fix(vram): report largest GGUF quant instead of whole HF repo for gallery size (#10700)

[Anai-Guo]

What / Why

Fixes #10700.

In the Install Models gallery, a model backed by a HuggingFace GGUF repo is reported with the size of the entire repository rather than the single GGUF file that would actually be downloaded. The reporter saw ornith-1.0-9b-mtp — a 9B model — listed as 71.2 GB · VRAM: 71.2 GB / "May not fit", because the repo ships many quantizations (Q4_K_M, Q5_K_M, Q8_0, F16, ...) and all of them were summed.

Root cause

pkg/vram/hf_estimate.go → sumWeightFileBytes adds up every weight file in the repo. That is correct for .safetensors/.bin shards (genuine parts of one model, all required together), but wrong for GGUF: the quantizations in a repo are mutually-exclusive alternatives — only one is ever downloaded and run.

Fix

sumWeightFileBytes now:

• Groups .gguf files by quantization variant, collapsing multi-part shards (model-00001-of-00003.gguf) onto a single variant key so shards of one quant are summed together.
• When GGUF files are present, reports the largest single variant as a conservative estimate instead of the whole repo.
• Leaves the non-GGUF path unchanged: .safetensors/.bin/.pt shards are still summed.

For the issue's repo this turns "sum of all quants" (~71 GB) into "largest quant" (a realistic single-file size), so the gallery no longer falsely flags small models as too large.

Tests

Added pkg/vram/hf_estimate_internal_test.go covering: multi-quant repo (largest variant wins), sharded variant (shards summed), safetensors-only repo (still summed), and LFS-size preference.

🤖 Generated with Claude Code

[mudler]

Good catch, thanks!