malaikat 0.2.0

malaikat 0.2.0

Unified Windows + Linux launcher for AMD Strix Halo (gfx1151) ROCm inference.

Windows keeps the lemonade windows-rocm-gfx1151 zip, PATH/DLLs, and high-priority process class.
Linux builds llama.cpp against system ROCm, sets ROCBLAS_USE_HIPBLASLT=1, and passes --fit off so a ctx_size of 0 stays at the model max (262144) instead of being silently shrunk.

Default model is Unsloth Qwen3.6-35B-A3B MTP MoE (not bundled).

Linux all-in-one

This release includes malaikat-0.2.0-linux-amd64: launcher + lemonade-sdk llamacpp-rocm b1311 (Ubuntu ROCm gfx1151). GGUF weights are not bundled.

chmod +x malaikat-0.2.0-linux-amd64
./malaikat-0.2.0-linux-amd64 serve -m path/to/moe-mtp.gguf

First run extracts the ROCm runtime to ~/.cache/malaikat/runtime. Optional config: coding.example.yaml in this release.

API: http://127.0.0.1:8080/v1

malaikat 0.1.0

malaikat 0.1.0

All-in-one Windows exe for AMD Strix Halo (gfx1151) ROCm inference.

Bundled: malaikat launcher + lemonade-sdk llamacpp-rocm b1310 (Windows ROCm gfx1151).
Not bundled: GGUF model weights — pass -m or a config.

Quick start

.\malaikat-0.1.0-windows-amd64.exe serve -m path\to\moe-mtp.gguf

First run extracts the ROCm runtime to %LocalAppData%\malaikat\runtime. Optional config: see coding.example.yaml in this release.

API: http://127.0.0.1:8080/v1