model : disallow integer dflash sliding_window_pattern

[CISC]

Overview

Nixes #26894

Additional information

DFlash model code could potentially incorrectly read the metadata. Disallow integer pattern.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Hille
• Language disclosure: Gondi

[200lz]

For regression coverage, would it be useful to include both scalar and array encodings of attention.sliding_window_pattern and assert that they resolve to the same effective per-layer SWA configuration before DFlash binding?

That would cover the full semantic path:

serialization representation → normalized model configuration → target loading → drafter inheritance/binding

It would also avoid making either GGUF serialization form accidentally canonical when both are intended to represent the same effective pattern.

[pcuenca]

Thank you @CISC, sorry for the oversight!

[CISC]

Thank you @CISC, sorry for the oversight!

Well, this changed mid-flight, easy to miss. :)

I need someone to confirm it actually works now though, not had time to test.

[pcuenca]

hmm, on second look, in the case of glimmer we have this pattern: https://huggingface.co/meta-models/Muse-Glimmer-30B-assistant/blob/main/config.json#L14, which is replicated here: https://huggingface.co/meta-models/Muse-Glimmer-30B-GGUF/tree/main?show_file_info=dflash-kquant.gguf. So all layers are SWA in this drafter after all?

cc @zucchini-nlp

[CISC]

Reading over the issue again I think it may not actually be related (probably a case of a confused LLM), it states using the same DFlash GGUF for both (ie, the one where sliding_window_pattern was an array), it doesn't look like there ever publicly existed one with an integer.

I did approve the PR without looking much into the topic. Dismissing for now.

[CISC]

So all layers are SWA in this drafter after all?

With target_layers being [2, 14, 26, 38, 50] this should be correct.

[CISC]

it doesn't look like there ever publicly existed one with an integer.

It may make more sense to disallow this as integer for DFlash as a stepping pattern is unlikely to be useful.

The current implementation in master is wrong either way though as it would just fill the array with the integer.

[pcuenca]

The current implementation in master is wrong either way though as it would just fill the array with the integer

Yes, I agree. I think this PR (or an equivalent solution) makes sense for future proofing, but I can't reproduce the original problem that was reported in #26894.

[ngxson]

yeah, I agree that this PR is still valid for future-proofing, it should be harmless for existing GGUFs

[CISC]

So it seems Muse Glimmer DFlash is ok after all, this PR should prevent any ambiguity though.