Thaw

v0.2.1 Breaking

This release includes 2 breaking changes for platform teams planning a safe upgrade.

Published 1mo AI Agents & Assistants

✓ No known CVEs patched

✓ No known CVEs patched in this version

Topics

agents inference kv-cache llm reinforcement-learning sglang

+1 more

vllm

Affected surfaces

breaking_upgrade

Summary

AI summary

Strict mode now raises on native extension load failure, eliminating the slow Python fallback path.

Full changelog

Freeze throughput: 0.19 → 19.62 GB/s on H100 (104× rewrite of the pipelined freeze path, O_DIRECT
pwrite, double-buffered WC pinned memory)
KV cache freeze/restore wired through the pipelined Rust path — no more Python-loop-bound
serialization for the competitive-moat KV snapshot
New on-disk KV format: standard .thaw file + .meta JSON sidecar. Legacy single-file readers
preserved for backward compat
Unified restore_from_bytes_auto — tries cudaHostRegister zero-copy first, falls back to
memcpy-through-staging on failure. One entry point, no caller-side flag
WC pinned-buffer thread-local cache — amortizes cudaHostAlloc across freeze/restore calls.
Critical for thaw serve hot-swap
Strict mode default flipped — failed native extension load now raises instead of silently falling
back to 50 MB/s Python path. Opt out via THAW_ALLOW_PYTHON_FALLBACK=1
Slow PyO3 siblings removed (freeze_to_file / restore_from_file) — silent 100× footgun
eliminated

pip install -U thaw-native thaw-vllm

Strict mode default flipped: failed native extension load now raises instead of silently falling back to the 50 MB/s Python path. Opt out via THAW_ALLOW_PYTHON_FALLBACK=1
Removed slow PyO3 siblings freeze_to_file and restore_from_file

Weekly OSS security release digest.

The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.

No spam, unsubscribe anytime.

Share this release

Track Thaw

Get notified when new releases ship.

About Thaw