Nvidia published TensorRT-LLM v1.3.0rc25 on 31 August at 03:24:43 UTC. The change with the widest blast radius is that KV cache manager V2 became the default for roughly 21 model families. The release is marked a prerelease, which is the mitigating fact here — but the notes are still internally at war with themselves.
A claim with no number behind it
The justification given for defaulting V2 is improved scalability and stability. There is no baseline, no benchmark, no throughput figure, no memory-footprint comparison, and no named workload. For a component that decides how the attention cache is allocated across a serving fleet, "more stable" without a measurement is not an engineering claim an operator can act on.
The same page contradicts it
The release notes carry a Known Issues section, and KV cache manager V2 appears in it repeatedly — including a note that the V2 scheduler may intermittently crash or run out of memory on B200, Nvidia's current flagship. So the component being promoted to default for stability reasons is, on the same page, documented as capable of crashing on the newest hardware most likely to be running it. Both statements may be true — V2 could be better in aggregate and still have an open scheduler bug — but the notes never reconcile them, and a reader who stops at the highlights sees only the first.
What the common framing gets wrong
Release-note summaries treat "default" as a maturity signal: if the vendor switched it on, it must be ready. In prerelease channels that inference does not hold. Defaulting is how a project gets test coverage, and rc-tagged builds are where that is supposed to happen. The error is reading a release-candidate default as a production recommendation — and the honest reading of this one is that Nvidia is broadening exposure to V2 precisely because it still has open issues.
What an operator should do
Read Known Issues before the highlights, and pin the previous KV cache manager if you are serving on B200 until the scheduler notes clear. The escape hatch exists; the release simply does not lead with the fact that you might want it.