v0.1.1

Highlights

• The expert cache follows the whole card's free memory (NVML): it shrinks when another program moves in and grows back when it leaves.
• scripts/agent_bench.py --squeeze replays an agent session under memory pressure. Results are in BENCHMARKS.md.

What's Changed

• chat: --prefill-thinking opens the think block on a fixed prefill by @joa in #12

Full Changelog: v0.1.0...v0.1.1

v0.1.0

First release of Phobos 🎉

• phobos-cli: Inference engine for GGUF models.; OpenAI-compatible or a simple REPL
• phobos-bench: Measure prefill and decode throughput
• phobos-compile: Compiles Phobos source code to PTX
• phobos-cache: Utilities for the local compiler cache

This release includes a pre-warmed cache for all supported SM architectures.

Dependencies: NVIDIA driver that supports CUDA 13.0.

toolchain-llvm-22.1.7

LLVM and MLIR 22.1.7, static, for the release workflow to link against. Not a Phobos release.