[bong-water-water-bong]

Bumps the private HIP-kernel add-on pin (config/gpu-kernels.config) to the merge of 1bit-MONSTER/gpu-kernels#48 (262c2f9aced4), for engine#782: the m128_w8_f2_sb / _sb tile family is now off by default, so the clang-24 spill that cost 3.3× of Q4_K prefill is no longer the selected path.

The attribution this bump closes

engine#782 (same llama.cpp pin, same dispatch): rc-2026-10-31c lib = clang-23 code objects (49× clang 23), tip = clang 24 (52×). Clang-24 _sb tiles spill 320-576 B/lane; the default GGML_HRX_HIP_DENSE_PREFILL_M128_SB=1 selected them. Q4_K M=16384 K=2560 N=9728, same hip.dense_prefill.mm in every arm:

arm	ms	TFLOP/s
rc lib (clang-23 objects)	20.33 / 20.48	40.1
tip, old default (_sb)	70.81 / 70.92	11.5 (3.47x)
tip + this default flip (mosaic_bench, rebuilt libggml-hrx with the rebased addon)	21.11 / 20.42	38.7 / 40.0
tip, M128_SB=1 (old behaviour, still available)	68.61 / 69.22	11.9

Clock-matched at ~1.94-1.97 GHz; the residual ~5 % vs the rc binary is clang-24 vs clang-23 codegen on the scratch-free tile (ROCm/llvm side, not a fork change).

The fork change

gpu-kernels #48 rebased onto current main (94bc42e) after #59/#1064 rewrote the same m128_sb/m128_fits region — both changes kept (swiglu_tile follows m128_sb(), and the default flip lives inside sb_enabled()). Verified on strixhalo: tools/check.sh gfx1151 "headers match the manifest (409 kernels)", 45 add-on objects, 93 artifacts (only the 10 pre-existing hyperconn/lightning fingerprint failures, engine#52); tools/check-probes.sh PASS with the plain table when unset and the old _sb table at =1; the bench above.

Merge gate

The engine's plain build does not compile the add-on; the change is host-side dispatch code whose targeted evidence is the add-on's check/probes plus the bench on the real tensors, recorded above (the lane also offered hw-gate, but the box's gpu-hold FIFO makes a gate cycle a >1 h wait for a non-device change; that policy question is parked in #1077). Merging on green build + this evidence, as the previous gpu-kernels bump (#1246) was.

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🎫 Ticket compliance analysis 🔶 1064 - Partially compliant Compliant requirements: The swiglu_tile() function now consults m128_sb() to honor the GGML_HRX_HIP_DENSE_PREFILL_M128_SB knob The PR updates the gpu-kernels pin to include the necessary changes for this behavior Non-compliant requirements: Verification that the plain tile (m128_w8_f1) exists for every format the fused matcher admits is not explicitly shown in the diff The m64_w8_f1 branch behavior is not explicitly addressed in the diff Requires further human verification: Confirmation that the plain tile m128_w8_f1 exists for all relevant formats Measurement of the m64_w8_f1 branch behavior to ensure it remains unchanged 1077 - Partially compliant Compliant requirements: The PR updates the gpu-kernels pin, which is a change to the HRX backend Non-compliant requirements: The PR does not explicitly address the CI job configuration or labeling requirements Requires further human verification: The CI job configuration needs to be verified to ensure the HRX backend is compiled The labeling of PRs touching third_party/llama.cpp or ggml/src/ggml-hrx needs to be confirmed 782 - Partially compliant Compliant requirements: The PR updates the gpu-kernels pin to include the change that disables _sb tiles by default The PR description confirms that the clang-24 spill no longer costs 3.3x of Q4_K prefill Non-compliant requirements: The PR does not include performance benchmarks to confirm the fix Requires further human verification: Performance benchmarks should be run to confirm the fix for the Q4_K prefill regression

⏱️ Estimated effort to review: 3 🔵🔵🔵⚪⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ Recommended focus areas for review Pin Update The PR updates the gpu-kernels pin to commit 262c2f9aced4f18d0f78c301c9f71030d4d44059. This change is critical for addressing performance issues related to the _sb tiles and ensuring that the HRX backend is properly compiled and tested. sha=262c2f9aced4f18d0f78c301c9f71030d4d44059