Skip to main content

PCIe Co-Processing

Run inference from a host connected to a Modalix PCIe Card. Use these tutorials in order.

The PCIe host API accepts either a model-ready tensor or decoded image pixels. Tensor mode keeps preprocessing on the host. I...

PCIeinferencetensorimagedetection

Synchronous `run()` is ideal for a first inference, but it waits for each result before submitting the next image. A producer...

PCIeasynchronousthroughputdetection
Run Multiple Models image Run Multiple Models 15 minutes

A Modalix PCIe Card exposes queues 0 through 3. Each active `Model` owns one queue, so independent models can execute concurr...

PCIequeuesconcurrencyclassificationdetection

The default PCIe route sends FP32 tensors to the card, where the input is quantized, the MLA runs, and the output is dequanti...

PCIeMLAINT8quantizationtensor

`Tensor::from_external()` lets a C++ application wrap an existing contiguous allocation instead of copying it into a new host...

PCIeC++tensorexternal memoryzero-copy wrapping