00566662c6
Inactive-cache rows now emit the newest snapshot holding config.json plus safetensors weights as their load_id, since the load consumes that path directly and a metadata-only newest revision would fail an eligible model. Ordinary offline guards no longer join a forced window (their block runs correctly under either env state), so concurrent online work sees the narrowest possible override; forced guards still share windows. The generation-time native template reload honors the load's local_files_only flag and resolved path, so a tool-calling turn on an auto-loaded model cannot download tokenizer files mid-Send. Every GGUF variant scan behind the model-kind gate now runs with an abortable 30s timeout (matching the inventory calls' own bound) so one hung request cannot gate Send forever. An HF cache snapshot registered as a custom scan folder dedupes against its cached row by expanding snapshot paths to their cache root, so shared files consume one load attempt. |
||
|---|---|---|
| .. | ||
| data_recipe | ||
| export | ||
| inference | ||
| rag | ||
| training | ||
| __init__.py | ||
| _torchao_stub.py | ||
| import_guards.py | ||
| tool_healing.py | ||