clm-0046
measured-heremed ●●○
citable URL: https://halobench.com/records/clm-0046/ — this address never moves; the anchor /records/#clm-0046 keeps resolving
The community BF16 flash-attention predictions (clm-0044) reproduce on this hardware under single-binary methodology: at 32k depth, stock Vulkan leads ROCm +17.4% prefill / +17.1% decode at f16 KV, and the PR-26856 patch inverts prefill to ROCm +46.9% with bf16 KV. At shallow depth every gap collapses (+2.6% / -1.0%), so backend comparisons are depth-statements or they are nothing. The fleet build delta is +2.8% pp / +0.9% tg with identical tau2 capability (anchor pair).
Note — the record's own working
One binary (77a9a66eb, HIP+Vulkan), device selection the only variable — tighter than either community source. Vulkan bf16 cells succeeded despite the device banner reporting bf16:0 — treated as a cast/fallback path, not native compute. The depth cells were re-run after a SIGPIPE/pipefail flag-detection bug silently substituted a prefill proxy (which would have shown +24.6% instead of +46.9% — the bug class is documented in methodology-lessons). PR 26856 remains unmerged and unreviewed upstream: adopt-watch, with the anchor-pair policy governing any adoption.
Cited by — computed at build time, never stored
model pages qwen35-122b