Give training its own GPU device

[lanxinger]

Summary

• Adapt ArthurBrussee/brush PR Give training its own GPU device ArthurBrussee/brush#526 through upstream head 65f19e0ada6221b6848bd9cda09b0a8ba83cf250.
• Give training a lazy dedicated GPU device so viewer work no longer shares its queue.
• Move displayed packed RGBA8 frames across devices through a host readback and viewer upload buffer.
• Generalize custom kernels over CubeRuntime and update the pinned Burn, CubeCL, and Cubek revisions.
• Make validation readbacks runtime-switchable so timing harnesses can disable synchronization.
• Reach CubeCL through Burn and centralize the remaining direct burn-cubecl surface in brush-cube.

Fork-specific adaptations

• Preserve native-MSL renderer paths, PPISP and appearance training, seeded training, LOD and refine behavior, JSONL metrics, iOS Metal selection, and robust PLY dataset detection.
• Route fork-only appearance, native-MSL optimizer/min-scale, and loss-test CubeCL consumers through Burn or brush-cube.
• Keep the Autodiff image-loss backward arm callable on the inner backend.
• Make repeated initExisting calls idempotent.
• Avoid a repository-wide typo exception by rewording the flate2 pin comment.

Upstream source: ArthurBrussee#526
Upstream head reviewed: 65f19e0ada6221b6848bd9cda09b0a8ba83cf250

Validation

• cargo fmt --all -- --check
• cargo check --workspace --all-targets --all-features
• cargo clippy --workspace --all-targets --all-features -- -D warnings
• cargo check --locked --workspace --no-default-features --all-targets
• Portable WASM and native-MSL WASM-fallback checks on wasm32-unknown-unknown
• Full Metal-backed all-feature workspace suite: 284 passed, 0 failed, 2 intentionally ignored
• All-feature private documentation with warnings denied
• Diff and whitespace checks
• Egg dataset smoke: 24 masked views at max resolution 512, 10 training steps, successful final export in 4.48 s; output was a valid 70 MiB binary PLY with 313,214 splats

The ignored tests are the documented large host-Metal appearance dispatch regression (~1.5 GiB peak) and native-Metal launch-geometry soak. The local cargo-deny binary is unavailable; the upstream head cargo-deny check is green.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-08-30T13:03:10.818538Z	c136345	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c13634540e

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Await registration of the host-provided device

tokio::sync::OnceCell::set is asynchronous, so assigning the returned future to _ drops it without ever storing burn. Consequently every initExisting(...) call appears successful, but the first later device() call opens the internal default device instead; training buffers then belong to that device and cannot be bound by the host WebGPU device as the brush-js API promises. Make this initialization path await set (or otherwise synchronously register the value) before returning success.

Useful? React with 👍 / 👎.

[lanxinger]

This is a false positive for the dependency version pinned by this PR. Cargo.lock resolves Tokio 1.53.1, where tokio::sync::OnceCell::set is synchronous: pub fn set(&self, value: T) -> Result<(), SetError>. Its implementation calls set_value before returning, so the host-provided device is stored before initExisting returns; no future is created or dropped here. The ignored Result only covers an earlier or in-flight initializer winning, consistent with the documented requirement that initExisting be called before anything touches the device. No code change is needed.