Thaw

thaw-native fix

The v0.3.0 wheel on PyPI was built before the plain-pinned-freeze fix (commit 2de24bf). On many hosts the write-combined pinned path capped freeze_to_file_pipelined at ~50 MB/s because CPU reads of WC memory are ~100× slower than plain-pinned reads. v0.3.1 uses plain pinned memory for the freeze pipeline (restore stays on WC, since its CPU-side work is writes).

Validated on H100 80 GB PCIe: freeze 0.05 → 2.96 GB/s, restore 0.05 → 3.99 GB/s. End-to-end ForkPool per-round: 21 s → 0.88 s.

ForkPool (new)

Pre-warmed subprocess pool: boot N vLLM engines once with real weights, hot-swap only KV per fork_completions call. Turns per-fork latency from ~340 s cold-boot into sub-second. See demos/fork_pool_rl.py and site/receipts/2026-04-20_h100_fork_pool_rl.json.

Install

pip install -U thaw-native thaw-vllm