[devin-ai-integration]

Summary

Both published skills install the kev package from a pinned commit: KEV_REF in skills/kev-deploy/scripts/kev_serve.py and skills/kev-finetune/scripts/kev_modal.py. On main both pins point at 84847f0 (#214). Three commits have touched kev/ since then, so a user who installs either skill today serves and trains with code main no longer runs. This PR moves both pins to the current tip of main, fc4a17e194bca35a48c57211392ec49ea54a94ac (#231). No SKILL.md or README names the commit (kev-finetune/SKILL.md only says "pinned commit"), so only the two constants change.

What the pins pick up (git log --oneline 84847f0..origin/main -- kev/ pyproject.toml uv.lock):

• serve: isolate accelerator probes and add an explicit device escape hatch #218 serve: isolate accelerator startup probes and allow explicit CPU selection. kev.device.select() runs one small allocation, matmul and sync in a fresh interpreter before any weights load, so a driver crash stays in that child process. kev.serve gets --device. CUDA graphs and fused kernels stay on by default, but only on NVIDIA (torch.version.hip is None), because those paths were only ever measured there.
• device probe: fail closed, try MPS after CUDA, and keep MLX off the torch probe #231 device probe: fail closed, try MPS after CUDA, and keep MLX off the torch probe. An explicit --device cuda on kev.train / kev.benchmark / kev.evaluate now runs the probe and fails closed: it raises instead of falling back to CPU. auto tries MPS if CUDA fails. Probe output is read from one marked line, so ROCm startup banners no longer break it.
• compare: write null diagnostics when none-of-the-above rows are absent #217 compare: write null diagnostics when none-of-the-above rows are absent. kev.compare writes null for an empty none-of-the-above diagnostic instead of NaN. Neither skill calls kev.compare.

What this means for each skill:

• kev-finetune: kev_modal.py launches kev.train ... --device cuda --dtype bf16, so every training container now runs the bf16 probe first. That costs one extra torch import in a child process (seconds). The upside: a broken GPU fails the trial before it starts instead of letting it train on CPU.
• kev-deploy: no change to what runs. kev_serve.py and kev_modal.load_server call Checkpoint.load("cuda", LoadOptions(dtype=bf16, cuda_graphs=True, fused=True)) and build kev.serve.Server directly. They never go through kev.serve.main() or select(). The pin still moves so both skills stay on one commit.

Serving numbers: not re-measured, and none changed. #218 and #231 touch kev/serve.py and kev/device.py, both on the serving list in AGENTS.md. That is why I checked whether the README serving table could be stale:

• Nothing on main is newer than the table. The reports it reads (runs/serve-*, runs/grouping-4b-h100*, runs/fused-*) were last committed 2026-09-24 and runs/serving-27b-r23 on 2026-09-28. The table section was last edited 2026-10-01 (Kev 1.0 release candidate: the four released checkpoints versioned as one family, with formal cards, a README family table, release notes and an asset builder (nothing published) #206).
• I don't think a re-measurement is needed. The only kev/serve.py change is in main(), where the device is picked: _serving_device() before the load instead of default_device(). For an NVIDIA CUDA device the LoadOptions that come out are the same as before (bf16, CUDA graphs on, fused when fla is installed, backend="auto"), and the timed request path (Server, kev.model, kev.cuda_graphs, kev.fused_qwen35) did not change. So no released checkpoint's serving path changed: 0.8B on L4, 4B on L40S, 9B on H100, 27B on B200.
• I also could not have re-measured. This environment has no MODAL_TOKEN_ID / MODAL_TOKEN_SECRET, so modal run modal_app.py::serving cannot run here.
• If this reasoning is wrong, the rows a re-measurement would replace are runs/serve-08b-l4, runs/serve-4b-l40s, runs/serve-9b-h100 and runs/fused-27b-b200. The table and model cards are left as they are and still trace to their committed reports (tests/test_conventions.py passes).

No earlier run of this automation has a PR open: #214 merged and #154 is closed, so this is a new PR rather than a push to an old branch. Nothing under evals/ or any frozen suite is touched.

Test plan

• git log --oneline 84847f0..origin/main -- kev/ pyproject.toml uv.lock: fc4a17e (device probe: fail closed, try MPS after CUDA, and keep MLX off the torch probe #231), 53cb0b5 (serve: isolate accelerator probes and add an explicit device escape hatch #218), 8ed2a2d (compare: write null diagnostics when none-of-the-above rows are absent #217)
• uv sync --extra serve (Python 3.12 venv; the box could not download 3.13) and uv run python -m pytest tests/test_skill_scripts.py tests/test_conventions.py -q: 40 passed
• git diff 84847f0..origin/main -- kev/serve.py: the only changes are device selection in main() and the NVIDIA-only guard on the CUDA graph / fused defaults; Server and the request path are unchanged
• Serving re-measurement: not run (no Modal credentials here; not expected to move, see above)

Link to Devin session: https://app.devin.ai/sessions/0b7ac5485ee34370a00de780897bfc50
Open in Devin Desktop: https://app.devin.ai/desktop/session/0b7ac5485ee34370a00de780897bfc50?variant=devin

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring