1813035d50
Follow-up review on the CPU-only / NUMA hardening: - Reserve a flat MTP fraction in the CPU context fit when MTP engages but the draft KV cannot be byte-sized (mtp_overhead_fn is None): budget_frac otherwise skips the flat reserve, so the fit could pick a context that OOMs once the draft allocates. - Bound the NUMA local-placement decision on the smallest node's free RAM, not the largest. The loader is not pinned, so first-touch may land on any node; interleave a footprint that fits only the larger node instead of gambling on placement. - Capture the first launch's scheduler abort before the no-spec MTP fallback resets the stdout tail, then fold it into the terminal memo decision, so a fallback that fails for another reason does not drop the memo and let the UI replay the load. Unit tests and the offline simulations updated and passing. |
||
|---|---|---|
| .. | ||
| backend | ||
| frontend | ||
| src-tauri | ||
| __init__.py | ||
| install_llama_prebuilt.py | ||
| install_node_prebuilt.py | ||
| install_python_stack.py | ||
| LICENSE.AGPL-3.0 | ||
| node_prebuilt_pins.json | ||
| package-lock.json | ||
| package.json | ||
| setup.bat | ||
| setup.ps1 | ||
| setup.sh | ||
| Unsloth_Studio_Colab.ipynb | ||