HF interop, spec-decoding selector, quant combobox, config safety, bonsai image #1

Merged

nirtal merged 1 commit from qwne38 into main

2026-10-08 18:34:05 +02:00

nirtal commented

2026-10-08 18:33:55 +02:00

Owner

tui/hf.py: download via hf_hub_download (standard HF cache layout:
refs/blobs/snapshots symlinked), hub/-style models_dir detection,
case-insensitive quant matching, search_models for the dashboard
'h' search (GGUF-tagged, downloads-sorted), list_repo_quants with
file sizes (repo-base prefix, filters mmproj/imatrix/shards/subdirs)
Model editor: split hf_spec into HF Repo + Quant combobox
(Enter opens greppable popup, arrows/Enter pick, custom quants allowed);
speculative decoding section replacing DFlash2:
off/mtp/dflash/ngram/mtp+ngram -> --spec-type mapping with per-type options
Dashboard: 'h' binds HF search -> prefilled add-model dialog;
sampling flags only passed when differing from llama.cpp defaults;
launch honors image env entries; detail view shows spec decoding
Image editor: environment variables (KEY=VALUE per line) +
preserve unknown keys (supports_fit) on save
Config safety: LLAMA_RUNNER_CONFIG_DIR/XDG isolation resolved per call,
save_config data-loss guard (ConfigDataLossError) + .bak + atomic write,
legacy dflash:true auto-migrates to spec_type
New image llama-cpp-bonsai:cuda13: PrismML fork prebuilt CUDA 13.3
binaries (PQ2_0/PTQ1_0 ternary GGUFs)
tui/testing.py: isolated_config() for headless tests

- tui/hf.py: download via hf_hub_download (standard HF cache layout: refs/blobs/snapshots symlinked), hub/-style models_dir detection, case-insensitive quant matching, search_models for the dashboard 'h' search (GGUF-tagged, downloads-sorted), list_repo_quants with file sizes (repo-base prefix, filters mmproj/imatrix/shards/subdirs) - Model editor: split hf_spec into HF Repo + Quant combobox (Enter opens greppable popup, arrows/Enter pick, custom quants allowed); speculative decoding section replacing DFlash2: off/mtp/dflash/ngram/mtp+ngram -> --spec-type mapping with per-type options - Dashboard: 'h' binds HF search -> prefilled add-model dialog; sampling flags only passed when differing from llama.cpp defaults; launch honors image env entries; detail view shows spec decoding - Image editor: environment variables (KEY=VALUE per line) + preserve unknown keys (supports_fit) on save - Config safety: LLAMA_RUNNER_CONFIG_DIR/XDG isolation resolved per call, save_config data-loss guard (ConfigDataLossError) + .bak + atomic write, legacy dflash:true auto-migrates to spec_type - New image llama-cpp-bonsai:cuda13: PrismML fork prebuilt CUDA 13.3 binaries (PQ2_0/PTQ1_0 ternary GGUFs) - tui/testing.py: isolated_config() for headless tests

nirtal added 1 commit

2026-10-08 18:33:55 +02:00

HF interop, spec-decoding selector, quant combobox, config safety, bonsai image 02621dd954

- tui/hf.py: download via hf_hub_download (standard HF cache layout:
  refs/blobs/snapshots symlinked), hub/-style models_dir detection,
  case-insensitive quant matching, search_models for the dashboard
  'h' search (GGUF-tagged, downloads-sorted), list_repo_quants with
  file sizes (repo-base prefix, filters mmproj/imatrix/shards/subdirs)
- Model editor: split hf_spec into HF Repo + Quant combobox
  (Enter opens greppable popup, arrows/Enter pick, custom quants allowed);
  speculative decoding section replacing DFlash2:
  off/mtp/dflash/ngram/mtp+ngram -> --spec-type mapping with per-type options
- Dashboard: 'h' binds HF search -> prefilled add-model dialog;
  sampling flags only passed when differing from llama.cpp defaults;
  launch honors image env entries; detail view shows spec decoding
- Image editor: environment variables (KEY=VALUE per line) +
  preserve unknown keys (supports_fit) on save
- Config safety: LLAMA_RUNNER_CONFIG_DIR/XDG isolation resolved per call,
  save_config data-loss guard (ConfigDataLossError) + .bak + atomic write,
  legacy dflash:true auto-migrates to spec_type
- New image llama-cpp-bonsai:cuda13: PrismML fork prebuilt CUDA 13.3
  binaries (PQ2_0/PTQ1_0 ternary GGUFs)
- tui/testing.py: isolated_config() for headless tests