[AbrahamCete]

What this changes

hip_records.zig has records for the planned walk's tf_sum_partials and tf_sum_partials_act, which came in with #1961. They held the system compiler's register counts. They now hold the counts from ROCm 10.1's hipcc (HIP 7.16.26385), the compiler the HIP records follow, taken from its hip-kernel-contract-limits RECORD lines:

• tf_sum_partials: 21 VGPRs and 20 SGPRs (22 on gfx1201). It held 23 VGPRs and 20 or 21 SGPRs.
• tf_sum_partials_act: 22 VGPRs and 22 SGPRs (24 on gfx1201). It held 24 VGPRs and 21 or 22 SGPRs.

Only these 8 usage lines change. No kernel or launch changes.

• Issue: n/a. There is no issue: @ashhart asked for this on perf(hip): a skewed-row matrix tile, three blocks a WGP on gfx1151 for prompts up to 8192 codes deep #2087 as a follow-up to perf(hip): the decode softmax stats and partial sums load sixteen ahead #1961.
• Stacked on: main.
• atc overlap --json (2026-11-05 06:18Z, on main 249e33b): this head shares hip_records.zig with hip flashnext: WMMA sparse attention for prompt chunks (sparse=wmma) #1622, hip flashnext: sparse attention with a lane a key over several heads, the scalar kernel's bits #1877, Nemotron Parse on HIP: NVIDIA's checkpoint, token-exact against its HF example #1941, LightOnOCR-3 (0.8B, 1B, 4B) on native HIP, with the model card's oracle #1965 and Clef-Flash: joint-schema decisions on /v1/systemone (HIP) #2048. Clef-Flash: joint-schema decisions on /v1/systemone (HIP) #2048 merges clean. The other four already conflict with main in that file without this change, and none of them touches the tf_sum_partials lines, so their conflicts are inherited.

Checks on this head

• Head commit: 0d96eac, on main 249e33b. Both sides were built from git archive trees.
• Lean checker, python3 tools/zig/lean_check.py, on the base and on this head: 425 and 425, 0 new (tools/zig/lean_new.py).
• Tests run, with the command and pass, fail and skip counts:
  • zig build test -Dhip=true --summary all: 575/575 steps, 7352/7414 tests passed, 62 skipped, 0 failed.
  • zig build hip-kernel-contract-limits -Dhipcc=<ROCm 10.1>/bin/hipcc: FAIL kernel-contract-limits: 14 problems, against main's 22.
    • The 8 lines of these two kernels now read record current on gfx1030, gfx1100, gfx1151 and gfx1201.
    • The 14 left belong to other kernels' records: GLM-OCR's text and vision RoPE, BF16 axial vision RoPE, the MoE combine gate, the 64-row item tile and Torch BF16 RoPE. This PR leaves them alone.
  • zig build hip-kernel-contract on gfx1151, built with that hipcc: SUMMARY gfx1151 93 kernels: 93 pass gates 1-4, 0 fail, 2 known findings; 11 controls: 11 red as planted, 0 missed; 4 limit problems, exit 1 for the limit problems. Main reads the same with 6 limit problems.
  • zig build test-golden --summary all: 5/5 steps.
  • zig fmt --check and git diff --check: clean.
• CI: green on 0d96eac, all 7 checks (zig, zig-windows, zig-macos-dev-tools, fmt, release-paths, and benchmark-clients on Ubuntu and Windows).
• Not run, and why: the GPU contract on gfx1030, gfx1100 and gfx1201, because I have no card. The offline limits above check their records.

Receipt

n/a. The change edits contract records only. It changes no kernel, launch or host code, so it cannot reach inference.

Checklist

• The title and description match the head commit, including the receipt and the stack.
• New tests fail before the change, pass after it, and skip cleanly without their dependency. (n/a: no test is added; hip-kernel-contract-limits is the check, 8 stale records before and 0 after.)
• Comments and docstrings are one line. No measurements or history in the source.
• No personal data, machine names, internal hosts or local paths in the diff or its commits. No AI attribution lines.

If it's useful, the other 14 records can follow the same way.

[ashhart]

Thanks @AbrahamCete for following up on #1961's records so quickly. Taking them from the release hipcc's own RECORD lines, and showing the 8 lines go current, is exactly the check. Ready to land, and it's in the queue.

[ashhart]

Thank you @AbrahamCete. This is merged into main with your commits under your own name, so you're listed as a TensorFold contributor. It ships in the next release, credited to you in the changelog and the release notes.

[ashhart]

Thanks @AbrahamCete! This is in TensorFold 1.2.3, out now: https://github.com/ashhart/TensorFold/releases/tag/v1.2.3