warpfront/hipfire

Tracks the two gates blocking Lloyd-MQ3 (qt=20) from replacing uniform MQ3 as the 3-bit default. PR #115 has the implementation; this issue is the punch list to clear it. Empirical state (from PR #115) 9B Qwen3.5 Lloyd-MQ3 wikitext2-test ppl = 18.52 vs uniform MQ3 = 42.03 — 2.27× quality reduction for +7.7% bandwidth (112 B/group vs 104). Closest hipfire has gotten any sub-4-bit to MQ4 (10.34 → 1.79× gap, vs uniform MQ3's 4.07×). Same lift across all sizes: 0.8B: 301 → 155 (1.94×) 4B: 45 → 22.6 (2.01×) 9B: 42 → 18.5 (2.27×) Decode took a 3.2× hit on 9B (141 → 44 tok/s) from the 8-way switch in gemv_mq3g256_lloyd.hip — the codebook lookup compiles to a longer dependency chain than uniform's scale*q + zero. Gate 1 — K4-unroll kernel (~1 day kernel work) Goal: 9B Lloyd-MQ3 decode ≥120 tok/s on gfx1100. Approach: mirror gemv_hfq3g256.gfx1100.hip's 4-accumulator K4-unroll pattern. Two ideas, either or both: LDS-resident codebook table — load 8 fp16 (16 B/wa

Reporter: @mikiadev in #79 — "in 7-10 attempts the last part of thinking 'I'm going to answer the user ' got infinitely looped" Reproduced: 7900 XTX, qwen3.6:35b-a3b.mq4 + DFlash draft, prompt "What is the answer to life, the universe, and everything?", DFlash greedy (temp=0), max_tokens=600. Hit rate 2/5 runs hit a clear block-level attractor in the last 1/3 of `` (Tier-2 / Tier-3 in `coherence-gate-dflash.sh`). Matches @mikiadev's ~2/10 from real pi traffic. Example signature from one stuck run: ``` 6× "context—which is why many people joke about how meaningless the answer actually was without" ``` The phrase rotates as a closed loop; draft and target keep agreeing on it because it's "consistent with itself" — a textbook DFlash sidecar-drift attractor (cf. CASK m-fold + DFlash 2026-04-26 in `feedback_dflash_coherence_gate_mandatory.md`). Why A3B specifically Per `feedback_a3b_dflash_regression.md` the global TriAttention sidecar for A3B-3.5 caps at R̄≈0.36 and A3B-3.6 at R

Status: PRD scoped at docs/plans/sliding-window-fa.prd; branch feat/sliding-window-fa (commit 6c59950); implementation queued. Background Lucebox-Org/lucebox-hub merged PR #26 on 2026-04-26 — "feat(dflash): sliding window FA + two-phase cache". On RTX 3090 with Qwen3.6-27B at 60K context, decode jumps from 25 tok/s (full attention) to 89.7 tok/s (window=2048), 3.48× with 100% speculative acceptance at all window sizes. The 100% acceptance claim is structural, not probabilistic: when target and draft both apply the same windowed mask, their argmax distributions agree by construction. Why this matters Empirically validated on the same target/draft class hipfire ships (Qwen3.6-27B + DFlash MQ4 draft). Cheaper than CASK / TriAttention at runtime: a single win_start shift in the FA mask, no separate scoring kernel pass per cycle. Compounds with DFlash spec-decode where verify/replay FA is the long-context hot path. Hipfire has no equivalent today Mechanism hipfire status Sliding

Status: open invitation, hardware-gated. Needs Strix Halo (Ryzen AI Max+ 395 / Radeon 8060S / any other gfx1151 SKU). Local maintainer doesn't have one. Context PR #59 (@KotDath) just landed the autodetect fix so hipfire diag and scripts/install.sh correctly recognize gfx1151 KFD topology (gfx_target_version 110501 → gfx1151). The arch now resolves to its existing archDefaults profile (asym2 KV, 16 GB shared memory budget). What's NOT shipped yet for gfx1151: tests/speed-baselines/gfx1151.txt — the speed-gate refuses to run on archs without a baseline file. Anyone with the hardware can bootstrap it in one command, see below. Documented bench numbers in docs/BENCHMARKS.md "Other arches" section — currently lists gfx1030 / gfx1010 / gfx1013 / gfx94x but not gfx1151. Coherence-gate confirmation across the model matrix — KotDath confirmed 0.8b + 9b + 9b-draft + DFlash work on Strix Halo, but a full ./scripts/coherence-gate.sh run would close the loop. Phase 1 — speed-baseline (