⚠ This notebook has stopped taking notes: newest record is 23 days old (eng-0280), against an expected cadence of 14 days.
clm-0074
measured-herehigh ●●●
citable URL: https://halobench.com/records/clm-0074/ — this address never moves; the anchor /records/#clm-0074 keeps resolving
On qwen36-27b-mtp Q4_K_M, ROCm wins prefill decisively at every tested depth over Vulkan (357.8 vs 302.7 t/s at d0, +18%; 210.5 vs 95.7 t/s at d32768, +120%) and is the ONLY backend that completed the deepest matrix cell (d131072) without a device-loss event — Vulkan device-lost on BOTH allowed attempts at that depth (vk::DeviceLostError, GPU wedged, recovered via kernel amdgpu ring reset each time; ROCm completed cleanly on its first attempt at 95.4 pp / 8.44 tg t/s). Decode is roughly matched between backends, Vulkan marginally ahead (12.83 vs 12.08 t/s at d0, +6%; 11.33 vs 10.86 t/s at d32768, +4%). ROCm is this bench's serving-config pick on both grounds: it is faster where it matters most (prefill, and by a wide margin at depth) and it is the only backend that reaches the full depth range this candidate's 262K native context makes plausible to serve at.
METHOD: llama-bench, stock 3653e6d (ROCm) / 3653e6d6d (Vulkan), explicit f16 KV, -fa 1, --load-mode none, pp1024/tg256. N=3 fresh-process reps at d0/d32768 (max CV 1.6% on ROCm prefill, 3.0% on Vulkan prefill at d32768 — at the house 3% flag threshold, but decode at the same depth was clean at 0.03%, read as ordinary prefill scatter rather than a defective-path signature), N=1 at d131072 per house convention.
DEVICE-LOSS EVIDENCE: dmesg -T (sudo) captured for both Vulkan d131072 attempts — "amdgpu 0000:c5:00.0: ring comp_1.2.0 timeout" / "Starting comp_1.2.0 ring reset" / "Ring comp_1.2.0 reset succeeded" / "device wedged, but recovered through reset" — both attempts hit the identical failure signature and both recovered cleanly (no reboot needed). This matches the laguna-s-21/deepseek-v4-flash device-loss precedent at comparable depth on this silicon; ornith-35b broke that streak once (its own Vulkan arm completed d131072 cleanly), this candidate does not — the streak-break was not a general fix, it was ornith-specific.
No MTP/speculation confound: this candidate's staged GGUF carries no MTP tensors at all (clm-0073), so every cell in this matrix is plain decode on both backends — the backend comparison is not entangled with a speculation lever.