On Vulkan, enabling Flash Attention can cut prompt processing by ~60% and decode by ~48%

On Strix Halo (RADV, warp size 64) llama-bench shows FA on costing more than it saves: llama-2-7b.Q4_0 pp512 drops 1252.63 -> 508.36 tok/s and tg128 drops 49.91 -> 25.80 tok/s. Qwen3.5-122B-A10B-GGUF Q4_K_XL-MTP drops pp512 325.58 -> 142.33 tok/s and tg128 22.42 -> 7.x tok/s.

Not established. The Vulkan Flash Attention path on this driver/GPU is slower than the non-FA path rather than faster; the issue is labelled bug-unconfirmed and stale upstream, so the mechanism is unidentified. Note the same log line reports bf16: 0 and warp size: 64, both of which plausibly matter for FA kernel selection.

Benchmark with and without -fa on your own GPU and quantization before enabling it; on this system leave FA off. Do not assume FA is a free win on Vulkan the way it is on CUDA.

Confidence is deliberately low: the numbers are reproducible but the root cause is unknown. Also beware conflating this with gotcha-vulkan-q4km-speed-regression, a separate bisected commit that changed Vulkan Q4_K_M decode.