Semlith Files

The CPU lane can be switched off again, beside another lane, and a Mac with defaults runs the Neural Engine alone from the first batch. No store format or settings.json change and no re-index.

Changed

• The CPU switch turns off beside another lane. From 0.37.0-rc.6 the CPU lane was always on and semlith accel off cpu always refused. It is now accepted while another lane (GPU, Neural Engine, CUDA, TensorRT for RTX, OpenVINO, llama.cpp or a registered remote lane) is on and can run on this machine, from the CLI or the Settings › Performance switch. With no such lane the CPU is on and locked, and accel off cpu exits non-zero with "the CPU lane stays on while no other lane is on; switch one on first". Turning off the last other lane is allowed; the CPU is then held on, and its saved off applies again when a lane comes back. accel on cpu now saves true instead of clearing the saved value.
• The CPU is off by default beside the Neural Engine. With nothing saved it is off while the Neural Engine is on and available (a Mac with Apple silicon) and on otherwise, so a Linux or Windows machine with a GPU lane keeps the CPU beside it. A saved choice wins, so a saved cpu: true now runs the CPU beside the Neural Engine, where 0.49.0 dropped it once the Neural Engine was ready. With no lane ready and none on its way the CPU still carries the run whatever its switch says, and it still re-embeds any vector a lane returned unsound.
• The GPU waits beside a starting Neural Engine. It embedded while the Neural Engine started or compiled. With the Neural Engine ready or on its way the GPU now runs beside it only when gpu-beside-ane is on. A failed Neural Engine is not used, and the GPU and CPU take over as before.
• GET /api/accel's CPU row tells the truth. saved is the saved choice or the default, source is default, saved or environment, locked is true only while no other lane is on, and locked_reason carries the new message only then, null otherwise. The portal's CPU row is a normal switch while unlocked and says "always on" only while locked. The library exports accel::CPU_LOCKED and keeps accel::CPU_ALWAYS_ON. SEMLITH_ACCEL still overrides every switch. Rolled back to 0.49.0, a saved cpu: false reads as on there again.
• An index run reads the lane switches every 100 ms, not on every turn. The scheduler's loop turns thousands of times a second while batches are out, and each turn read settings.json and probed every lane switched on. On the M1 that cost a run about a tenth more CPU time than lanes pinned by SEMLITH_ACCEL. Lane readiness and failure are still read on every turn, and a switch lands within 100 ms.
• tests/airgap.sh checks the note, not the file name. The pictured query must print no "image results skipped" note, instead of the results being grepped for logo.png.

Install

macOS and Linux

:::sh
curl --proto '=https' --tlsv1.2 -fsSL https://raw.githubusercontent.com/semlith/semlith/main/install.sh | sh

Windows

:::powershell
irm https://raw.githubusercontent.com/semlith/semlith/main/install.ps1 | iex

Or download the archive for your platform below, unpack it, and put semlith on your PATH. Intel macOS has no prebuilt binary: ONNX Runtime no longer publishes x86_64 macOS builds.

Verify a download against SHA256SUMS before running it, SHA256SUMS against its signature (release key SHA256:a5oJmrakIwDHEWTDWqWe51Ttt6m2tQ0dsGzIqyxDuxs) with echo "semlith-release ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAICMTHiHMzeSHzzc+8UDZHUzJcvb6XGUBb+TJdFes91xm" > allowed_signers and ssh-keygen -Y verify -f allowed_signers -I semlith-release -n file -s SHA256SUMS.sig < SHA256SUMS, and its provenance with gh attestation verify <file> --repo semlith/semlith --signer-workflow semlith/semlith/.github/workflows/release.yml.