[ggerganov]

Overview

-hf <repo>:F16 picked the BF16 model when a repo had both, because the tag was matched anywhere in the path: the pattern F16[.-] also matches inside Llama-2-7B-BF16.gguf. The tag now has to start at a token boundary (start of path or after -, ., /), which also stops :Q4_XS from matching IQ4_XS and similar. A tag no file carries now resolves nothing with the usual error listing the available files, instead of silently downloading the wrong dtype.

tests/test-model-resolution.cpp: new F16/BF16 and Q4_XS/IQ4_XS cases, a pinned listing for ggml-org/Llama-2-7B-GGUF which ships both, and 17 pins corrected that had frozen the old behavior. The suite passes on all 172 pinned real listings.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES
  • Harness: dsh:llama.cpp
  • Hardware: Mac Studio, M5 Ultra, 256GB
  • Model: ggml-org/GLM-5.3-Flash-GGUF
  • Stats: ◷ 33.5m · ⚙ 19.3s · ↑ 91k · ↓ 66k · ⇈ 423.6 tok/s · ⇊ 36.1 tok/s · ≡ 112 · ⌨ 12

[ServeurpersoCom]

LGTM, and good to see the real listings harness bite: 17 pins had frozen the F16 -> BF16 bug!

Non-blocking suggestion: the left boundary could also accept _. The harness only pins spec-conformant names so it does not cover this, but some real repos use it as separator, e.g. stepfun-ai/Step-3.5-Flash-GGUF-Q4_K_S ships step3p5_flash_Q4_K_S-00001-of-00012.gguf, where :Q4_K_S would no longer match since an explicit tag has no fallback. (^|[-./_]) would keep it working while still separating F16/BF16 and Q4_XS/IQ4_XS.