thaw-vllm

The fork primitive for LLM inference. Snapshot a running session — weights + KV cache + scheduler state — and hydrate it into N divergent children that skip prefill. For RL rollouts, parallel coding agents, agent branching. Supports vLLM and SGLang.

thaw-vllm has been downloaded 9,245 times in total on PyPI, including 801 in the last 30 days. The latest version is 0.5.1, released Apr 23, 2026. It is distributed under the Apache-2.0 license.

Version0.5.1
Downloads
9.24k
LicenseApache-2.0
AuthorNils Matteson
UpdatedApr 23, 2026

Downloads

Weekly, last 90d.
Includes CI traffic.

VersionsTotal0.*
Range
View
Granularity
Group by
CI traffic
Stack: OffCI: Included2 / 17 series
Selected total18.49k
9.2kAll-time
801Last 30 days
13Last 24 h
<0.01/sPer second

Version distribution

Share of downloads by released version. Computed over the last quarter.

  • 01

    0.5.1

    825 downloads
    8.9%
  • 02

    0.3.0

    814 downloads
    8.8%
  • 03

    0.1.0

    695 downloads
    7.5%
  • 04

    0.2.1

    680 downloads
    7.4%
  • 05

    0.6.0

    602 downloads
    6.5%
  • 06

    0.5.0

    590 downloads
    6.4%
  • 07

    0.1.4

    542 downloads
    5.9%
  • 08

    0.2.0

    542 downloads
    5.9%
  • 09

    Other

    4.0k downloads
    42.8%

Guess the next day

Thirteen recent days of thaw-vllm downloads. Drag the green handle on the right to guess where day fourteen lands.

TRUTH1317SATSUNMONTUEWEDTHUFRISATSUNMONTUEWEDTHU