-
整理中文 model card 與 SSD streaming 定位
-
initial working 4k validation: SSD-streamed Qwen3.8-27B with 3GB RAM limit
Auto Upload Agent
- Fixed llama.cpp AMX compile flags to disable 13.9 GiB weight copy
- Added SSD streaming patch (LLAMA_MMAP_PREFETCH=0) to keep weights on SSD
- Fixed memwatch.py anonymous memory calculation (uses /proc/PID/smaps_rollup Anonymous field)
- Disabled weight repack (--no-repack) to avoid 3.5 GiB allocation
- Validated 4K context with 0.84 GiB anonymous memory (within 3 GB limit)
- Model weights (15.33 GiB) stay on SSD via mmap, faulted in on demand
- Pi coding agent configured with theoretical max context 262144
