[ServeurpersoCom]

Overview

This PR is a study to let #29584 land without regressions: it pins model resolution on real Hugging Face repo layouts, and along the way it fixes the master bugs this corpus revealed, from imatrix files listed as models and missed split names to unpaired -vision mmproj, MTP heads in subfolders, _ separated quants and draft-only repos.

Additional information

I scanned the listings of 1237 GGUF repos from the major publishers on Hugging Face, ran the resolver on 7474 requests over them, and checked every pick against the file's actual GGUF header read with HTTP range requests. From that corpus, 63 real listings covering all 94 file naming patterns are pinned in test-model-resolution: master fails 25 of these checks, and the fix passes them all with no regression across the full corpus.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES

[angt]

find_token() keeps only the leftmost token, so Vision masks a later mmproj:

before:

$ LLAMA_CACHE=$(mktemp -d) llama download -hf leafspark/Llama-3.2-11B-Vision-Instruct-GGUF:F16
Downloading Llama-3.2-11B-Vision-Instruct-mmproj.f1… ───              22%
Downloading Llama-3.2-11B-Vision-Instruct.f16.gguf                     0%
^C

after:

$ LLAMA_CACHE=$(mktemp -d) build/bin/llama download -hf leafspark/Llama-3.2-11B-Vision-Instruct-GGUF:F16
Downloading Llama-3.2-11B-Vision-Instruct-mmproj.f1… ─────────────── 100%

[angt]

I still think it's the opportunity to follow the initial rules more strictly, like each component is delimitated by a -, etc.

Almost everything was specified here: https://github.com/ggml-org/ggml/blob/master/docs/gguf.md

[ServeurpersoCom]

Thanks, that is exactly what this PR must not do. I'll rework it to pin master's current behavior only, known bugs included, so it changes nothing in the resolver and stays a pure test harness. Improvements can come later, one at a time, on top of it.

[angt]

Thanks, that is exactly what this PR must not do. I'll rework it to pin master's current behavior only, known bugs included, so it changes nothing in the resolver and stays a pure test harness. Improvements can come later, one at a time, on top of it.

Okay, I agree

I was confused by the download.cpp patch that has been removed.

[ServeurpersoCom]

Rebased on master. This harness helped me spot a few resolution regressions in other PRs, so it might be worth having in the tree, no rush though.