Struct Flux2Runner 

pub struct Flux2Runner { /* private fields */ }

FLUX.2 denoiser runner — native CPU or compiled HIR on any Device.

Implementations

impl Flux2Runner

pub fn builder() -> Flux2RunnerBuilder

pub fn drop_text_encoder_weights(&self) -> Result<(), Error>

pub fn config(&self) -> &Flux2Config

pub fn device(&self) -> Device

pub fn batch(&self) -> usize

pub fn img_seq(&self) -> usize

pub fn txt_seq(&self) -> usize

pub fn uses_nvfp4(&self) -> bool

pub fn has_text_encoder(&self) -> bool

pub fn has_vae(&self) -> bool

pub fn uses_compiled_denoiser(&self) -> bool

True when denoiser forwards use compiled HIR (Self::device).

pub fn uses_compiled_text_encoder(&self) -> bool

True when text encoding uses compiled HIR on Self::device.

pub fn uses_compiled_vae(&self) -> bool

pub fn warmup_denoiser( &self, img_ids: &[f32], txt_ids: &[f32], ) -> Result<(), Error>

Pre-compile the denoiser HIR for the given position ids (RoPE tables are baked in).

pub fn encode_prompt(&self, prompt: &str) -> Result<(Vec<f32>, Vec<f32>), Error>

Encode a text prompt into FLUX.2 encoder_hidden_states and txt_ids.

Uses compiled HIR on Metal / MLX when Self::uses_compiled_text_encoder; native CPU on CUDA / ROCm / wgpu / Vulkan and CPU otherwise.

pub fn encode_prompt_native( &self, prompt: &str, ) -> Result<(Vec<f32>, Vec<f32>), Error>

Native CPU text encoder (no IR compile).

pub fn encode_prompt_compiled( &self, prompt: &str, ) -> Result<(Vec<f32>, Vec<f32>), Error>

Encode via compiled text-encoder HIR on Self::device.

pub fn forward( &self, hidden_states: &[f32], encoder_hidden_states: &[f32], timestep: &[f32], guidance: Option<&[f32]>, img_ids: &[f32], txt_ids: &[f32], ) -> Result<Flux2Output, Error>

One denoiser forward: latents + text context → noise prediction.

pub fn vae_encode_rgb( &self, rgb: &[f32], pixel_h: usize, pixel_w: usize, ) -> Result<Vec<f32>, Error>

VAE encode RGB planar [-1,1] NCHW → latent (compiled on GPU when enabled).

pub fn encode_rgb_to_packed( &self, rgb: &[f32], pixel_h: usize, pixel_w: usize, latent_h: usize, latent_w: usize, eff_h: usize, eff_w: usize, ) -> Result<Vec<f32>, Error>

Encode planar RGB [-1,1] NCHW to packed transformer latents.

pub fn has_vae_encoder(&self) -> bool

pub fn prepare_img2img_packed( &self, rgb: &[f32], pixel_h: usize, pixel_w: usize, latent_h: usize, latent_w: usize, eff_h: usize, eff_w: usize, noise: &[f32], image_strength: f32, num_inference_steps: usize, ) -> Result<Vec<f32>, Error>

img2img: encode source RGB and blend with noise at the strength-derived sigma.

pub fn prepare_edit_conditioning( &self, images: &[(&[f32], usize, usize)], eff_h: usize, eff_w: usize, latent_h: usize, latent_w: usize, ) -> Result<Flux2ReferenceConditioning, Error>

Edit mode: encode reference images into concat conditioning tokens.

pub fn forward_noise( &self, hidden_states: &[f32], encoder_hidden_states: &[f32], timestep: &[f32], guidance: Option<&[f32]>, img_ids: &[f32], txt_ids: &[f32], ) -> Result<Vec<f32>, Error>

Denoiser noise prediction (compiled on Self::device when not CPU-native).

pub fn forward_noise_native( &self, hidden_states: &[f32], encoder_hidden_states: &[f32], timestep: &[f32], guidance: Option<&[f32]>, img_ids: &[f32], txt_ids: &[f32], ) -> Result<Vec<f32>, Error>

Native CPU reference forward (no IR compile).

pub fn forward_noise_dual_native( &self, hidden_states: &[f32], encoder_hidden_states: &[f32], timestep: &[f32], timestep_target: &[f32], guidance: Option<&[f32]>, img_ids: &[f32], txt_ids: &[f32], ) -> Result<Vec<f32>, Error>

Native forward with dual-time embedding (flow-map).

pub fn forward_noise_compiled( &self, hidden_states: &[f32], encoder_hidden_states: &[f32], timestep: &[f32], guidance: Option<&[f32]>, img_ids: &[f32], txt_ids: &[f32], ) -> Result<Vec<f32>, Error>

Compiled HIR denoiser on Self::device (Metal / MLX / CUDA / CPU).

pub fn forward_noise_dual_compiled( &self, hidden_states: &[f32], encoder_hidden_states: &[f32], timestep: &[f32], timestep_target: &[f32], guidance: Option<&[f32]>, img_ids: &[f32], txt_ids: &[f32], ) -> Result<Vec<f32>, Error>

Compiled forward with dual-time temb (flow-map).

pub fn forward_cfg( &self, hidden_states: &[f32], pos_encoder: &[f32], neg_encoder: &[f32], timestep: &[f32], guidance: Option<&[f32]>, img_ids: &[f32], pos_txt_ids: &[f32], neg_txt_ids: &[f32], cfg_scale: f32, ) -> Result<Flux2Output, Error>

Classifier-free guidance: positive + negative text, then neg + cfg_scale * (pos - neg) on the noise prediction.

pub fn encode_prompt_pair( &self, prompt: &str, negative_prompt: Option<&str>, ) -> Result<(Vec<f32>, Vec<f32>, Option<Vec<f32>>, Option<Vec<f32>>), Error>

Tokenize and encode positive + optional negative prompts.

pub fn vae_config(&self) -> Option<&Flux2VaeConfig>

pub fn decode_to_rgb( &self, packed_latents: &[f32], img_ids: &[f32], latent_h: usize, latent_w: usize, ) -> Result<(Vec<u8>, u32, u32), Error>

Decode denoised packed latents to interleaved RGB u8 (HWC) and pixel (height, width).

pub fn decode_to_rgb_native( &self, packed_latents: &[f32], img_ids: &[f32], latent_h: usize, latent_w: usize, ) -> Result<(Vec<u8>, u32, u32), Error>

Native CPU decode (unpack / BN / unpatchify + VAE decoder).

pub fn decode_to_rgb_compiled( &self, packed_latents: &[f32], img_ids: &[f32], latent_h: usize, latent_w: usize, ) -> Result<(Vec<u8>, u32, u32), Error>

Compiled VAE decoder on Self::device (unpack/BN/unpatchify stay on CPU).