This release includes 1 breaking change 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
auth
rce_ssrf
Summary
AI summaryRemoves the VLLM_ENABLE_V1_MULTIPROCESSING=0 requirement, enabling thaw under vLLM's default V1 multiprocessing mode.
Full changelog
Summary
- Removes the
VLLM_ENABLE_V1_MULTIPROCESSING=0requirement for freeze / restore / load / pool / TP. - thaw now works under vLLM's default V1 multiprocessing mode.
- KV cache path still sets V1 MP=0 internally (scheduler state only reachable in V1-inproc/V0) — users don't have to know.
- Auto-sets
VLLM_ALLOW_INSECURE_SERIALIZATION=1on import so vLLM 0.19's msgspec IPC accepts function callables via cloudpickle fallback.
Validation
2×A40 RunPod, vLLM 0.19.0, Llama-3-8B:
- [x] weights-only freeze under V1 MP default
- [x]
thaw_vllm.load()+ generate under V1 MP default - [x]
thaw servepool + slot-warm + two completions - [x] TP=2 freeze + restore under V1 MP default
- [x] KV cache regression (conditional V1 MP=0 auto-set still works)
5/5 pass.
Test plan
- [ ] Merge
- [ ] Update README to drop the env-var note
- [ ] Bump minor version on PyPI
Breaking Changes
- Removed requirement for `VLLM_ENABLE_V1_MULTIPROCESSING=0`; thaw now works under vLLM's default V1 multiprocessing mode.
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
About Thaw
All releases →Related context
Related tools
Beta — feedback welcome: [email protected]