Int8 quantization for RF-DETR seg support?
#2
by nvriese1 - opened
Question in title, are there full int8 quantized safetensors / pth, tensorrt (.engine) weights available? If not, is there documentation for how to quantize from fp32 to int8 from the seg (not object detection) weights via the rfdetr package or similar?
A standard full int8 quant via trtexec yields a non-performant graph due to the presence of precision sensitive layers.