A device package produced on the host
Ultralytics 8.4.107 adds a `format=ascend` path that runs YOLO models through Huawei's CANN compiler and produces `.om` packages. The release notes cover detection, segmentation, pose, oriented boxes, classification, and depth. Models for named targets such as Ascend 310P3 and 310B4 use static shapes and FP16 precision. The concrete delivery is an artifact format that can be compiled for specified hardware targets, rather than a new speed claim.[1]
Export can run on a Linux host without an attached Ascend device. That separation matters on the milestone ladder: the move from source model to a hardware-specific package can finish before the physical accelerator is installed. The package alone says nothing about correct output on an NPU, acceptable latency, or sustainable power behavior. Compilation is the first observable rung; an operating system belongs to a different environment.[1]
The dependency on the device side
Inference requires the CANN runtime and the `ais_bench` tool. The delivered chain therefore reads: source model, compilation on Linux, `.om` package, CANN runtime, measurement tool, and target Ascend NPU. “Ascend support” is no longer one software switch; it is a deployment boundary joining several versioned components. A narrower counterpoint is that a standardized package and tool may make most of those dependencies routine for an operator.[1]
The notes provide no speed, energy-use, or task-accuracy result for this path. A performance comparison with other accelerators therefore lacks a common workload, input shape, batch size, latency target, and wall-power boundary. Producing a hardware-specific file is a shipment-like milestone, but it is not a performance-per-watt result. Package count and useful compute are different quantities.[1]
The next rung is a measured device run
The same release reduces port collisions in distributed-training rendezvous, prevents empty segmentation annotations from corrupting boxes, and corrects class/AP misalignment in some evaluation plots. Those fixes do not measure Ascend speed, but they expose two separate failure boundaries: while the hardware artifact is produced, the meaning of training and evaluation measurements must remain intact. A file that runs on a device is not useful if its boxes or reported measurements are wrong.[1]
The release sits on a clear rung: the hardware-specific artifact and execution path have shipped, while a comparable operating result on the device has not. A run on an Ascend 310P3 or 310B4 that reports task, input shape, batch size, latency, accuracy, and wall power together would move the chain into measurable deployment. Until that measurement exists, the defensible statement is that Ultralytics has opened the route to Ascend; the capacity behind that route has not been quantified.[1]