FastFlowLM targets AMD Ryzen AI NPUs. Its move into the ROCm organization does not turn every ROCm-compatible GPU into a supported FastFlowLM device.
What version 1.0 adds
The August 11 release introduces SmolVLA support and expands flmbench testing across context sizes. The benchmark’s default iteration count is two; its test range is not a universal maximum context length for every model.
The project README identifies XDNA2 Ryzen AI NPUs as the target. Model assets and optimized kernels require an initial download. A small runtime installer therefore does not describe the total disk space or initial network access needed. Local inference and initial provisioning are separate stages.
Two Linux checks answer different questions
The Linux setup guide explicitly separates the kernel’s amdxdna device from XRT userspace. Its prerequisites include compatible firmware, a suitable kernel driver and the XRT stack. A Windows driver version copied into a Linux article is not an installation requirement for that system.
The read-only command flm validate checks the kernel DRM path. Model execution uses XRT, so a successful validation can coexist with “No such device with index '0'” from a model run. The guide recommends checking xrt-smi examine separately. Our diagram shows those two gates rather than inventing a successful terminal capture.
If the first gate passes and the second fails, repeatedly downloading the same model does not address the indicated device-discovery problem. Record the kernel, driver path, firmware and XRT device listing first. If both pass, a model-specific failure becomes a different investigation.
Measure the workload you intend to use
Compare the same model, prompt, output length and warm/cold state. Time to first output and sustained generation answer different questions; a robotics policy adds action-quality and control-loop concerns. We have not measured this release’s energy use. An NPU workload still belongs to a computer with CPU orchestration, memory and background services, so “runs on the NPU” cannot establish zero total CPU load.
September 8: separate Linux and Windows prerequisites, kernel detection and XRT execution; remove unqualified efficiency and zero-CPU claims.