·
3 commits
to main
since this release
a27868c
This commit was signed with the committer’s verified signature.
raffaelschneider
Raffael Schneider
Per-token inference benchmarking. New `SYS_PERF_COUNTER` syscall (71) exposes RDTSC to shards, gated by `PLEDGE_PERF`. The llama-inference shard measures per-token cycle counts with machine-parseable output.
Includes native Linux baseline (`scripts/bench-native.c`) using the same LCG-generated model for direct comparison, and `coconut-prof.py --baseline` for side-by-side reporting.
Phase 3: Inference Stack (6 of 6) — Phase complete
What's next
Phase 4 focuses on hardening and multi-vendor support:
- Security audit (external)
- Fuzzing campaign — already started, found and fixed 2 bugs
- NVIDIA GPU HAL (Hopper/Blackwell)
- ARM64 port + Apple GPU
- Network shard (RDMA/GPU-Direct)
- Formal verification (Verus)
See issues for details.