perf(gfx1100): add opt-in MQ4V2 K5120 gate/up decode specialization

[HUSRCF]

perf(gfx1100): add opt-in MQ4V2 K5120 gate/up decode specialization

Summary

Add an opt-in specialization of the existing MQ4V2 gate/up decode kernel for exact gfx1100 and (gate_m, up_m, K) = (17408, 17408, 5120). It fixes the group count at compile time while preserving the existing weight representation and arithmetic. Other architectures and shapes retain their existing routes; there is no additional resident weight copy.

The setting is kernel.mq4v2_gateup_k5120, with compatibility override HIPFIRE_MQ4V2_GATEUP_K5120=0/1. It defaults off and is resolved through ProcessConfig/FeatureFlags at initialization. The new symbol is registered for kernel packaging and replay; compiler-less installations need an updated pack to opt in.

Scope

• Tested base: beta 45f1abd1d. Current integration includes beta a193e4491 through merge commit 3123f3865; original commits/history are preserved.
• Commits: cb2f84ae3, 671e88e02, f689f559c.
• Surfaces: kernel dispatch/registration/replay, configuration, documentation.
• No Architecture trait, attention, quantization, or speculative-decoding changes.
• Independent of the split-KV attention verifier work.

Measured Result

Primary final check: W7900 Dual Slot GPU1 / gfx1100 / HIP 7.15, same pinned model, Q8/VMM, AR, explicit max_think_tokens=1, TG4096. ABBA off/on/on/off = 38.9 / 39.6 / 39.6 / 38.8 tok/s. Median 38.85 -> 39.60 (+1.93%). All four requests generated 4096 visible-output tokens with no reasoning or cached prompt tokens; request and decoded-text hashes match. This is separate from the earlier GPU0 reasoning run below; do not pool the two devices or modes.

Reproducer and summary: benchmarks/scripts/mq4v2_k5120_abba.sh and benchmarks/results/mq4v2-k5120-20261005/README.md.

Raw evidence: mq4v2-k5120-evidence-20261005.tar.gz (337179 bytes), SHA256 38e70e8d1bbf1540920482b77f5cec432ea248f3c8804623f37974ce19e954f8.

Latest-beta integration boundary

After the full validation below, beta advanced to a193e4491. I merged it without rewriting the measured commits. The latest conflicts were confined to generated crate maps and the environment inventory count, regenerated with the official scripts; no K5120 kernel or dispatch conflict required manual resolution. On the merged tree, cargo test --locked -p rdna-compute --lib k5120 passes all four tests (config lifecycle, exact arch/shape guard, inventory and replay contract), cargo test --locked -p hipfire-config --lib passes 107 tests, and lifecycle, all 48 crate maps and whitespace checks pass. Full CI and GPU ABBA results below belong to the pre-merge candidate, not a newly benchmarked latest-beta binary. This PR remains a draft pending full integration validation and remote CI.

Earlier GPU0 Reasoning Run

W7900 gfx1100, physical GPU0, HIP 7.15; Qwen3.8-27B MQ4-XT; Q8 KV/VMM; speculative decoding off. Each fresh process had a 10-second DPM warmup and a distinct 128-token warmup request, followed by the same TG4096 request. Arms ran A/B/B/A with 15-second idle intervals. Rates are daemon-reported decode throughput, excluding TTFT/load.

Arm	Setting	Decode tok/s
A1	off	39.0
B1	on	39.6
B2	on	39.6
A2	off	38.9

Median: 38.95 -> 39.60 tok/s, +1.67%. This is a small workload-specific observation from two samples per condition, not a universal speedup claim. All requests generated 4096 tokens with zero cached prompt tokens; decoded reasoning/content hashes matched. Token-ID equality is not inferred from text hashes.

The model contract ignored the named thinking-off budget: these measured requests produced reasoning, not final answers. This benchmark is separate from the successful functional serving battery. No per-arm profiler symbol trace was collected during timing; dispatch/replay registration has separate validation.

Artifact SHA256: 80e7c624424fd1d363ba86681d3dc1e5ac5534e0e064306a32be204c4843d0f3.
CLI MD5: 894739f626fe3bf5cd4e077bf7743a6d.
Daemon MD5: 529a278af991815057690ff384cd5846.

Verification Status

• Product CLI/daemon release build.
• Configuration tests: 107 passed.
• K5120 config roundtrip and arch/shape guard tests: 2 passed.
• Registry/replay and source-fingerprint checks from rebase validation.
• Official serve battery: off 5/5, on 5/5; matching requests, token counts and decoded responses.
• Official Redline harness: four consecutive positions with bit-exact HIP/blob/PM4 logits, KV and recurrent/GDN state; 659 launches, 16 families.
• Lifecycle documentation checker and whitespace checks.
• Independent narrow code review (gpt-6-sol high).
• Complete official no-GPU CI with PYTHONSAFEPATH=1 PYTHONPATH="$PWD"; workspace lib tests: 4336 passed, 67 ignored. The initial Python import-shadowing failure and successful rerun are both retained.
• Refreshed crate maps, scratch-growth and ratchet-diff checks pass.
• All structural CI green: daemon_lines=5392 exceeds 5377, and two historical changelog sections differ. Both failures reproduce unchanged on clean beta 45f1abd1d; no policy/threshold edits included. Remote CI and maintainer review remain required.
• Locked global speed-gate: not claimed by this targeted ABBA experiment.
• Commit portable reproducer, original prompt and summary; prepare raw-evidence archive with file hashes.
• Publish the evidence archive and link it in the PR.

Unrelated local research and unsuccessful optimization experiments are excluded from this submission.

[HUSRCF]

Updated against beta a193e44 in 3123f38, preserving the existing history. The conflicts were generated crate maps and the environment inventory count; regenerated them with the official scripts. GitHub now reports mergeable=true. Rechecked: 4 K5120 tests, 107 config tests, lifecycle inventory, all 48 crate maps and whitespace checks pass. No K5120 arithmetic/dispatch changes were needed. The attached GPU ABBA evidence remains explicitly tied to the previously tested commit; this remains draft pending full latest-beta integration validation.