Skip to content

release: 0.231.0 - #853

Merged
dmmdea merged 1 commit into
mainfrom
release/0.231.0
Nov 11, 2026
Merged

dmmdea merged 1 commit into
mainfrom
release/0.231.0

Conversation

@dmmdea

@dmmdea dmmdea commented Nov 11, 2026

Copy link
Copy Markdown
Owner

One change, merged to main through its own pull request (#852) after one
security pass and green CI:

  • A text lease on a node's GPUs no longer refuses an accel job: an NPU
    or TPU is not a card a lease names.
  • The owner of a shared accelerator tool is tried first, and when it
    cannot take the call the next device that serves the name answers
    through its own tool row; the result carries
    fallback{from,reason,served_by}. An embedding moves to another device
    only with any_space:true.
  • out_path is checked before a fallback runs, and the Coral sidecar
    guards its own writes (the two findings of the security pass).
  • ADR 0102, which supersedes one consequence of ADR 0068.

Deletes: changes/unreleased/fix-accel-independent-and-fallback.md

🤖 Generated with Claude Code

One change, merged to main through its own pull request (#852) after one
security pass and green CI:

- A text lease on a node's GPUs no longer refuses an accel job: an NPU
  or TPU is not a card a lease names.
- The owner of a shared accelerator tool is tried first, and when it
  cannot take the call the next device that serves the name answers
  through its own tool row; the result carries
  fallback{from,reason,served_by}. An embedding moves to another device
  only with any_space:true.
- out_path is checked before a fallback runs, and the Coral sidecar
  guards its own writes (the two findings of the security pass).
- ADR 0102, which supersedes one consequence of ADR 0068.

Deletes: changes/unreleased/fix-accel-independent-and-fallback.md
@dmmdea
dmmdea merged commit 1bfbd8a into main Nov 11, 2026
6 checks passed
@dmmdea
dmmdea deleted the release/0.231.0 branch November 11, 2026 19:58
@dmmdea

dmmdea commented Nov 11, 2026

Copy link
Copy Markdown
Owner Author

Live run of 0.231.0 (the run the pull request body promised)

A fresh process of the installed 0.231.0 binary on the delegator, driven over MCP stdio. The node that carries the NPU card is "node E", the Coral's is "node L", the RK3588 is "node O".

1. A GPU text lease held on node E (/fleet/health lease.class: text, read before and after the calls):

Call Result
offload_object_detect ran on the NPU card: 2 cats, 2 remotes, a couch; 56.9 ms on-card
offload_depth ran on the NPU card: PNG landed beside the image; 558 ms wall
control: an image-gen dispatch to node E 503 node leased (gpu lease class=text ...)

Repeated under gpu reserve --exclusive (lease_exclusive: true before and after): detect and classify ran on the NPU card. Before this release both calls got the control's 503.

2. Node E unreachable (a scratch config without it in delegate_remotes; each answer's probe list shows it was not asked):

Call Served by Result Wall
offload_object_detect Coral, node L 3 cat boxes 1,607 ms
offload_classify_image Coral tabby 0.375 518 ms
offload_semantic_segment with out_path Coral person 29,239 px; a 640x425 PNG at out_path, opened and correct 529 ms
offload_image_embed nobody deferred, fallback_withheld
offload_image_embed any_space:true Coral 1280-d, space: efficientnet-edgetpu-s 529 ms
offload_semantic_segment, out_path = the input image nobody deferred, fallback_withheld; the image unchanged
offload_depth (no second device) nobody plain defer
offload_object_detect, node L dropped too RK3588 NPU, node O 2 cats, 2 remotes, a couch, a bed 2,112 ms

Every fallback answer carries fallback{from: axera-ax650, reason, served_by}. The agent loop's tools were called for real through pipeline.NewLoopAccel with the same two configs: the Coral served detect, the embedding was withheld, and with the real config the NPU card answered as before.

Operator note. The fallback order is the listing order of fleet_accelerators. The run showed why it matters: the Coral missed both remotes and the RK3588 found them, so the delegator's list was reordered to put rknpu first (a config change, no code). A per-tool fallback order would be the complete form of this and is not in this release.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant