I built thaw because forking an LLM agent is absurdly wasteful today. When an agent explores N branches — RL rollouts, best-of-N, parallel coding attempts — each branch re-runs prefill over the same shared context. You pay for the same prompt N times.

thaw snapshots a live inference session — weights, KV cache, scheduler state, and the prefix-hash table — and hydrates N children that diverge from the fork point without re-prefilling. It's `git branch` for a running model.

The receipt (H100 80GB, Llama-3.1-8B, real hardware): a pre-warmed pool boots once in 22.3s, then each fork round of 4 branches × 64 tokens runs in 0.88s median. Cold-boot equivalent would be ~340s/round — ~400× amortized. All rounds bit-identical at the fork boundary. Full JSON receipt + reproducer in the repo, nothing hand-waved.

NVIDIA shipped Dynamo Snapshot last week for fast pod cold-starts — and they free the KV cache before checkpoint, by design. thaw is the opposite bet: preserve the KV cache so a fork is near-free. Different problem, opposite mechanic.

pip install thaw-vllm. Works with vLLM and SGLang, Apache-2.0.

https://github.com/thaw-ai/thaw

I'm a solo dev and this is the thing I most want feedback on: is the fork primitive the right shape, or do people want it wrapped in a framework(LangGraph/TRL) node instead? Happy to go deep on the KV-restore internals.