[Jordan Olshevski (jveski)]

Makes GPUDirect model loading with Racer and vLLM work out of the box. The released racer-dataplane image now includes CUDA and cuFile, and in the default auto mode it touches no GPU until the first GPU read, so one image serves GPU and CPU nodes. The operator now checks racer-dataplane-config, the webhook mounts the client socket into init containers, and GPU, cuFile, and landing metrics are exported. The racer-vllm plugin is now a single module, and the Python packages ship as wheels on each GitHub release instead of PyPI.

[Copilot]

🔵 Needs a closer look

Human assessment is needed because an unpublished finding indicates that required GPU E2E runs can pass without exercising the dataplane when the socket is unset.

No open issues 🎉

_{Review effort: Balanced}

[Copilot]

🔵 Needs a closer look

Human assessment is needed to evaluate a critical gap in signed release inventory and digest-pinned signature-verification coverage for the new racer-vllm image.

No open issues 🎉

_{Review effort: Balanced}

[Copilot]

🔵 Needs a closer look

Human assessment is needed because checkpoint layout changes can bypass artifact pinning and silently mix new weights with previously fetched configuration and tokenizer files.

No open issues 🎉

_{Review effort: Balanced}

[Copilot]

🟡 Changes recommended

Shard discovery mishandles formatted rank placeholders, omitting other ranks’ ETags and causing valid models to fail pin checks with ModelChangedError.

1 open issue

• Detect all format fields when discovering tensor-parallel model files · New

_{Review effort: Balanced}

[Copilot]

🔵 Needs a closer look

Native CUDA/cuFile interfaces, cross-container GPU identity, and multi-GPU RDMA buffer lifetimes require hardware review because simulation and same-process tests cannot establish driver/NIC correctness or DMA safety.

No open issues 🎉

1 resolved since last review

• Detect all format fields when discovering tensor-parallel model files

_{Review effort: Balanced}