[deriknel]

Problem and result

With --pipeline-windows 2 on SYCL, greedy text forks from the serial loop's within the first few hundred characters. A pipelined verifier (set_always_publish) captures its one-token window with the in-graph state advance, and pl_commit_async then runs the commit graph on top of it, so every one-token window advances the first stage's recurrent (GDN) state twice. c81f998 fixed this in the CUDA verifier on 2026-10-09; the SYCL verifier still has the old condition. With this change the pipelined text matches the serial loop's.

Related discussion

Needs #2433, without which --pipeline-windows does not run on SYCL at all. Ports c81f998. No scope agreement needed: a focused fix.

Approach

sycl/src/core/verify.cpp, the two conditions c81f998 changed in src/core/verify.cpp:

• Verifier::record_window: self_commit is off when always_publish_ is set.
• Verifier::commit: the branch that skips the commit graph for a one-token window is not taken when always_publish_ is set.

Only --pipeline-windows calls set_always_publish(true), so the serial loop does not change. The branch still merges cleanly into main.

Validation

Setup for every run: main e88ee39 built for SYCL (AOT bmg-g31, oneAPI 2026.1, Compute Runtime 26.35.39758.10, IGC 2.41.5) in a container; host CachyOS, kernel 7.2.11, xe. 2x Intel Arc Pro B70 32 GB, layer split 26, Swift 1.5 Flash-Next IQ3_S with every expert in VRAM, --max-context 262144, int8 KV, --spec 4 --spec-min-p 0.6, --prefill 4096. Each build has only this PR's commit on top of main. Greedy check: 5 fixed prompts (prose, code, maths, an edit, a story), temperature 0, 400 tokens, thinking off, compared character by character. The serial reference is main with pipeline windows off.

Passed:

• main + [Engine] sycl: a pipelined window still on the GPU is polled again, not failed #2433 + this PR, --pipeline-windows 2: all 5 greedy texts identical to the serial loop's, in two rounds.
• Same build: decode 88.8 / 126.4 tok/s (prose / code) pipelined, against 83.0 / 106.6 serial on main (+7% / +19%).
• Same build: a ~148K-token prompt (answer found) and 10 prompts of ~13K tokens (all found), no verify: error, no lost device, no xe reset.
• ctest on both cards: the same 3 failures as on main; iq_parity with fixtures, 0 failures.

Failed:

• main + [Engine] sycl: a pipelined window still on the GPU is polled again, not failed #2433 only, --pipeline-windows 2 (the bug): pipelined text differs from the serial loop's at character 72 (prose), 146 (code) and 50 (story); maths and the edit are the same.

Not run:

• CUDA, HIP and Metal builds: the change is in sycl/ only.

Limits

None known. On 0.1.43.1 we also confirmed the cause with STRATA_ONE_TOKEN_COMMIT=0 (identical text and state hashes), not repeated on 0.2.0.

We investigated (myself and Claude Opus 5.5 in Claude Code) and I reviewed the findings and ran every test on my two Arc Pro B70s on my CachyOS machine.

[Niko1221]

Thanks for the pull request.

We're resetting the tracker for 0.2.0-beta. There are more than 1,300 open issues and pull requests, too many for us to review properly, so we're closing everything that isn't being worked on right now. Closed here does not mean rejected.

To reopen it takes about two minutes:

1. Rebase on main if it no longer applies, and re-run your checks on 0.2.0-beta.
2. Start the title with its area: [Engine], [GUI], [TUI], [Benchmark], [Docs] or [Other].
3. Fill in the pull request template: the problem and result, and the validation (what you ran, on which revision, what you saw).
4. Comment /reopen. A check runs and either reopens the pull request or lists what is still missing.

The rules are in CONTRIBUTING.md.

[deriknel]

/reopen

[github-actions]

Reopened, thank you. A maintainer will pick it up from here. If it conflicts with main, rebase it first; that speeds up review.