On Intel Arc dGPUs, llama.cpp agentic workloads stall and enabling FA can cut prompt processing ~15x

On a Ryzen 5600g with Intel Arc Pro B50 (Battlemage), agentic coding with Devstral 2 small 24b or GPT-OSS 120b stalls in the first few agent steps; prompt processing drops to single digits and effectively stops once context reaches 4k-8k tokens. Measured without FA: ~300 tok/s pp; with FA enabled: ~15-20 tok/s. SYCL is worse - black screen and OOM when context fills, with 8 GB still free on the GPU.

Not established. The reporter tried Vulkan and SYCL, latest devel kernels and mesa 26.0, multiple llama.cpp builds including one with XE2 improvements, FA on/off, ctk/ctv q8_0 on/off, various UB/B combinations and several offloading strategies, on both Windows and Linux, with no configuration that works.

No reliable workaround reported. Intel dGPU support in llama.cpp should be treated as experimental for long-context or agentic workloads; validate on your exact SKU before committing to it.

Confidence is medium: it is one user's hardware-specific report with no maintainer diagnosis, but the configuration sweep is thorough and the FA regression is quantified. The FA-on pp drop parallels gotcha-vulkan-flash-attn-slower on a different vendor's stack, which suggests FA dispatch is the thing to A/B first on any non-CUDA backend.