[dlangbe]

JIRA ID : AICK-1895

Motivation

lds_k_group_pad was added to AttentionDenseSpec in PR #10102 with a fixed
default of 8. That PR committed to a follow-up sweep over the full pad space to
confirm the default and record whether any shape benefits from a different value.
This PR closes that follow-up by documenting the sweep methodology, results, and
decision in-repo, and making the knob visible in the runbook's knob catalog.

Technical Details

Three documentation edits:

library/builders/gfx950/attention/prefill/README.md

• Fixed stale line: "block_n and waves_per_eu are the only performance knobs"
  — there are now three (block_n, waves_per_eu, lds_k_group_pad).
• Added ## Tuning — lds_k_group_pad section documenting:
  • Sweep axes: pad ∈ {0, 8, 16, 24, 32} × bn ∈ {64, 128} × 3 GQA ratios ×
    causal/swa/varlen/persistent × S ∈ {2048, 4096, 8192} (840 configs,
    all correctness-passed).
  • The whole-wave bank-conflict model explaining why pad=8, 16, and 24 tie at
    the 4-lanes-per-bank floor, and what it would take to separate them.
  • Decision: default stays at 8. pad=24's plurality is sub-1% and noise-grade
    without an ABBA repeat; pad=16 has an unacceptable regression on a bn=128
    sliding-window shape that makes it unsafe as a global or shape-conditioned
    default.
  • The pad=16 regression recorded as a named constraint for future work.
  • The multiple-of-8 invariant (smem_load_vN stamps align 16
    unconditionally; __post_init__ enforces it).

platform/dsl_docs/optimization/optimization_runbook.md §12.1.F

• Added lds_k_group_pad knob-catalog row, sibling to the existing lds_k_pad
  and lds_layout rows. Points to the gfx950 README for the full decision
  record. Makes the knob visible to the Step 0 exhaustive lever sweep.

library/builders/gfx942/attention/prefill/README.md

• Added one row to the "Lever record → Adopted" table noting that gfx942 reuses
  the shared AttentionDenseSpec.lds_k_group_pad field and inherits default 8,
  with a cross-reference to the gfx950 sweep.

Per platform/AGENTS.md §Compliance, absolute TFLOPS figures and per-lever
deltas are kept on the protected results page only; the in-repo docs contain
methodology, axes, decision, and qualitative direction.

Test Plan

• Documentation-only change; no kernel IR or dispatch logic modified.
• Verified the stale knob-count sentence (L26-28 of the gfx950 README) is
  corrected to reflect three knobs.
• Confirmed the new runbook row is correctly placed in §12.1.F alongside sibling
  LDS-layout knobs.
• Confirmed the gfx942 lever table row accurately describes the inheritance
  relationship and points to the canonical decision record.

Test Result

No functional changes. All existing parity and byte-identity gates are unaffected.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[therock-pr-bot]

✅ All Checks Passed — Ready for Review

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	PR does not contain code files — Unit Test auto-passed
🔎 pre-commit	✅ Pass	—
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—
🤖 therock-pr-bot	✅ Pass	—

🎉 All checks passed! This PR is ready for review.

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

🙋 Wish to Override Policy?

[therock-pr-bot]

🎉 All checks passed! This PR is ready for review.