[bartekxk]

D256 bf16 decode dispatcher candidate enablement.

Motivation

Technical Details

• register D256 bf16 decode dispatcher candidate
• tune num_segments

Test Plan

run_test.py, wire test

Test Result

Passed.

Submission Checklist

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

JIRA ID: AICK-1496

[therock-pr-bot]

✅ All Checks Passed — Ready for Review

Check	Status	Details
🌿 Branch Name	✅ Pass	—
📝 PR Title/Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	—
🔎 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.

[therock-pr-bot]

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

[Copilot]

Pull request overview

This PR enables a new specialized D256 bf16 decode attention dispatcher candidate (targeting gfx950/gfx942) and adjusts unified-attention tuning to improve/validate decode performance, along with accompanying wiring tests and updated benchmark/shape inputs.

Changes:

• Register a new d256_decode attention candidate in the dispatcher and add a shared cohort predicate for gating eligible decode problems.
• Tune _num_segments for gfx942 decode so D256 can use the formula-derived segment count (avoiding an overly-low cap).
• Expand/update parity scenarios and benchmark shape sets/results to include D256 bf16 decode coverage.

Reviewed changes

Copilot reviewed 7 out of 8 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
dnn-providers/hip-kernel-provider/rocke/library/tests/dispatch/attention/test_d256_decode_wiring.py	Adds unit tests verifying candidate registration, gating, and routing to the new d256_decode path.
dnn-providers/hip-kernel-provider/rocke/library/kernels/common/attention_unified.py	Tunes _num_segments for gfx942 decode and introduces _d256_decode_cohort predicate.
dnn-providers/hip-kernel-provider/rocke/library/dispatch/attention.py	Registers the new attention_d256_decode candidate and wires selection/spec naming.
dnn-providers/hip-kernel-provider/rocke/library/builders/gfx950/attention/prefill/parity_unified_attention.py	Expands gfx950 parity scenarios to include more D256 bf16 decode seq_len pairs.
dnn-providers/hip-kernel-provider/rocke/library/builders/gfx942/attention/prefill/parity_unified_attention.py	Reworks gfx942 parity scenario selection to inline scenario generation and adds bf16 D256 decode shapes.
dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx950/attention/decode/decode_shapes.json	Adds Qwen3-Next-80B-A3B bf16 D256 decode benchmark shapes.
dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx942/attention/decode/decode_shapes.json	Adds Qwen3-Next-80B-A3B bf16 D256 decode benchmark shapes.
dnn-providers/hip-kernel-provider/rocke/library/benchmarks/gfx942/attention/decode/decode_shapes_perf.csv	Updates recorded perf sweep results (including new D256 bf16 entries).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[AnaghaRaoAMD]

Please run parity_unified_attention.py --scenario on the bf16 D256 decode shapes on a real gfx942/gfx950 GPU and post the max_abs numbers — ideally with a larger input magnitude than uniform(-0.1, 0.1) so the absolute tolerance doesn't mask a relative error

[AnaghaRaoAMD]

follow-up: when PR#9496 lands the problem-driven prove_parity, the D256 decode cohort and the num_segments D256 branch should be added to that matrix — and if/when the C++ attention engine goes live, the d256_decode routing + segment logic will need a C++ mirror to stay in lockstep.

[AnaghaRaoAMD]

some nit comments. Overall, LGTM.