Struct Sam 

pub struct Sam { /* private fields */ }

Full SAM model — owns the compiled image encoder + all decoder weights. Stateless wrt prompts: every call to Sam::forward runs the cached encoder + a fresh decoder forward.

Implementations

impl Sam

pub fn from_safetensors( weights_path: &str, cfg: SamConfig, ) -> Result<Sam, Error>

Load SAM ViT-B (or L/H — pass the matching config) from a safetensors checkpoint, compiling the image encoder for the CPU backend. For GPU/Metal/MLX, use Sam::from_safetensors_on.

pub fn from_safetensors_on( weights_path: &str, cfg: SamConfig, device: Device, ) -> Result<Sam, Error>

Same as Sam::from_safetensors but compiles the image encoder for the given device. Requires the matching backend feature on rlx-models:

feature	backend
metal	Device::Metal
mlx	Device::Mlx
gpu	Device::Gpu
cuda	Device::Cuda
rocm	Device::Rocm
tpu	Device::Tpu

pub fn encode_image(&mut self, image_nchw: &[f32]) -> Vec<f32>

Encode an image into the [256, 64, 64] image embedding. image_nchw: pre-padded [3, 1024, 1024] NCHW f32 tensor (see super::preprocess::preprocess_image).

pub fn predict_masks( &mut self, image_embeddings: &[f32], points: Option<(&[f32], &[f32])>, boxes: Option<&[f32]>, masks: Option<&[f32]>, multimask_output: bool, ) -> Result<MaskPrediction, Error>

Run the prompt encoder + mask decoder on a pre-encoded image.

pub fn forward( &mut self, rgb: &[u8], h_in: usize, w_in: usize, points: Option<(&[f32], &[f32])>, boxes: Option<&[f32]>, masks: Option<&[f32]>, multimask_output: bool, ) -> Result<(MaskPrediction, (usize, usize)), Error>

End-to-end forward: image bytes → masks. rgb is HWC u8.

pub fn config(&self) -> &SamConfig

pub fn mask_side(&self) -> usize

Spatial side length of the predicted mask logits (= 4 · hw = 256 for ViT-B at 1024 input).

pub fn input_image_size(&self) -> usize

Image side that the model operates on internally.