Export YOLO to executorch for RISC-V Baremetal environment

[vaibhavmindgrove22]

I am trying to do object detection using YOLO on a RISC-V baremetal environment. The RISC-V core I am using is single core. It supports IMAFDC. I saw that exporting YOLO to executorch requires XNNPack backend. I have a few questions.

1. Is this realistically possible to be done without using XNNPack (which relies on pthreads and pthreads require an OS)?
2. Or can I run YOLO with XNNPack backend without using pthreads?
3. Also if there is no OS, will this run in single threaded mode in baremetal and not give any errors?

Any help regarding this would be appreciated.

cc @psiddh @AdrianLundell @digantdesai

[GregoryComer]

Hi, we don't have much acceleration in place for bare-metal RISC-V. You can, at a minimum, run with ExecuTorch's portable kernels, which are written to run with minimal requirements (including bare metal, so long as you can compile C++17), but will be quite slow. You could try this as a starting point - export the model with no partitioner.

XNNPACK itself supports RISC-V in some capacity, but this path isn't well tested through ExecuTorch. For the immediate pthreadpool blocker, ExecuTorch can be built threadless, but XNNPACK has deps on it. In theory, it might be possible to make ET + XNNPACK build for bare metal, as XNNPACK already supports a no-threadpool mode at runtime, but it would require deeper build + source changes to fully gate it out. Depending on how much effort you want to put it, you could maybe stub out pthreadpool as a no-op and find a build configuration to make that work.

[psiddh]

Hi @vaibhavmindgrove22 ,
Great to see you progressing from MNIST (#18573) to YOLO on your RISC-V bare metal platform!

As @GregoryComer mentioned, the most practical first step is to export YOLO with no partitioner (portable kernels only). You can adapt the conversion script from the executorch-examples YOLO Android app by simply removing the XnnpackPartitioner from to_edge_transform_and_lower(), just pass partitioner=[].

Also, I recently cane across this video FOSDEM 2026 talk on bringing up bare metal ExecuTorch on RISC-V (CORE-V CV32E40Pv2) , perhaps covers exactly the OS-dependency removal challenges you may face.. Worth a watch.

One practical note: YOLO26n at 640x640 needs (I think) ~5 MB for the input tensor alone plus model weights and activations. You may want to try imgsz=320 or smaller to fit your memory budget. What RAM/clock speed does your platform have?

[vaibhavmindgrove22]

@GregoryComer @psiddh
Thanks for your reply. I really appreciate your guidance. Will give you guys an update if I get a breakthrough.

[vaibhavmindgrove22]

Hi @vaibhavmindgrove22 , Great to see you progressing from MNIST (#18573) to YOLO on your RISC-V bare metal platform!

As @GregoryComer mentioned, the most practical first step is to export YOLO with no partitioner (portable kernels only). You can adapt the conversion script from the executorch-examples YOLO Android app by simply removing the XnnpackPartitioner from to_edge_transform_and_lower(), just pass partitioner=[].

Also, I recently cane across this video FOSDEM 2026 talk on bringing up bare metal ExecuTorch on RISC-V (CORE-V CV32E40Pv2) , perhaps covers exactly the OS-dependency removal challenges you may face.. Worth a watch.

One practical note: YOLO26n at 640x640 needs (I think) ~5 MB for the input tensor alone plus model weights and activations. You may want to try imgsz=320 or smaller to fit your memory budget. What RAM/clock speed does your platform have?

Yes @psiddh I watched the FOSDEM 2026 talk before doing the mnist classifier on ExecuTorch. It was very useful to me.
My microcontroller has 128KB SRAM (on chip memory) and 16MB PSRAM. The clock frequency is 700MHz to my knowledge.