Skip to content

release: 0.70.2 #4890

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
noahgift wants to merge 5 commits into from
Closed

release: 0.70.2 #4890

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
be512e8
chore(release): bump workspace to 0.70.2
noahgift Oct 7, 2026
b11fd23
release(0.70.2): crux-smoke emergency scope, carried certification, n…
noahgift Oct 7, 2026
f6d343a
fix(crux): serve cell runs the hf/vllm comparators first, alone on th…
noahgift Oct 7, 2026
70c9b9c
fix(blis): the 8x6 asm kernel read up to 128 bytes past its A panel o…
Oct 7, 2026
89e261c
release(0.70.2): notes name the CPU train/finetune crash fix (#4891)
noahgift Oct 7, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
27 changes: 27 additions & 0 deletions CHANGELOG.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,6 +29,33 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
host still owes every Qwen3 cell. Planned for 0.70.1: native sm_121 SASS, then a re-test of batched
prefill on GB10.

## [0.70.2] - 2026-10-07

0.70.2 ships under a recorded operator emergency scope, `crux-smoke`, as 0.70.1 did: "0.70.2 ships on
CRUX smoke on lambda and gx10 GPU. Everything bigger is nightly." The record is the 0.70.2 entry under
`emergency_scopes` in `contracts/model-capability-ladder-v1.yaml`. The prompt certification is carried
from 0.70.1 unchanged (same models, prompts and oracle engines). The full notes, with the reason for
carrying it, are in `evidence/release/0.70.2/RELEASE-NOTES-0.70.2.md`.

### Fixed
- `apr serve`: tools in a chat request reach the chat template, and Qwen3.5 XML tool calls are parsed (#4650).
- `apr serve`: KV caches no longer commit every byte when they are built (#4769).
- `apr train` / `apr finetune` on CPU: an intermittent crash (SIGSEGV) in the AVX2 matrix-multiply kernel
is fixed (#4891).

### Release
- The PP-26 witness marker is re-measured on this release's base, so its release-phase check judges fresh
evidence (#4888).
- The coverage gate reads the nightly's sha-keyed receipt and is checked before the tag (#4734); merge, tag
and publish requirements are one list in code (#4688); no public GitHub release before its assets,
clean-room and preflight (#4690).

### Known issues (moved to 0.71.0)
- Six models still fail and ship as known failures, each tracked on its ticket: #4661 and #4662 (cuda serve
/api/chat gibberish), #4663 (no measured thinking budget for an IQ4_XS quant), #4664 (an IQ2_XXS GPU forward
falls back to CPU), #4665 (no CUDA forward for qwen35moe), #4666 (think block unclosed within 2048 tokens).
- One unattended autopilot pass: the autopilot's readiness step and tag cut do not read a recorded scope.

## [0.69.3] - 2026-09-24

0.69.3 is an emergency early release, authorized by the operator: "we need near parity apr serve
Expand Down
Loading
Loading