llama.cpp mmproj naming debate exposes duplication
The discussion argues that naming multimodal projector files as `[model-name]-mmproj-BF16.gguf` would make local model management cleaner and reduce duplicated `mmproj` blobs across quantized variants. The issue sits in the rough edges of multimodal GGUF packaging, where file names and folder structure still do too much work.
This is a small naming tweak with outsized UX impact: the ecosystem has outgrown the one-file-per-folder habit, and the current convention leaks implementation details into the user workflow. Hugging Face repos already show the mess: many multimodal models ship a shared `mmproj-BF16.gguf`, which makes downloads and local organization noisy. A model-prefixed filename would help batch downloads, avoid collisions, and make UI pickers easier to scan. The deeper fix is probably metadata-aware tooling, not just better names, so inference engines can resolve the right projector automatically. The fact that users are debating file naming at all is a sign multimodal support is mature enough to need packaging standards, not just conversion scripts.
DISCOVERED
5h ago
2026-04-24
PUBLISHED
6h ago
2026-04-24
RELEVANCE
AUTHOR
yc22ovmanicom