Update Buildkite pipeline for the new JuliaGPU cluster

[maleadt]

The JuliaGPU Buildkite agents have moved to a dedicated cluster, where the single juliagpu queue has been split into per-backend queues. Steps now select agents using queue: "cuda", queue: "rocm" or queue: "oneapi" instead of queue: "juliagpu" combined with a cuda/rocm/intel tag. Steps that targeted queue: "juliagpu" without a backend tag now use the cluster's default queue.

The old benchmark queue is gone as well; benchmark agents now live in the backend queues with a benchmark tag, so the CUDA benchmark step in benchmarks.yml now targets queue: "cuda" with benchmark: "*" (keeping the gpu: "rtx2070" pin).

Queues are cluster-scoped, so steps targeting the juliaecosystem queue (the macOS Metal tests of MLDataDevices and WeightInitializers, and the CPU benchmark step) can no longer be uploaded from within a pipeline running on the JuliaGPU cluster. They have been moved verbatim into a new .buildkite/pipeline-julia.yml, which will be wired up to a separate ecosystem-cluster pipeline once it is back online. Until then those steps will not run; in particular, the (currently disabled) benchmark aggregation step temporarily only sees CUDA results, no CPU ones.

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request restructures the Buildkite CI configuration by moving macOS Metal and CPU benchmark steps into a new dedicated pipeline file (.buildkite/pipeline-julia.yml) and updating queue names (e.g., from juliagpu to specific queues like cuda, rocm, and oneapi) across various testing configurations. The review feedback highlights that the hardcoded encrypted SECRET_CODECOV_TOKEN in the new pipeline file will fail to decrypt on the juliaecosystem cluster due to differing decryption keys, and recommends configuring this token via the Buildkite UI instead.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

Since the juliaecosystem queue runs on a separate cluster, it will use a different Buildkite agent stack with different decryption keys. Consequently, the encrypted SECRET_CODECOV_TOKEN copied verbatim from the GPU cluster pipeline will fail to decrypt on the new cluster.

Recommendation:
Remove this hardcoded encrypted token and instead configure CODECOV_TOKEN as a pipeline-level environment variable in the Buildkite UI for the new ecosystem-cluster pipeline. This ensures portability and avoids decryption errors.

  XLA_REACTANT_GPU_PREALLOCATE: "false"