Repository navigation
Part of release 2026.10.6 (liblloyal 3.2.1).
CI tests the CPU that ships
No change to the headers: 3.2.1 is 3.2.0's code. What changes is the CPU llama.cpp is built for in CI.
Both CI builds of llama.cpp used its default native build, so each run compiled whatever the runner's CPU offered. The tests run for 3.2.0 landed on an Intel AMX runner, which compiled llama.cpp's AMX matrix multiply. At b11559 that kernel corrupts every sequence but the first that shares a batch on Qwen3.5 (upstream ggml-org/llama.cpp#29670, with its fix in #29671), and two Qwen3.5 cases with several branches decoding together failed. On every other runner they pass.
CI now builds llama.cpp for the AVX2 floor lloyal.node's Linux package ships, so the integration suites test the kernels that ship and give the same result on any runner. lloyal.node's builds don't reach the AMX kernel: its standard Linux package doesn't compile it, and its GPU pack offloads every layer. The llama.cpp pin moves once the fix is released upstream.
Use 3.2.1 rather than 3.2.0: its tests on main are green.
Engine: llama.cpp b11559.
What's Changed
- ci: llama.cpp is built for the AVX2 floor that ships, not the runner's CPU by @lloyal-research in #53
Full Changelog: v3.2.0...v3.2.1
