q4nx-build for the cited conversion workflow. A useful outcome would need a concrete, maintainer-approved scope for documenting structural changes or opening conversion tooling; the issue names no repository file or test.