[mrsergiopascal]

Summary

Keep decode-sized replay position pointers aligned without changing their values. An odd replayCut offset currently leaves POS at 8 modulo 16; split sparse attention aliases its default RBASE/CBASE to POS, and the served aligned AOT variants correctly reject that launch with MissingTritonVariant.

Only when the cut view is unaligned and the surviving chunk has n <= tri.decode_rows, upload the exact surviving host_pos slice to the existing aligned ch.pos0, through the existing stream-ordered upload. Even cuts and larger chunks keep their original view. No kernel math, replay geometry, AOT captures, alignment matcher, deployment scripts or runtime settings change.

Target and duplicate check

Based on bertholomus/TensorFold:deepseek-v41-zig-tp4 at b65c8b92d2403dfdefccf15a3b23ba609a4553e2, the active TP4 engine branch. The affected family is absent from this fork's main and current ashhart main. Checked current source, fork issues/PRs and upstream replay/alignment reports: no equivalent repair found. The prompt-indexer VIS alignment repair (#4/#8) and generic safe AOT matching changes are different callsites/problems.

Regression and reproduction

• python3 tools/dsv41-zig/test_replay_positions.py (Zig 0.17.0): 3/3 passed. The actual extracted replayCut is compiled, with GPU collaborators stubbed; exact positions, noncontiguous sentinels, host immutability, odd/even/large cuts and early returns are checked.
• Running that same test against the unmodified base function: 2 failed, 1 passed, with expected 0, found 8 on position alignment.
• Added opt-in, standard-library synthetic Responses reproducer and instructions in tools/dsv41-zig/replay_positions.md. It sizes with /tokenize, preserves identical first-request bytes for before/after, checks SSE completion, exact tool arguments and actual completed continuation. Its generated fixture objects were verified equal to the four previously GPU-qualified synthetic fixtures. The public helper's CLI/syntax were checked; no new GPU run was performed while preparing this PR.
• zig fmt --check on the changed Zig file and git diff --check: passed.
• python3 tools/zig/lean_check.py: 1615 inherited findings before and after; zero added (nonzero exit, not a clean full-tree claim).
• zig build test test-golden -Dcpu=apple_m1: blocked on this macOS host, by missing Xcode Metal compiler and the pre-existing Linux-only c.SO.BUSY_POLL reference in dsv41/link.zig. No unrelated build/policy/CI edits made.
• Independent read-only CLI code review completed before publication.

Earlier real TP4 before/after qualification

These results are from the recipe engine base 59402086381a42c74d53053443224eec2790b4c5 plus the four runs/2026-10-17-1017-tensorfold patches in tonyd2wild/DeepSeek-V4.1-Flash-vLLM-DGX-Spark, with/without this narrow fix. They do not claim GPU qualification of the newer PR base. Tested on four GB10 Sparks, DeepSeek-V4.1-Flash 1017, TP4, 2048-row chunks and 128-row replay window.

Rendered input tokens	Baseline	Identical first request after fix	Candidate cached tokens
10353	MissingTritonVariant, zero output tokens	completed + tool continuation	0
10363	MissingTritonVariant, zero output tokens	completed + tool continuation	0
39039	MissingTritonVariant, zero output tokens	completed + tool continuation	36864
123003	MissingTritonVariant, zero output tokens	completed + tool continuation	120832

Neighbors 39038/39040 and 123002/123004 and prefix warmers also passed with exact tool arguments and ALIGNMENT_OK continuations. The failing launch has ring_size=144, n_idx=512 and POS/RBASE/CBASE all 8 modulo 16. Existing H16/H32 AOT records were inspected for the aligned requirement. Replay offsets are source-derived; prompt lengths, prefix resumption and completions were observed.

Only synthetic fixtures and four-Spark TP4 were tested. No TP2/other hardware qualification, benchmark, soak, private-workload replay, broad long-job stability or whole-model token-exactness claim. No raw prompts, private logs, infrastructure identifiers or deployment material are included.

[mrsergiopascal]

Recipe heads-up for @tonyd2wild: this narrow engine PR addresses decode-sized long-context replay alignment in the runs/2026-10-17-1017-tensorfold recipe. Earlier four-Spark TP4 tests reproduced MissingTritonVariant at 10353/10363/39039/123003 rendered tokens and passed the identical requests after this repair, including tool continuations and cached-prefix cases. The PR includes synthetic reproduction instructions and explicitly distinguishes that pinned-build evidence from this newer branch. Linking here rather than opening a duplicate recipe issue; no recipe/runtime change is requested until engine review and qualification.

[bertholomus]

Merged, thank you @mrsergiopascal! A replay cut that leaves 1 to 15 rows at an odd offset now copies those rows' positions into the chunk's aligned buffer, and every other cut keeps the old view, so replies that worked before stay the same. Here the host tests, the server tests and an aarch64 + CUDA build pass, and your CPU test passes on the fix and fails on the old code, as you wrote. Thanks also for the clear notes and the reproducer!