GPU & half precision

With predict / process

Move the model; the batch follows. predict reads the device and dtype off the model's first floating-point parameter and moves the processor's output to match, so there is nothing else to do:

from nobg import AutoModel

model = AutoModel.from_pretrained("feyninc/FeyNobg").to("cuda")
model.process("input.jpg").save("output.png")

A half-precision model works the same way:

import torch

model = AutoModel.from_pretrained("feyninc/FeyNobg").to("cuda", torch.bfloat16)
model.process("input.jpg").save("output.png")

Integer inputs (SAM3's input_ids, attention_mask) are moved to the device but keep their dtype. predict also handles eval mode for you — it calls model.eval(), runs under no_grad, and restores training mode afterwards if the model was in it.

Manually, with autocast

model = AutoModel.from_pretrained("feyninc/FeyNobg").eval().to("cuda")
inputs = processor(image, return_tensors="pt").to("cuda")

with torch.no_grad(), torch.autocast("cuda", dtype=torch.bfloat16):
    outputs = model(pixel_values=inputs["pixel_values"])

Autocast keeps the weights in float32 and casts per-op, which is usually the safer half-precision route than casting the model outright.

Notes

• Post-processing is dtype-agnostic. post_process_alpha_matting sigmoids and interpolates in whatever dtype the logits arrive in; predict casts the mattes to CPU float32 before compositing.
• refine_foreground always computes in ≥ float32, whatever you hand it — its internal divisions overflow in half precision, which shows up as black patches — then casts back to the input dtype. See Refining edges.
• SAM3 needs an exact input size. pixel_values must be config.image_size square regardless of device or dtype; the vision tower's rotary tables are fixed-size buffers.
• ONNX exports run on CPU float32 by default. OnnxModel reports no parameters, which is how the shared predict reads it as CPU float32 — matching what the session accepts. Install onnxruntime-gpu and the loader will pick up the GPU provider. See ONNX export.