[bong-water-water-bong]

gpu-kernels 4f586cac adds, since cf222e7d:

• Docs: the onebit recipe lives in our Lemonade fork #24: hip.fa_decode_d256.gate, one dispatch for the gated full attention of Qwen3.5 / Qwen3.6 / Qwen3.8
  (FLASH_ATTN_EXT -> RESHAPE -> MUL by SIGMOID(gate)) at 1..8 query tokens (decode and MTP verify), head
  256, f16 KV, n_kv >= 1024. The Loom gated kernel ran it on one workgroup per query head (16-24
  workgroups); profile at 26,644 tokens: 55% of a Qwen3.5-4B decode step and 34% of a Qwen3.8-27B one,
  about 41 GB/s of KV. The HIP kernel splits the cache over T x Hkv x S workgroups, reads each K/V row
  once per KV head and merges with agent-scope fences and runtime-zeroed counters (about 217 GB/s at
  n_kv 32768 in its bench).

Measured on strixhalo gfx1151, performance mode, private build (llama.cpp af04c9f2 + hrx-system 82804197 +
the add-on), llama-bench -fa 1 -p 0 -n 64 -d D, interleaved x3, medians, tok/s (ours):
d=0 4096 16384 32768
Qwen3.5-4B before 61.13 50.93 34.17 23.71
Qwen3.5-4B after 61.20 59.19 53.31 47.07
Qwen3.8-27B before 12.34 11.36 9.27 7.49
Qwen3.8-27B after 12.32 12.11 11.62 10.97
(Lemonade's Vulkan llama.cpp b11393, same command and files: 4B 64.03 / 60.99 / 54.80 / 48.27,
27B 12.47 / 9.06 / 11.78 / 11.14; 27B d<32768 rows on the 08bb2241 base.)
Qwen3.8-27B --mtp, llama-server, 26,644-token prompt, 64 greedy tokens: 16.40 -> 23.78 tok/s (medians of 4).
Correctness: kernel bench vs a double-precision reference on 17 shapes (G 1..8, T 1..8, gate on/off) PASS,
5 repeats bit-identical; greedy text vs the Loom kernel identical on Qwen3.5-4B at 180 / 6,662 / 26,644
tokens (mean top-10 KL <= 1.7e-4), Qwen3.8-27B 26,644 tokens mean KL 3.7e-5 then a near-tie flip at token
10 (logprob gap 0.008); identical requests repeat exactly; no GPU fault in dmesg.

Needs the hw-gate run (label) before merge: goldens with Qwen3.8-27B decode past 1024 KV cells may drift on
near-ties.

🤖 Generated with Claude Code

[github-actions]

PR Reviewer Guide 🔍

(Review updated until commit 6b898ff)

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 24 - Partially compliant Compliant requirements: Update docs/lemonade.md to reflect that the Lemonade recipe lives in the 1bit-MONSTER/lemonade fork Update the README status to reflect the change Update PORTING step 1 to reflect the change Non-compliant requirements: None Requires further human verification: None

⏱️ Estimated effort to review: 2 🔵🔵⚪⚪⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Outdated Reference The documentation still refers to the old commit cf222e7d for the private HIP kernels, while the actual pinned commit in config/gpu-kernels.config is 4f586caca4147b072b5056756a24852cd4a08b1b. This discrepancy could mislead users into thinking the wrong version is being used. | private HIP kernels | [1bit-MONSTER/gpu-kernels](https://github.com/1bit-MONSTER/gpu-kernels) (not a submodule: a build points at a checkout with `-DONEBIT_GPU_PRIVATE`) | `4f586cac` — the commit the release builds and the hardware gate were run at |

[github-actions]

Persistent review updated to latest commit 5b979e4

[github-actions]

Persistent review updated to latest commit 6b898ff