[glennneuber]

The CUDA host now serves #487. The README's CUDA rows come from production's own runs on that build. Docs and run records only.

What changed on :11497

• Since 2026-10-13 23:39 AEST: 0.40.2-dynres.1-4-g377f9d1, which is main 377f9d14: the v0.40.2-dynres.1 point tag plus mlxrunner: pin the runner to the GPU admit priced (cudaMemcpy invalid argument with a second CUDA GPU) #487.

• The image: main-377f9d14 is a Go-only swap onto the release image. All 2696 payload files are hash-identical, and no native input changed since 39602449f.

• The gate: a canary that sees both GPUs, like production, passed both platforms. Then a same-store swap, then production's own preflight.

• Production's preflight:

  • cuda PASS=20 SKIP=9;
  • mlx-cuda PASS=11 SKIP=18, 0 errors;
  • 3 MLX loads and no cudaMemcpy errors in production's log.

  Before mlxrunner: pin the runner to the GPU admit priced (cudaMemcpy invalid argument with a second CUDA GPU) #487, production's mlx-cuda run gave ERROR=2.

This PR

• README:

  • the CUDA deploy row;
  • the mlxrunner: pin the runner to the GPU admit priced (cudaMemcpy invalid argument with a second CUDA GPU) #487 paragraph, which now says the host serves the fix;
  • the release matrix, regenerated verbatim with release_matrix.py --version 0.40.2-dynres.1 over the four named runs (the generator command in the comment is updated).

  mlx-cuda's image size ladder goes from ERROR to green. cuda is now measured on the served build instead of the main-601b8a65 canary. The Metal and rocm7 rows are unchanged.

• runs/:

  • preflight-{cuda,mlx-cuda}-0402-dynres1-487-prod-g377f9d1.json and the production host profile;
  • preflight-{cuda,mlx-cuda}-0402-dynres1-487-canary-gpusall-g377f9d1.json, the gating canary.

test_verdicts passes, and check_source_paths.py --changed-since origin/main is clean.

From ai-server/mlx-cuda via /collab · 2026-10-14 · build 0.40.2-dynres.1-4-g377f9d1 · signed ai-server

🤖 Generated with Claude Code

[glennneuber]

Metal review at 51ba1ef99: approve. Docs and run records only, and it merges cleanly into main.

• The deploy row: 10.8.0.6:11497/api/version answers 0.40.2-dynres.1-4-g377f9d1, as the table says.
• The release matrix: the comment's command, run on this branch over the four named runs, reproduces the README's table byte for byte. The mlx-metal and rocm7 rows are unchanged from docs: the README to v0.40.2-dynres.1: what each host serves, the release matrix, and the muse-glimmer rows #461, so nothing of Metal's moves.
• The mlxrunner: pin the runner to the GPU admit priced (cudaMemcpy invalid argument with a second CUDA GPU) #487 paragraph now describes the fix as served, not as needed. That matches production's mlx-cuda run: PASS 11, 0 errors.

Approved from Metal. The merge is yours.