implement cumprod

.gitignore

@@ -66,7 +66,3 @@ scraps/
 # These should be built by the platform
 windows/sqlite/sqlite3.exp
 windows/sqlite/sqlite3.lib
-
-# ArtCraft Windows DLLs
-crates/desktop/src-tauri/*.dll
-windows_dlls/

.idea/dictionaries/bt.xml

@@ -2,7 +2,6 @@
   <dictionary name="bt">
     <words>
       <w>arpabet</w>
-      <w>artcraft</w>
       <w>backfill</w>
       <w>backfilling</w>
       <w>backfills</w>

Cargo.toml

@@ -23,8 +23,6 @@ members = [
 
   # Desktop apps
   "crates/desktop/src-tauri",
-  "crates/desktop/ml_models",
-  "crates/desktop/ml_weights_registry",
 
   # Shared libraries
   "crates/lib/actix_cors_configs",
@@ -98,17 +96,6 @@ members = [
   "crates/service/plugins/user_traits_component",
 
   # Vendored code
-  "crates/vendor/candle/candle-core",
-  "crates/vendor/candle/candle-datasets",
-  "crates/vendor/candle/candle-examples",
-  "crates/vendor/candle/candle-flash-attn",
-  "crates/vendor/candle/candle-kernels",
-  "crates/vendor/candle/candle-metal-kernels",
-  "crates/vendor/candle/candle-nn",
-  "crates/vendor/candle/candle-onnx",
-  "crates/vendor/candle/candle-pyo3",
-  "crates/vendor/candle/candle-transformers",
-  "crates/vendor/candle/tensor-tools",
   "crates/vendor/limitation",
   # NB(bt,2023-12-12): Newrelic keeps breaking with upgrades. Kill for now.
   #"crates/vendor/newrelic-telemetry-sdk-rust", 

app_config.yaml

@@ -1,8 +1,7 @@
-image_width: 1024
-image_height: 1024
+image_width: 512
+image_height: 512
 #scheduler_steps: 15
 # NB(bt): It looks like we can get down to 2 sampling steps, perhaps ?
-scheduler_steps: 8
+scheduler_steps: 2
 scheduler_samples: 1
-cfg_scale: 15.0
-seed: 42
+cfg_scale: 1.1

crates/desktop/ml_models/Cargo.toml

@@ -1,79 +0,0 @@
-[package]
-name = "ml_models"
-edition = "2021"
-version = "0.0.1"
-authors = ["storyteller"]
-publish = false
-description = ""
-license = ""
-repository = ""
-
-[lib]
-name = "ml_models"
-path = "src/lib.rs"
-crate-type = ["staticlib", "cdylib", "rlib"]
-
-[package.metadata.cargo-udeps.ignore]
-normal = ["my-workspace-hack"]
-
-[dependencies]
-
-# Internal
-filesys = { path = "../../lib/files/filesys" }
-memory_caching = { path = "../../lib/caching/memory_caching" }
-ml_weights_registry = { path = "../ml_weights_registry" }
-primitives = { path = "../../lib/data/primitives" }
-
-# Bypass cudarc issues on Windows https://github.com/coreylowman/cudarc/issues/332
-cudarc = "0.13.7"
-
-# External
-anyhow = "1.0"
-base64 = "0.22.1"
-bytes = "1.10"
-candle-core = { path = "../../vendor/candle/candle-core", features = ["cuda"] }
-candle-nn = { path = "../../vendor/candle/candle-nn", features = ["cuda"] }
-candle-transformers = { path = "../../vendor/candle/candle-transformers", features = ["cuda"] }
-const_format = "0.2.34"
-directories = "6.0.0"
-dirs = "6.0.0"
-futures = "0.3"
-half = { version = "2.3.1", features = ["num-traits", "use-intrinsics", "rand_distr"] }
-hf-hub = "0.3.2"
-image = "0.25.5"
-log = "0.4"
-ndarray = "0.15.6"
-nvidia-checker = "0.1.2"
-once_cell = "1.20.3"
-ort = { version = "1.16.2", features = ["cuda"] }
-rand = "0.8.5"
-# NB(bt,2025-03-10): default-features = false on Windows while I debug openssl / rustls
-reqwest = { version = "0.12.12", features = ["stream"], default-features = false }
-serde = { version = "1.0", features = ["derive"] }
-serde_json = "1.0"
-serde_yaml = "0.9"
-tauri = { version = "2.2.4", features = [] }
-tauri-plugin-log = "2"
-tempdir = "0.3.7"
-tempfile = "3.18.0"
-tokenizers = "0.15.2"
-tokio = { version = "1.42", features = ["rt", "rt-multi-thread", "fs"] }
-tokio-util = { version = "0.7", features = ["codec"] }
-
-# tracing-subscriber = { version = "0.3", features = [ "env-filter", "fmt" ] }
-
-[features]
-default = ["cuda"]
-cuda = [
-    "candle-core/cuda",
-    "candle-nn/cuda",
-    "candle-transformers/cuda",
-    # NB(bt,2025-03-10): Building without Flash attention on Windows for now to bypass linker issues
-    "candle-transformers/flash-attn",
-    "candle-core/cudnn"
-]
-
-[dev-dependencies]
-libloading = "0.8.6"
-serial_test = "3.2.0"
-assertor = "0.0.3"

crates/desktop/ml_models/src/lib.rs

@@ -1 +0,0 @@
-pub mod ml;

crates/desktop/ml_models/src/ml/background_removal/mod.rs

@@ -1,2 +0,0 @@
-pub mod onnx_session;
-pub mod remove_image_background;

crates/desktop/ml_models/src/ml/background_removal/onnx_session.rs

@@ -1,19 +0,0 @@
-//! Adapted from https://github.com/dnanhkhoa/rust-background-removal/blob/main/src/onnx.rs
-
-use std::path::Path;
-use ort::{Environment, ExecutionProvider, GraphOptimizationLevel, SessionBuilder};
-
-pub fn onnx_session(onnx_model_file: &Path) -> Result<ort::Session, anyhow::Error> {
-  let environment = Environment::default().into_arc();
-  let session = SessionBuilder::new(&environment)?
-    .with_optimization_level(GraphOptimizationLevel::Level1)? // Configure model optimization level
-    .with_intra_threads(1)? // Configure the number of threads used for inference
-    .with_execution_providers([
-      // Configure execution providers (e.g., CUDA, CoreML, CPU)
-      ExecutionProvider::CUDA(Default::default()),
-      ExecutionProvider::CoreML(Default::default()),
-      ExecutionProvider::CPU(Default::default()),
-    ])?
-    .with_model_from_file(onnx_model_file)?;
-  Ok(session)
-}

crates/desktop/ml_models/src/ml/background_removal/remove_image_background.rs

@@ -1,61 +0,0 @@
-use crate::ml::background_removal::onnx_session::onnx_session;
-use image::{imageops, DynamicImage};
-use ndarray::{Array, CowArray};
-use ort::Value;
-use std::path::Path;
-
-pub async fn remove_image_background<P: AsRef<Path>>(model_path: P, image: DynamicImage) -> anyhow::Result<DynamicImage> {
-  // NB: Real time needed cuDNN 9, but this uses cuDNN 8:
-  // sudo apt-get install libcudnn8-dev libcudnn8
-  // update-alternatives: warning: forcing reinstallation of alternative /usr/include/x86_64-linux-gnu/cudnn_v9.h because link group libcudnn is broken
-  // update-alternatives: using /usr/include/x86_64-linux-gnu/cudnn_v8.h to provide /usr/include/cudnn.h (libcudnn) in manual mode
-  let session = onnx_session(model_path.as_ref())?;
-  let image = process_dynamic_image(&session, image)?;
-  Ok(image)
-}
-
-fn process_dynamic_image(
-  session: &ort::Session,
-  dynamic_img: DynamicImage,
-) -> Result<DynamicImage, anyhow::Error> {
-  let input_shape = session.inputs[0]
-    .dimensions()
-    .map(|dim| dim.unwrap())
-    .collect::<Vec<usize>>();
-  let input_img = dynamic_img.into_rgba8();
-  let scaling_factor = f32::min(
-    1., // Avoid upscaling
-    f32::min(
-      input_shape[3] as f32 / input_img.width() as f32, // Width ratio
-      input_shape[2] as f32 / input_img.height() as f32, // Height ratio
-    ),
-  );
-  let mut resized_img = imageops::resize(
-    &input_img,
-    input_shape[3] as u32,
-    input_shape[2] as u32,
-    imageops::FilterType::Triangle,
-  );
-  let input_tensor = CowArray::from(
-    Array::from_shape_fn(input_shape, |indices| {
-      let mean = 128.;
-      let std = 256.;
-
-      (resized_img[(indices[3] as u32, indices[2] as u32)][indices[1]] as f32 - mean) / std
-    })
-      .into_dyn(),
-  );
-  let inputs = vec![Value::from_array(session.allocator(), &input_tensor)?];
-  let outputs = session.run(inputs)?;
-  let output_tensor = outputs[0].try_extract::<f32>()?;
-  for (indices, alpha) in output_tensor.view().indexed_iter() {
-    resized_img[(indices[3] as u32, indices[2] as u32)][3] = (alpha * 255.) as u8;
-  }
-  let output_img = imageops::resize(
-    &resized_img,
-    (input_img.width() as f32 * scaling_factor) as u32,
-    (input_img.height() as f32 * scaling_factor) as u32,
-    imageops::FilterType::Triangle,
-  );
-  Ok(DynamicImage::ImageRgba8(output_img))
-}

crates/desktop/ml_models/src/ml/model_cache.rs

@@ -1,60 +0,0 @@
-use anyhow::anyhow;
-use candle_transformers::models::stable_diffusion::unet_2d::UNet2DConditionModel;
-use candle_transformers::models::stable_diffusion::vae::AutoEncoderKL;
-use std::sync::{Arc, RwLock};
-// TODO: This data structure is gross. Generalize the locking and loading semantics for reuse.
-
-// Simple registry for now. We can build complex machinery that aids in 
-// VRAM utilization, disk space saving, intelligent scheduling, etc. in the future.
-pub struct ModelCache {
-  vae: RwLock<Option<Arc<AutoEncoderKL>>>,
-  unet2: RwLock<Option<Arc<UNet2DConditionModel>>>,
-}
-
-impl ModelCache {
-  pub fn new() -> Self {
-    Self { 
-      vae: RwLock::new(None),
-      unet2: RwLock::new(None),
-    }
-  }
-  
-  pub fn set_vae(&self, vae: Arc<AutoEncoderKL>) -> anyhow::Result<()> {
-    match self.vae.write() {
-      Err(err) => Err(anyhow!("{}", err)),
-      Ok(mut lock) => {
-        *lock = Some(vae);
-        Ok(())
-      }
-    }
-  }
-
-  pub fn get_vae(&self) -> anyhow::Result<Option<Arc<AutoEncoderKL>>> {
-    match self.vae.read() {
-      Err(err) => Err(anyhow!("{}", err)),
-      Ok(ref lock) => {
-        Ok(lock.as_ref().map(|v|v.clone()))
-      }
-    }
-  }
-
-  pub fn set_unet(&self, vae: Arc<UNet2DConditionModel>) -> anyhow::Result<()> {
-    match self.unet2.write() {
-      Err(err) => Err(anyhow!("{}", err)),
-      Ok(mut lock) => {
-        *lock = Some(vae);
-        Ok(())
-      }
-    }
-  }
-
-  pub fn get_unet(&self) -> anyhow::Result<Option<Arc<UNet2DConditionModel>>> {
-    match self.unet2.read() {
-      Err(err) => Err(anyhow!("{}", err)),
-      Ok(ref lock) => {
-        Ok(lock.as_ref().map(|v|v.clone()))
-      }
-    }
-  }
-  
-}

crates/desktop/ml_models/src/ml/model_config.rs

@@ -1,61 +0,0 @@
-use crate::ml::model_file::StableDiffusionVersion;
-use candle_core::{DType, Device};
-use candle_transformers::models::stable_diffusion::StableDiffusionConfig;
-use hf_hub::api::sync::Api;
-use log::info;
-
-const DEFAULT_SD_IMAGE_WIDTH: usize = 1024;
-const DEFAULT_SD_IMAGE_HEIGHT: usize = 1024;
-const DEFAULT_SD_NUMERIC_DATATYPE: DType = DType::F32;
-const SD_VERSION: StableDiffusionVersion = StableDiffusionVersion::V1_5;
-
-pub struct ModelConfig {
-
-  /// We only support one device type currently, and we can't multiplex
-  /// over several devices yet.
-  pub device: Device,
-
-  /// The core numeric type to use for models.
-  pub dtype: DType,
-
-  pub sd_version: StableDiffusionVersion,
-  pub sd_config: StableDiffusionConfig,
-
-  /// Probably shouldn't be here.
-  pub hf_api: Api,
-}
-
-impl ModelConfig {
-  pub fn init() -> anyhow::Result<Self> {
-    let device = Device::new_cuda(0)
-      .unwrap_or_else(|e| {
-          info!("CUDA not available ({}), falling back to CPU", e);
-          Device::Cpu
-        });
-
-    let sd_config = match SD_VERSION {
-      StableDiffusionVersion::Turbo => {
-        StableDiffusionConfig::sdxl_turbo(None, Some(DEFAULT_SD_IMAGE_HEIGHT), Some(DEFAULT_SD_IMAGE_WIDTH))
-      }
-      StableDiffusionVersion::V1_5 | StableDiffusionVersion::V1_5Inpaint => {
-        StableDiffusionConfig::v1_5(None, Some(DEFAULT_SD_IMAGE_HEIGHT), Some(DEFAULT_SD_IMAGE_WIDTH))
-      }
-      StableDiffusionVersion::V2_1 | StableDiffusionVersion::V2Inpaint => {
-        StableDiffusionConfig::v2_1(None, Some(DEFAULT_SD_IMAGE_HEIGHT), Some(DEFAULT_SD_IMAGE_WIDTH))
-      }
-      StableDiffusionVersion::Xl | StableDiffusionVersion::XlInpaint => {
-        StableDiffusionConfig::sdxl(None, Some(DEFAULT_SD_IMAGE_HEIGHT), Some(DEFAULT_SD_IMAGE_WIDTH))
-      }
-    };
-
-    let hf_api = Api::new()?;
-
-    Ok(Self {
-      device,
-      dtype: DEFAULT_SD_NUMERIC_DATATYPE,
-      sd_version: SD_VERSION,
-      sd_config,
-      hf_api,
-    })
-  }
-}

crates/desktop/ml_models/src/ml/model_file.rs

@@ -1,180 +0,0 @@
-#[cfg(feature = "accelerate")]
-extern crate accelerate_src;
-#[cfg(feature = "mkl")]
-extern crate intel_mkl_src;
-
-use candle_transformers::models::stable_diffusion;
-use anyhow::{Error as E, Result};
-use hf_hub::api::sync::Api;
-#[derive(Debug, Clone, Copy, PartialEq, Eq)]
-pub enum StableDiffusionVersion {
-    V1_5,
-    V1_5Inpaint,
-    V2_1,
-    V2Inpaint,
-    Xl,
-    XlInpaint,
-    Turbo,
-}
-
-#[derive(Debug, Clone, Copy, PartialEq, Eq)]
-pub enum ModelFile {
-    Tokenizer,
-    Tokenizer2,
-    Clip,
-    Clip2,
-    Unet,
-    UnetLcm,
-    Vae,
-}
-
-impl StableDiffusionVersion {
-    pub fn repo(&self) -> &'static str {
-        match self {
-            Self::XlInpaint => "diffusers/stable-diffusion-xl-1.0-inpainting-0.1",
-            Self::Xl => "stabilityai/stable-diffusion-xl-base-1.0",
-            Self::V2Inpaint => "stabilityai/stable-diffusion-2-inpainting",
-            Self::V2_1 => "stabilityai/stable-diffusion-2-1",
-            Self::V1_5 => "Lykon/dreamshaper-7",
-            Self::V1_5Inpaint => "stable-diffusion-v1-5/stable-diffusion-inpainting",
-            Self::Turbo => "Lykon/dreamshaper-7",
-        }
-    }
-
-    fn unet_file(&self, use_f16: bool) -> &'static str {
-        match self {
-            Self::V1_5
-            | Self::V1_5Inpaint
-            | Self::V2_1
-            | Self::V2Inpaint
-            | Self::Xl
-            | Self::XlInpaint
-            | Self::Turbo => {
-                if use_f16 {
-                    "unet/diffusion_pytorch_model.fp16.safetensors"
-                } else {
-                    "unet/diffusion_pytorch_model.safetensors"
-                }
-            }
-        }
-    }
-
-    fn vae_file(&self, use_f16: bool) -> &'static str {
-        match self {
-            Self::V1_5
-            | Self::V1_5Inpaint
-            | Self::V2_1
-            | Self::V2Inpaint
-            | Self::Xl
-            | Self::XlInpaint
-            | Self::Turbo => {
-                if use_f16 {
-                    "vae/diffusion_pytorch_model.fp16.safetensors"
-                } else {
-                    "vae/diffusion_pytorch_model.safetensors"
-                }
-            }
-        }
-    }
-
-    fn clip_file(&self, use_f16: bool) -> &'static str {
-        match self {
-            Self::V1_5
-            | Self::V1_5Inpaint
-            | Self::V2_1
-            | Self::V2Inpaint
-            | Self::Xl
-            | Self::XlInpaint
-            | Self::Turbo => {
-                if use_f16 {
-                    "text_encoder/model.fp16.safetensors"
-                } else {
-                    "text_encoder/model.safetensors"
-                }
-            }
-        }
-    }
-
-    fn clip2_file(&self, use_f16: bool) -> &'static str {
-        match self {
-            Self::V1_5
-            | Self::V1_5Inpaint
-            | Self::V2_1
-            | Self::V2Inpaint
-            | Self::Xl
-            | Self::XlInpaint
-            | Self::Turbo => {
-                if use_f16 {
-                    "text_encoder_2/model.fp16.safetensors"
-                } else {
-                    "text_encoder_2/model.safetensors"
-                }
-            }
-        }
-    }
-}
-
-impl ModelFile {
-    pub fn get(
-        &self,
-        filename: Option<String>,
-        version: StableDiffusionVersion,
-        use_f16: bool,
-    ) -> Result<std::path::PathBuf> {
-        match filename {
-            Some(filename) => Ok(std::path::PathBuf::from(filename)),
-            None => {
-                let (repo, path) = match self {
-                    Self::Tokenizer => {
-                        let tokenizer_repo = match version {
-                            StableDiffusionVersion::V1_5
-                            | StableDiffusionVersion::V2_1
-                            | StableDiffusionVersion::V1_5Inpaint
-                            | StableDiffusionVersion::V2Inpaint => "openai/clip-vit-base-patch32",
-                            StableDiffusionVersion::Xl
-                            | StableDiffusionVersion::XlInpaint
-                            | StableDiffusionVersion::Turbo => {
-                                // This seems similar to the patch32 version except some very small
-                                // difference in the split regex.
-                                "openai/clip-vit-large-patch14"
-                            }
-                        };
-                        (tokenizer_repo, "tokenizer.json")
-                    }
-                    Self::Tokenizer2 => {
-                        ("laion/CLIP-ViT-bigG-14-laion2B-39B-b160k", "tokenizer.json")
-                    }
-                    Self::Clip => (version.repo(), version.clip_file(use_f16)),
-                    Self::Clip2 => (version.repo(), version.clip2_file(use_f16)),
-                    Self::UnetLcm => (
-                        "SimianLuo/LCM_Dreamshaper_v7",
-                        if use_f16 {
-                            "unet/diffusion_pytorch_model.fp16.safetensors"
-                        } else {
-                            "unet/diffusion_pytorch_model.safetensors"
-                        },
-                    ),
-                    Self::Unet => (version.repo(), version.unet_file(use_f16)),
-                    Self::Vae => {
-                        // Override for SDXL when using f16 weights.
-                        // See https://github.com/huggingface/candle/issues/1060
-                        if matches!(
-                            version,
-                            StableDiffusionVersion::Xl | StableDiffusionVersion::Turbo,
-                        ) && use_f16
-                        {
-                            (
-                                "madebyollin/sdxl-vae-fp16-fix",
-                                "diffusion_pytorch_model.safetensors",
-                            )
-                        } else {
-                            (version.repo(), version.vae_file(use_f16))
-                        }
-                    }
-                };
-                let filename = Api::new()?.model(repo.to_string()).get(path)?;
-                Ok(filename)
-            }
-        }
-    }
-}

crates/desktop/ml_models/src/ml/stable_diffusion/infer_clip_text_embeddings.rs

@@ -1,158 +0,0 @@
-#[cfg(feature = "accelerate")]
-extern crate accelerate_src;
-#[cfg(feature = "mkl")]
-extern crate intel_mkl_src;
-
-use candle_transformers::models::stable_diffusion;
-use std::path::{Path, PathBuf};
-
-use crate::ml::model_file::{ModelFile, StableDiffusionVersion};
-use ml_weights_registry::weights_registry::weights::{CLIP_JSON, SDXL_TURBO_CLIP_TEXT_ENCODER, SDXL_TURBO_CLIP_TEXT_ENCODER_2};
-use anyhow::Error as E;
-use candle_core::{DType, Device, Module, Tensor, D};
-use log::info;
-use tokenizers::Tokenizer;
-
-
-pub struct ClipArgs<'a, P: AsRef<Path>,  Q: AsRef<Path>> {
-  pub prompt: &'a str, 
-  pub uncond_prompt: &'a str, 
-  pub tokenizer: Option<String>, 
-  pub clip_weights: Option<String>, 
-  pub clip2_weights: Option<String>, 
-  pub sd_version: StableDiffusionVersion, 
-  pub sd_config: &'a stable_diffusion::StableDiffusionConfig, 
-  pub use_f16: bool, 
-  pub device: &'a Device, 
-  pub dtype: DType, 
-  pub use_guide_scale: bool, 
-  pub clip_json_path: P,
-  pub clip_weights_path: Q,
-}
-
-#[allow(clippy::too_many_arguments)]
-pub fn infer_clip_text_embeddings<P: AsRef<Path>, Q: AsRef<Path>>(args: ClipArgs<'_, P, Q>) -> anyhow::Result<Tensor> {
-  let ClipArgs {
-    prompt, 
-    uncond_prompt, 
-    tokenizer, 
-    clip_weights, 
-    clip2_weights, 
-    sd_version, 
-    sd_config, 
-    use_f16, 
-    device, 
-    dtype, 
-    use_guide_scale, 
-    clip_json_path,
-    clip_weights_path,
-  } = args;
-
-  info!("Preparing text embeddings... for {:?}", sd_version);
-
-  let which = match sd_version {
-    StableDiffusionVersion::Xl | StableDiffusionVersion::XlInpaint | StableDiffusionVersion::Turbo => vec![true, false],
-    _ => vec![true],
-  };
-
-  let text_embeddings = which
-    .iter()
-    .map(|first| {
-      do_infer_clip_text_embeddings(
-        &prompt,
-        &uncond_prompt,
-        None, // tokenizer
-        clip_weights.clone(),
-        None, // clip2_weights
-        sd_version,
-        &sd_config,
-        false, // use_f16
-        &device,
-        dtype,
-        use_guide_scale,
-        *first,
-        clip_json_path.as_ref(),
-        clip_weights_path.as_ref(),
-      )
-    })
-    .collect::<anyhow::Result<Vec<_>>>()?;
-
-  let text_embeddings = Tensor::cat(&text_embeddings, D::Minus1)?;
-
-  Ok(text_embeddings)
-}
-
-#[allow(clippy::too_many_arguments)]
-fn do_infer_clip_text_embeddings(
-  prompt: &str, 
-  uncond_prompt: &str, 
-  tokenizer: Option<String>, 
-  clip_weights: Option<String>, 
-  clip2_weights: Option<String>, 
-  sd_version: StableDiffusionVersion, 
-  sd_config: &stable_diffusion::StableDiffusionConfig, 
-  use_f16: bool, 
-  device: &Device, 
-  dtype: DType, 
-  use_guide_scale: bool, 
-  first: bool, 
-  clip_json_path: &Path,
-  clip_weights_path: &Path,
-) -> anyhow::Result<Tensor> {
-  info!("do_infer_clip_text_embeddings called with args {:?}", (prompt, uncond_prompt, tokenizer.clone().unwrap_or_else(|| "None".to_string()), clip_weights.clone().unwrap_or_else(|| "None".to_string()), clip2_weights.clone().unwrap_or_else(|| "None".to_string()), sd_version, sd_config, use_f16, device, dtype, use_guide_scale, first));
-
-  let tokenizer = clip_json_path;
-
-  info!("Loading Clip Tokenizer path: {:?}", tokenizer);
-
-  let tokenizer = Tokenizer::from_file(tokenizer).map_err(E::msg)?;
-
-  let pad_id = match &sd_config.clip.pad_with {
-    Some(padding) => *tokenizer.get_vocab(true).get(padding.as_str()).unwrap(),
-    None => *tokenizer.get_vocab(true).get("<|endoftext|>").unwrap(),
-  };
-
-  println!("Running with prompt \"{prompt}\".");
-  let mut tokens = tokenizer.encode(prompt, true).map_err(E::msg)?.get_ids().to_vec();
-  if tokens.len() > sd_config.clip.max_position_embeddings {
-    anyhow::bail!("the prompt is too long, {} > max-tokens ({})", tokens.len(), sd_config.clip.max_position_embeddings)
-  }
-  while tokens.len() < sd_config.clip.max_position_embeddings {
-    tokens.push(pad_id)
-  }
-  let tokens = Tensor::new(tokens.as_slice(), device)?.unsqueeze(0)?;
-
-  println!("Building the Clip transformer.");
-  
-  // TODO(bt,2025-03-23): These used to be two different clip models. We should only need one.
-  let clip_weights = if first {
-    println!(">>>>> CLIP 1");
-    clip_weights_path
-  } else {
-    println!(">>>>> CLIP 2");
-    clip_weights_path
-  };
-  
-  let clip_config = if first { &sd_config.clip } else { sd_config.clip2.as_ref().unwrap() };
-  let text_model = stable_diffusion::build_clip_transformer(clip_config, clip_weights, device, DType::F32)?;
-
-  let text_embeddings = text_model.forward(&tokens)?;
-
-  let text_embeddings = if use_guide_scale {
-    let mut uncond_tokens = tokenizer.encode(uncond_prompt, true).map_err(E::msg)?.get_ids().to_vec();
-    if uncond_tokens.len() > sd_config.clip.max_position_embeddings {
-      anyhow::bail!("the negative prompt is too long, {} > max-tokens ({})", uncond_tokens.len(), sd_config.clip.max_position_embeddings)
-    }
-    while uncond_tokens.len() < sd_config.clip.max_position_embeddings {
-      uncond_tokens.push(pad_id)
-    }
-
-    let uncond_tokens = Tensor::new(uncond_tokens.as_slice(), device)?.unsqueeze(0)?;
-    let uncond_embeddings = text_model.forward(&uncond_tokens)?;
-
-    Tensor::cat(&[uncond_embeddings, text_embeddings], 0)?.to_dtype(dtype)?
-  } else {
-    text_embeddings.to_dtype(dtype)?
-  };
-  Ok(text_embeddings)
-}

crates/desktop/ml_models/src/ml/stable_diffusion/lcm_pipeline.rs

@@ -1,281 +0,0 @@
-use std::path::{Path, PathBuf};
-use std::sync::{Arc, RwLock, RwLockReadGuard, RwLockWriteGuard};
-
-use anyhow::{anyhow, Error as E, Result};
-use candle_core::{DType, Device, IndexOp, Tensor, D};
-use candle_nn::VarBuilder;
-use candle_transformers::models::stable_diffusion::StableDiffusionConfig;
-use candle_transformers::models::stable_diffusion::lcm::{LCMScheduler, LCMSchedulerConfig};
-use candle_transformers::models::stable_diffusion::unet_2d::BlockConfig;
-use candle_transformers::models::stable_diffusion::unet_2d::UNet2DConditionModel;
-use candle_transformers::models::stable_diffusion::unet_2d::UNet2DConditionModelConfig;
-use candle_transformers::models::stable_diffusion::vae::{AutoEncoderKL, DiagonalGaussianDistribution};
-use crate::ml::image::dynamic_image_to_tensor::dynamic_image_to_tensor;
-use crate::ml::image::tensor_to_image_buffer::{tensor_to_image_buffer, RgbImage};
-use crate::ml::model_cache::ModelCache;
-use crate::ml::model_config::ModelConfig;
-use crate::ml::model_file::{ModelFile, StableDiffusionVersion};
-use crate::ml::prompt_cache::PromptCache;
-use crate::ml::stable_diffusion::get_vae_scale::get_vae_scale;
-use crate::ml::stable_diffusion::infer_clip_text_embeddings::{infer_clip_text_embeddings, ClipArgs};
-use crate::ml::stable_diffusion::remap_lcm_strength_range::remap_lcm_strength_range;
-use image::DynamicImage;
-use log::info;
-use ml_weights_registry::weights_registry::weights::{LYKON_DEAMSHAPER_7_TEXT_ENCODER_FP16, LYKON_DEAMSHAPER_7_VAE, SIMIANLUO_LCM_DREAMSHAPER_V7_UNET};
-use rand::Rng;
-
-const DEFAULT_STRENGTH : f64 = 75.0;
-
-pub struct Args<'a, P1: AsRef<Path>, P2: AsRef<Path>, P3: AsRef<Path>, P4: AsRef<Path>> {
-  pub image: &'a DynamicImage,
-  pub prompt: String,
-  pub uncond_prompt: String,
-  pub cfg_scale: Option<f64>,
-  pub img2img_strength: Option<f64>,
-  pub model_cache: &'a ModelCache,
-  pub prompt_cache: &'a PromptCache,
-  pub use_flash_attn: bool,
-  pub model_config: &'a ModelConfig,
-  pub maybe_seed: Option<u64>,
-  pub scheduler_steps: usize,
-  pub vae_path: P1,
-  pub unet_path: P2,
-  pub clip_json_path: P3,
-  pub clip_weights_path: P4,
-}
-
-pub fn lcm_pipeline<P1, P2, P3, P4>(args: Args<'_, P1, P2, P3, P4>) -> Result<RgbImage> 
-  where P1: AsRef<Path>, P2: AsRef<Path>, P3: AsRef<Path>, P4: AsRef<Path>
-{
-  let Args { 
-    prompt, 
-    uncond_prompt, 
-    cfg_scale,
-    img2img_strength, 
-    model_cache, 
-    prompt_cache, 
-    image, 
-    use_flash_attn ,
-    model_config: ModelConfig {
-      device,
-      dtype,
-      sd_version,
-      sd_config,
-      hf_api,
-    },
-    maybe_seed,
-    scheduler_steps,
-    vae_path,
-    unet_path,
-    clip_json_path,
-    clip_weights_path,
-  } = args;
-  
-  let dtype = *dtype;
-  let sd_version = *sd_version;
-
-  let img2img_strength = img2img_strength
-    .map(|s| remap_lcm_strength_range(s))
-    .unwrap_or(DEFAULT_STRENGTH);
-  
-  let img2img_strength = img2img_strength / 100.0;
-
-  let use_f16 = true;
-
-  // Use LCM Scheduler instead of Euler Ancestral for better speed and quality
-  let mut scheduler = LCMScheduler::new(scheduler_steps, img2img_strength, LCMSchedulerConfig::default())?;
-
-  let seed = maybe_seed.unwrap_or_else(|| rand::thread_rng().gen());
-
-  let seed = 42;
-  
-  device.set_seed(seed)?;
-
-  let guidance_scale = match cfg_scale {
-    Some(guidance_scale) => guidance_scale,
-    None => match sd_version {
-      StableDiffusionVersion::V1_5 | StableDiffusionVersion::V2_1 | StableDiffusionVersion::Xl => 7.5,
-      StableDiffusionVersion::Turbo => 0.,
-      _ => 0., // NB: Not sure what the other model families should use, so sticking with "0"
-    },
-  };
-
-  // let use_guide_scale = guidance_scale > 1.0;
-  let use_guide_scale = false;
-
-  let maybe_cached = prompt_cache.get_copy(&prompt)?;
-
-  let mut text_embeddings = if let Some(tensor) = maybe_cached {
-    tensor
-  } else {
-    info!("Prompt is NOT cached! Calculating embedding...");
-    let tensor = infer_clip_text_embeddings(ClipArgs {
-      prompt: & prompt,
-      uncond_prompt: &uncond_prompt,
-      tokenizer: None, // tokenizer
-      clip_weights: None,
-      clip2_weights: None, // clip2_weights
-      sd_version,
-      sd_config: &sd_config,
-      use_f16, // use_f16
-      device,
-      dtype,
-      use_guide_scale,
-      clip_json_path,
-      clip_weights_path,
-    })?;
-
-    prompt_cache.store_copy(&prompt, &tensor)?;
-    tensor
-  };
-
-  info!("Text embeddings shape: {:?}", text_embeddings.shape());
-
-  info!("Loading input image into tensor...");
-
-  let input_image = dynamic_image_to_tensor(image, &device, dtype)?;
-
-  info!("Reference image shape: {:?}", input_image.shape());
-
-  let vae_scale = get_vae_scale(sd_version);
-
-  let maybe_vae = model_cache.get_vae()?;
-
-  let vae = match maybe_vae {
-    Some(vae) => vae,
-    None => {
-      info!("Building VAE model from file {:?}...", vae_path.as_ref());
-      
-      let vae = sd_config.build_vae(vae_path, &device, dtype)?;
-
-      let vae = Arc::new(vae);
-
-      model_cache.set_vae(vae.clone())?;
-
-      vae
-    }
-  };
-
-  let maybe_unet = model_cache.get_unet()?;
-
-  let unet = match maybe_unet {
-    Some(unet) => unet,
-    None => {
-      info!("No unet found in cache; loading...");
-
-      let in_channels = match sd_version {
-        StableDiffusionVersion::XlInpaint | StableDiffusionVersion::V2Inpaint | StableDiffusionVersion::V1_5Inpaint => 9,
-        _ => 4,
-      };
-
-      // For LCM models, ensure that:
-      // 1. We're using the correct prediction type (epsilon)
-      // 2. The cross-attention dimension is correct (768 for Dreamshaper)
-      // 3. The in_channels matches the expected input (4 for standard, 9 for inpainting)
-      let unet_config = UNet2DConditionModelConfig {
-        // LCM specific config values
-        cross_attention_dim: 768,      // Dreamshaper v7 uses 768
-        time_cond_proj_dim: Some(256), // LCM models use 256-dimensional guidance embeddings
-        // Rest of configuration...
-        center_input_sample: false,
-        flip_sin_to_cos: true,
-        freq_shift: 0.0,
-        blocks: vec![
-          BlockConfig { out_channels: 320, use_cross_attn: Some(1), attention_head_dim: 8 },
-          BlockConfig { out_channels: 640, use_cross_attn: Some(1), attention_head_dim: 8 },
-          BlockConfig { out_channels: 1280, use_cross_attn: Some(1), attention_head_dim: 8 },
-          BlockConfig { out_channels: 1280, use_cross_attn: None, attention_head_dim: 8 }
-        ],
-        layers_per_block: 2,
-        downsample_padding: 1,
-        mid_block_scale_factor: 1.0,
-        norm_num_groups: 32,
-        norm_eps: 1e-5,
-        sliced_attention_size: None,
-        use_linear_projection: false,
-        time_embed_dim: Some(1280),
-        // time_embed_dim: Some(256),
-      };
-
-      // Load the model directly from the safetensors file
-      let vs_unet = unsafe { VarBuilder::from_mmaped_safetensors(&[unet_path], dtype, &device)? };
-
-      let unet = UNet2DConditionModel::new(vs_unet, in_channels, 4, use_flash_attn, unet_config)?;
-
-      let unet = Arc::new(unet);
-
-      model_cache.set_unet(unet.clone())?;
-
-      unet
-    },
-  };
-
-  let init_latent_dist: DiagonalGaussianDistribution = vae.encode(&input_image)?;
-
-  // Create LCM guidance scale embeddings
-  let embedding_dim = 1280;
-  let guidance_scale_embedding = scheduler.get_guidance_scale_embedding(guidance_scale, embedding_dim, &device, dtype)?;
-
-  // Generate latents from input image using the LCM approach
-  info!("Generating latents from input image...");
-  let init_latents = (init_latent_dist.sample()? * vae_scale)?;
-  info!("Initial latents shape: {:?}", init_latents.shape());
-
-  // Calculate timesteps for LCM with proper img2img handling
-  let timesteps = LCMScheduler::get_timesteps_for_steps(scheduler_steps, img2img_strength);
-  let t_start = timesteps[0]; // First denoising step
-
-  // Add noise at the appropriate timestep for img2img
-  info!("Adding noise to latents for t_start={}", t_start);
-  let noise = init_latents.randn_like(0f64, 1f64)?;
-  let latents = scheduler.add_noise(&init_latents, noise, t_start)?;
-
-  let mut latents = latents;
-  let timesteps: Vec<_> = scheduler.timesteps().iter().copied().collect();
-
-  for (timestep_index, &timestep) in timesteps.iter().enumerate() {
-    info!("Processing step {}/{} @ timestamp = {}", timestep_index + 1, timesteps.len(), timestep);
-
-    let latent_model_input = if use_guide_scale { Tensor::cat(&[&latents, &latents], 0)? } else { latents.clone() };
-
-    let latent_model_input = scheduler.scale_model_input(latent_model_input, timestep);
-
-    // Use the guidance scale embedding in the UNet inference with the proper method name
-    let mut noise_pred = match unet.forward_with_guidance(&latent_model_input, timestep as f64, &text_embeddings, Some(&guidance_scale_embedding)) {
-      Ok(pred) => pred,
-      Err(e) => {
-        info!("UNet inference failed with error: {}", e);
-        return Err(anyhow::anyhow!("UNet inference failed: {}", e));
-      },
-    };
-
-    if use_guide_scale {
-      let chunks = noise_pred.chunk(2, 0)?;
-      let (noise_pred_uncond, noise_pred_text) = (&chunks[0], &chunks[1]);
-      noise_pred = (noise_pred_uncond + (guidance_scale * (noise_pred_text - noise_pred_uncond)?)?)?;
-    }
-
-    // Apply the LCM scheduler step
-    latents = scheduler.step(&noise_pred, timestep, &latents, timestep_index, scheduler_steps)?;
-  }
-
-  info!("Diffusion process completed, decoding image...");
-  let image = vae.decode(&(latents / vae_scale)?)?;
-
-  info!("VAE decode completed");
-
-  info!("Scaling image back");
-  let image = ((image / 2.)? + 0.5)?;
-
-  info!("Normalized image values");
-  let image = (image.clamp(0f32, 1.)? * 255.)?;
-
-  info!("Convert to int8");
-  let image = image.to_dtype(DType::U8)?;
-
-  info!("Converted to 8-bit format");
-
-  let image = tensor_to_image_buffer(&image.i(0)?)?;
-
-  Ok(image)
-}

crates/desktop/ml_models/src/ml/stable_diffusion/mod.rs

@@ -1,5 +0,0 @@
-pub mod get_vae_scale;
-pub mod infer_clip_text_embeddings;
-pub mod stable_diffusion_pipeline;
-pub mod lcm_pipeline;
-mod remap_lcm_strength_range;

crates/desktop/ml_models/src/ml/stable_diffusion/remap_lcm_strength_range.rs

@@ -1,54 +0,0 @@
-use primitives::numerics::remap_range_f64::{remap_range_f64, Args};
-
-/// Our LCM implementation uses strength [13, 100]. Strength below 13 will take 8 steps, 
-/// so we need to limit ourselves to this range from an input range of [0, 100].
-pub fn remap_lcm_strength_range(value: f64) -> f64 {
-  remap_range_f64(Args {
-    old_min: 0.0,
-    old_max: 100.0,
-    new_min: 13.0,
-    new_max: 100.0,
-    value,
-  })
-}
-
-#[cfg(test)]
-mod test {
-  use crate::ml::stable_diffusion::remap_lcm_strength_range::remap_lcm_strength_range;
-  use assertor::{assert_that, FloatAssertion};
-
-  #[test]
-  fn old_max() {
-    assert_eq!(remap_lcm_strength_range(100.0), 100.0);
-  }
-  
-  #[test]
-  fn old_min() {
-    assert_eq!(remap_lcm_strength_range(0.0), 13.0);
-  }
-
-  #[test]
-  fn under_bounds() {
-    assert_eq!(remap_lcm_strength_range(-100.0), 13.0);
-  }
-  
-  #[test]
-  fn over_bounds() {
-    assert_eq!(remap_lcm_strength_range(200.0), 100.0);
-  }
-
-  #[test]
-  fn other_values() {
-    assert_that!(remap_lcm_strength_range(0.0)).with_abs_tol(0.001).is_approx_equal_to(13.000);
-    assert_that!(remap_lcm_strength_range(5.0)).with_abs_tol(0.001).is_approx_equal_to(17.350);
-    assert_that!(remap_lcm_strength_range(11.0)).with_abs_tol(0.001).is_approx_equal_to(22.570);
-    assert_that!(remap_lcm_strength_range(13.0)).with_abs_tol(0.001).is_approx_equal_to(24.310);
-    assert_that!(remap_lcm_strength_range(14.0)).with_abs_tol(0.001).is_approx_equal_to(25.180);
-    assert_that!(remap_lcm_strength_range(20.0)).with_abs_tol(0.001).is_approx_equal_to(30.400);
-    assert_that!(remap_lcm_strength_range(25.0)).with_abs_tol(0.001).is_approx_equal_to(34.750);
-    assert_that!(remap_lcm_strength_range(50.0)).with_abs_tol(0.001).is_approx_equal_to(56.500);
-    assert_that!(remap_lcm_strength_range(75.0)).with_abs_tol(0.001).is_approx_equal_to(78.250);
-    assert_that!(remap_lcm_strength_range(99.0)).with_abs_tol(0.001).is_approx_equal_to(99.130);
-    assert_that!(remap_lcm_strength_range(100.0)).with_abs_tol(0.001).is_approx_equal_to(100.000);
-  }
-}

crates/desktop/ml_weights_registry/Cargo.toml

@@ -1,18 +0,0 @@
-[package]
-name = "ml_weights_registry"
-edition = "2021"
-version = "0.0.1"
-authors = ["storyteller"]
-publish = false
-description = ""
-license = ""
-repository = ""
-
-[lib]
-name = "ml_weights_registry"
-path = "src/lib.rs"
-
-# External
-[dependencies]
-anyhow = "1.0"
-const_format = "0.2.34"

crates/desktop/ml_weights_registry/src/lib.rs

@@ -1 +0,0 @@
-pub mod weights_registry;

crates/desktop/ml_weights_registry/src/weights_registry/mod.rs

@@ -1,3 +0,0 @@
-mod weight_descriptor_builder;
-pub mod weight_descriptor;
-pub mod weights;

crates/desktop/ml_weights_registry/src/weights_registry/weight_descriptor.rs

@@ -1,45 +0,0 @@
-
-// TODO(bt,2025-03-13): When it's time for dynamic, non-const allocation, refer to
-//  http::header::name::HeaderName: https://github.com/hyperium/http/blob/master/src/header/name.rs
-//  as a pattern that will support both compile-time constants and dynamically allocated String-based variants.
-
-#[derive(Clone, Copy)]
-pub enum WeightFunction {
-  ImageSegmentation,
-  TextEncoder,
-  TextTokenizer,
-  Unet,
-  Vae,
-}
-
-pub struct WeightDescriptor {
-  /// Name of the weight, eg. "Lykon Dreamshaper 7"
-  pub name: &'static str,
-
-  /// Name of the file on the filesystem, eg. lykon_dreamshaper_7_text_encoder.fp16.safetensors
-  pub filename: &'static str,
-
-  /// URL of the file in the R2 bucket,
-  /// eg. https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/lykon_dreamshaper_7_text_encoder.fp16.safetensors
-  pub r2_download_url: &'static str,
-
-  pub function: WeightFunction,
-}
-
-impl WeightFunction {
-  pub fn as_descriptive_str(&self) -> &'static str {
-    match self {
-      WeightFunction::ImageSegmentation => "image segmentation",
-      WeightFunction::TextEncoder => "text encoder",
-      WeightFunction::TextTokenizer => "text tokenizer",
-      WeightFunction::Unet => "UNET",
-      WeightFunction::Vae => "VAE",
-    }
-  }
-}
-
-impl WeightDescriptor {
-  pub fn to_descriptive_name(&self) -> String {
-    format!("{} ({})", self.name, self.function.as_descriptive_str())
-  }
-}

crates/desktop/ml_weights_registry/src/weights_registry/weight_descriptor_builder.rs

@@ -1,17 +0,0 @@
-pub (super) const R2_BUCKET_URL : &str = "https://pub-bc5e2bc0cdee4bb5ae8fca9d641ca0d6.r2.dev/";
-
-// TODO(bt,2025-03-13): Try to use keyword args from the call site.
-// NB: We're using a macro because we can't use a builder pattern and then concat string literals
-//  and variables (even if they're known to be &'static) and keep allocations static.
-macro_rules! weight {
-    ($name: literal, $file_name:literal, $weight_function:expr) => {
-      $crate::weights_registry::weight_descriptor::WeightDescriptor {
-        name: $name,
-        filename: $file_name,
-        r2_download_url: const_format::concatcp!(crate::weights_registry::weight_descriptor_builder::R2_BUCKET_URL, $file_name),
-        function: $weight_function,
-      }
-    };
-}
-
-pub(crate) use weight;

crates/desktop/ml_weights_registry/src/weights_registry/weights.rs

@@ -1,65 +0,0 @@
-use crate::weights_registry::weight_descriptor::{WeightDescriptor, WeightFunction};
-use crate::weights_registry::weight_descriptor_builder::weight;
-
-pub const CLIP_JSON : WeightDescriptor = weight!(
-  "CLIP",
-  "clip_vit_bigg_14_lion2b_39b_b160k.tokenizer.json",
-  WeightFunction::TextTokenizer
-);
-
-/// For LCM-based image generation
-pub const LYKON_DEAMSHAPER_7_TEXT_ENCODER_FP16 : WeightDescriptor = weight!(
-  "Lykon Dreamshaper 7",
-  "lykon_dreamshaper_7_text_encoder.fp16.safetensors",
-  WeightFunction::TextEncoder
-);
-
-/// For LCM-based image generation
-pub const LYKON_DEAMSHAPER_7_VAE : WeightDescriptor = weight!(
-  "Lykon Dreamshaper 7",
-  "lykon_dreamshaper_7_vae.safetensors",
-  WeightFunction::Vae
-);
-
-/// For LCM-based image generation
-pub const SIMIANLUO_LCM_DREAMSHAPER_V7_UNET : WeightDescriptor = weight!(
-  "SimianLuo LCM Dreamshaper v7",
-  "simianluo_lcm_dreamshaper_v7_unet.safetensors",
-  WeightFunction::Unet
-);
-
-/// Image segmentation and background removal
-/// dis_medium.onnx
-pub const DIS_MEDIUM_ONNX : WeightDescriptor = weight!(
-  "Dichotomous Image Segmentation Medium",
-  "dis_medium.onnx",
-  WeightFunction::ImageSegmentation
-);
-
-#[deprecated]
-pub const SDXL_TURBO_CLIP_TEXT_ENCODER : WeightDescriptor = weight!(
-  "SDXL Turbo CLIP",
-  "clip_text_encoder.safetensors",
-  WeightFunction::TextEncoder
-);
-
-#[deprecated]
-pub const SDXL_TURBO_CLIP_TEXT_ENCODER_2 : WeightDescriptor = weight!(
-  "SDXL Turbo CLIP #2",
-  "clip_text_encoder_2.safetensors",
-  WeightFunction::TextEncoder
-);
-
-#[deprecated]
-pub const SDXL_TURBO_VAE : WeightDescriptor = weight!(
-  "SDXL Turbo",
-  "diffusion_pytorch_model.vae.safetensors",
-  WeightFunction::Vae
-);
-
-#[deprecated]
-pub const SDXL_TURBO_UNET : WeightDescriptor = weight!(
-  "SDXL Turbo",
-  "diffusion_pytorch_model.unet.safetensors",
-  WeightFunction::Unet
-);

crates/desktop/src-tauri/.taurignore

@@ -1,4 +0,0 @@
-# Don't restart the dev server on Rust changes
-crates/
-Cargo.toml
-Cargo.lock

crates/desktop/src-tauri/Cargo.toml

@@ -7,7 +7,6 @@ license = ""
 repository = ""
 edition = "2021"
 rust-version = "1.77.2"
-publish = false
 
 # See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
 
@@ -23,59 +22,35 @@ tauri-build = { version = "2.0.4", features = [] }
 # Internal
 filesys = { path = "../../lib/files/filesys" }
 memory_caching = { path = "../../lib/caching/memory_caching" }
-ml_weights_registry = { path = "../ml_weights_registry" }
-primitives = { path = "../../lib/data/primitives" }
 
-# Internal ML
-# TODO: Only on Windows and Linux
-#ml_models = { path = "../ml_models", optional = true }
+# Bypass cudarc issues on Windows https://github.com/coreylowman/cudarc/issues/332
+cudarc = "0.13.7"
 
 # External
-#openssl = { version = "0.10", features = ["vendored"] }
 anyhow = "1.0"
-base64 = "0.22.1"
 bytes = "1.10"
-const_format = "0.2.34"
-directories = "6.0.0"
-dirs = "6.0.0"
-futures = "0.3"
-half = { version = "2.3.1", features = ["num-traits", "use-intrinsics", "rand_distr"] }
+base64 = "0.22.1"
+candle-core = { version = "0.8.2", features = ["cuda"] }
+candle-nn = { version = "0.8.2", features = ["cuda"] }
+candle-transformers = { version = "0.8.2", features = ["cuda"] }
+hf-hub = "0.3.2"
 image = "0.25.5"
+serde_yaml = "0.9"
 log = "0.4"
 once_cell = "1.20.3"
-rand = "0.8.5"
-# NB(bt,2025-03-10): default-features = false on Windows while I debug openssl / rustls
-reqwest = { version = "0.12.12", features = ["stream"], default-features = false }
+rand = "0.8"
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1.0"
-serde_yaml = "0.9"
 tauri = { version = "2.2.4", features = [] }
 tauri-plugin-log = "2"
-tempdir = "0.3.7"
-tempfile = "3.18.0"
-tokio = { version = "1.42", features = ["rt", "rt-multi-thread", "fs"] }
-tokio-util = { version = "0.7", features = ["codec"] }
+tokenizers = "0.15.2"
 
-# tracing-subscriber = { version = "0.3", features = [ "env-filter", "fmt" ] }
-
-[target.'cfg(not(target_os = "macos"))'.dependencies]
-ml_models = { path = "../ml_models" }
-
-[dev-dependencies]
-libloading = "0.8.6"
-serial_test = "3.2.0"
-assertor = "0.0.3"
-
-#[target.'cfg(target_os = "macos")'.features]
-#default  = []
-#ml_models = ["dep:ml_models"]
-
-[profile.dev]
-incremental = true # Compile your binary in smaller steps.
-
-[profile.release]
-codegen-units = 1 # Allows LLVM to perform better optimization.
-lto = true # Enables link-time-optimizations.
-opt-level = 3 # `s` prioritizes small binary size. Use `3` if you prefer speed.
-panic = "abort" # Higher performance by disabling panic handlers.
-strip = true # Ensures debug symbols are removed.
+[features]
+default = ["cuda"]
+cuda = [
+  "candle-core/cuda",
+  # TODO(bt,2025-02-22): cuDNN might speed up Stable Diffusion
+  # "candle-core/cudnn",
+  "candle-nn/cuda",
+  "candle-transformers/cuda",
+]

crates/desktop/src-tauri/src/endpoints/download_models.rs

@@ -1,12 +0,0 @@
-use crate::state::app_dir::AppDataRoot;
-use tauri::State;
-
-#[tauri::command]
-pub async fn download_models(
-  _app_data_root: State<'_, AppDataRoot>,
-) -> Result<String, String> {
-
-  // NB: This endpoint has been removed for now as we now download models in the app lifecycle.
-
-  Ok(("downloaded".to_string()))
-}

crates/desktop/src-tauri/src/endpoints/image_endpoint.rs

@@ -0,0 +1,117 @@
+use crate::ml::model_cache::ModelCache;
+use crate::ml::prompt_cache::PromptCache;
+use crate::ml::stable_diffusion::stable_diffusion_pipeline::{stable_diffusion_pipeline, Args};
+use crate::state::app_config::AppConfig;
+use base64::{engine::general_purpose::STANDARD as BASE64_STANDARD, Engine};
+use bytes::BytesMut;
+use image::imageops::FilterType;
+use image::{DynamicImage, ImageFormat, ImageReader};
+use log::{error, info};
+use std::io::Cursor;
+use std::path::PathBuf;
+use tauri::State;
+
+const PROMPT_FILENAME : &str = "prompt.txt";
+const PROMPT: &str = "A beautiful landscape with mountains and a lake";
+
+const NEGATIVE_PROMPT: &str = "bad quality, bad faces, poor quality, blurry faces, watermark";
+
+/// This handler takes an image (as a base64 encoded string) and a prompt and returns
+/// an image (as a base64-encoded string).
+#[tauri::command]
+pub fn infer_image(
+  image: &str,
+  prompt: Option<String>,
+  strength: Option<u8>,
+  model_config: State<AppConfig>,
+  model_cache: State<ModelCache>,
+  prompt_cache: State<PromptCache>,
+) -> Result<String, String> {
+
+  let prompt = get_prompt_or_fallback(prompt);
+
+  info!("Prompt: {}", prompt);
+
+  let image = hydrate_base64_image(image)
+    .map_err(|err| format!("Couldn't hydrate image from base64: {}", err))?;
+
+  let result = do_infer_image(&prompt, image, strength, &model_config, &model_cache, prompt_cache);
+  
+  if let Err(err) = result.as_deref() {
+    error!("There was an error: {:?}", err);
+  }
+  
+  result
+}
+
+fn do_infer_image(
+  prompt: &str,
+  image: DynamicImage,
+  strength: Option<u8>,
+  config: &AppConfig,
+  model_cache: &ModelCache,
+  prompt_cache: State<PromptCache>,
+) -> Result<String, String> {
+  println!("infer_image called; generating image with SDXL Turbo...");
+  
+  let args = Args {
+    image: &image,
+    prompt: prompt.to_string(),
+    uncond_prompt: NEGATIVE_PROMPT.to_string(),
+    //guidance_scale: Some(0.0),
+    model_cache,
+    configs: config,
+    prompt_cache: &prompt_cache,
+    i2i_strength: strength,
+    cfg_scale: config.cfg_scale,
+  };
+
+  match stable_diffusion_pipeline(args) {
+    Ok(image) => {
+      let mut bytes = Vec::with_capacity(1024*1024);
+      
+      image.write_to(&mut Cursor::new(&mut bytes), ImageFormat::Png)
+        .map_err(|err| format!("failure to encode image: {:?}", err))?;
+
+      let bytes = BASE64_STANDARD.encode(bytes);
+
+      Ok(bytes)
+    },
+    Err(e) => Err(format!("Failed to generate image: {}", e)),
+  }
+}
+
+fn get_prompt_or_fallback(user_prompt: Option<String>) -> String {
+  let user_prompt = user_prompt.map(|prompt| prompt.trim().to_string())
+    .filter(|prompt| !prompt.is_empty());
+
+  if let Some(prompt) = user_prompt {
+    return prompt;
+  }
+
+  let prompt_file = PathBuf::from(PROMPT_FILENAME)
+    .canonicalize()
+    .unwrap_or_else(|_| PathBuf::from(PROMPT_FILENAME));
+
+  std::fs::read_to_string(&prompt_file)
+    .map_err(|err| format!("Failed to read prompt file: {}", err))
+    .unwrap_or_else(|_| {
+      error!("Failed to read prompt file: {:?}", prompt_file);
+      PROMPT.to_string()
+    })
+    .trim()
+    .to_string()
+}
+
+fn hydrate_base64_image(base64_image: &str) -> anyhow::Result<DynamicImage> {
+  let bytes = BASE64_STANDARD.decode(base64_image)?;
+
+  let image = ImageReader::new(Cursor::new(bytes))
+    .with_guessed_format()?
+    .decode()?;
+
+  // TODO(bt,2025-02-17): Running out of vram with full image buffer size
+  let image = image.resize(512, 512, FilterType::CatmullRom);
+
+  Ok(image)
+}

crates/desktop/src-tauri/src/endpoints/inpaint_image_endpoint.rs

@@ -1,24 +0,0 @@
-use crate::ml::model_cache::ModelCache;
-use crate::ml::prompt_cache::PromptCache;
-use crate::state::app_config::AppConfig;
-use crate::state::app_dir::AppDataRoot;
-use image::{ImageReader, RgbImage};
-use log::info;
-use tauri::{AppHandle, State};
-
-
-#[tauri::command]
-pub async fn inpaint_image(
-  image: &str,
-  mask: &str,
-  prompt: String,
-  model_config: State<'_, AppConfig>,
-  model_cache: State<'_, ModelCache>,
-  prompt_cache: State<'_, PromptCache>,
-  app_data_root: State<'_, AppDataRoot>,
-  app: AppHandle,
-) -> Result<String, String> {
-  info!("inpaint_endpoint endpoint called.");
-
-  Ok(mask.to_string())
-}

crates/desktop/src-tauri/src/endpoints/mod.rs

@@ -1,6 +1,3 @@
-pub mod download_models;
 pub mod flip_image;
-pub mod inpaint_image_endpoint;
-pub mod realtime_image_endpoint;
-pub mod remove_background_endpoint;
-pub mod text_to_image_endpoint;
+pub mod image_endpoint;
+pub mod test_counter;

crates/desktop/src-tauri/src/endpoints/realtime_image_endpoint.rs

@@ -1,143 +0,0 @@
-#[cfg(not(target_os = "macos"))]
-use ml_models::ml::stable_diffusion::lcm_pipeline::{lcm_pipeline, Args};
-
-use crate::state::app_config::AppConfig;
-use crate::state::app_dir::AppDataRoot;
-use crate::stubs::model_cache::ModelCache;
-use crate::stubs::prompt_cache::PromptCache;
-use crate::utils::image::decode_base64_image::decode_base64_image;
-use crate::utils::image::encode_dynamic_image_base64_png::encode_dynamic_image_base64_png;
-use crate::utils::image::encode_rgb_image_base64_png::encode_rgb_image_base64_png;
-use image::imageops::FilterType;
-use image::{DynamicImage, RgbImage};
-use log::{error, info};
-use ml_weights_registry::weights_registry::weights::{CLIP_JSON, LYKON_DEAMSHAPER_7_TEXT_ENCODER_FP16, LYKON_DEAMSHAPER_7_VAE, SDXL_TURBO_CLIP_TEXT_ENCODER, SIMIANLUO_LCM_DREAMSHAPER_V7_UNET};
-use std::path::PathBuf;
-use tauri::{AppHandle, State};
-
-
-const PROMPT_FILENAME : &str = "prompt.txt";
-const PROMPT: &str = "shiba inu, cute dog, detailed, walking in a wooded forest, photorealistic, 8k";
-
-const NEGATIVE_PROMPT: &str = "bad quality, bad faces, poor quality, blurry faces, watermark";
-
-const RANDOM_SEED: u32 = 42;
-
-
-/// This handler takes an image (as a base64 encoded string) and a prompt and returns
-/// an image (as a base64-encoded string).
-#[tauri::command]
-pub async fn infer_image(
-  image: &str,
-  prompt: Option<String>,
-  strength: Option<f64>,
-  model_config: State<'_, AppConfig>,
-  model_cache: State<'_, ModelCache>,
-  prompt_cache: State<'_, PromptCache>,
-  app_data_root: State<'_, AppDataRoot>,
-  app: AppHandle,
-) -> Result<String, String> {
-  info!("infer_image endpoint called.");
-
-  let prompt = get_prompt_or_fallback(prompt);
-
-  info!("Strength: {:?}; Prompt: {}", strength, prompt);
-
-  let image = decode_base64_image(image)
-    .map_err(|err| format!("Couldn't hydrate image from base64: {}", err))?;
-  
-  // TODO(bt,2025-02-17): Running out of vram with full image buffer size
-  let mut image = image.resize(512, 512, FilterType::CatmullRom);
-
-  #[cfg(not(target_os = "macos"))]
-  {
-    let result = infer_image_impl(
-      &prompt,
-      image,
-      strength,
-      &model_config,
-      &model_cache,
-      prompt_cache,
-      &app_data_root
-    ).await;
-
-    image = match result {
-      Ok(img) => DynamicImage::ImageRgb8(img),
-      Err(err) => {
-        error!("There was an error generating the image: {:?}", err);
-        return Err(format!("There was an error generating the image: {}", err));
-      }
-    };
-  }
-
-  let bytes = encode_dynamic_image_base64_png(image)
-    .map_err(|err| format!("failure to encode image: {:?}", err))?;
-
-  info!("Inference successful; image converted to base64, serving back to browser...");
-
-  Ok(bytes)
-}
-
-#[cfg(not(target_os = "macos"))]
-async fn infer_image_impl(
-  prompt: &str,
-  image: DynamicImage,
-  strength: Option<f64>,
-  config: &AppConfig,
-  model_cache: &ModelCache,
-  prompt_cache: State<'_, PromptCache>,
-  app_data_root: &AppDataRoot,
-) -> Result<RgbImage, String> {
-  
-  let weights_dir = app_data_root.weights_dir();
-  let vae_path = weights_dir.weight_path(&LYKON_DEAMSHAPER_7_VAE);
-  let unet_path = weights_dir.weight_path(&SIMIANLUO_LCM_DREAMSHAPER_V7_UNET);
-  let clip_json_path = weights_dir.weight_path(&CLIP_JSON);
-  //let clip_weights_path= weights_dir.weight_path(&LYKON_DEAMSHAPER_7_TEXT_ENCODER_FP16);
-  let clip_weights_path= weights_dir.weight_path(&SDXL_TURBO_CLIP_TEXT_ENCODER);
-
-  let args = Args {
-    image: &image,
-    prompt: prompt.to_string(),
-    uncond_prompt: NEGATIVE_PROMPT.to_string(),
-    model_cache,
-    prompt_cache: &prompt_cache,
-    img2img_strength: strength,
-    cfg_scale: config.cfg_scale,
-    use_flash_attn: true,
-    model_config: &config.model_config,
-    maybe_seed: None,
-    scheduler_steps: config.scheduler_steps,
-    vae_path: &vae_path,
-    unet_path: &unet_path,
-    clip_json_path: &clip_json_path,
-    clip_weights_path: &clip_weights_path,
-  };
-
-  let image = lcm_pipeline(args)
-    .map_err(|err| format!("failure to encode image: {:?}", err))?;
-
-  Ok(image)
-}
-
-fn get_prompt_or_fallback(user_prompt: Option<String>) -> String {
-  let user_prompt = user_prompt.map(|prompt| prompt.trim().to_string())
-    .filter(|prompt| !prompt.is_empty());
-
-  if let Some(prompt) = user_prompt {
-    return prompt;
-  }
-
-  let prompt_file = PathBuf::from(PROMPT_FILENAME)
-    .canonicalize()
-    .unwrap_or_else(|_| PathBuf::from(PROMPT_FILENAME));
-
-  std::fs::read_to_string(&prompt_file)
-    .map_err(|err| format!("Failed to read prompt file: {}", err))
-    .unwrap_or_else(|_| {
-      error!("Failed to read prompt file: {:?}", prompt_file);
-      PROMPT.to_string()
-    })
-    .trim()
-    .to_string()
-}

crates/desktop/src-tauri/src/endpoints/remove_background_endpoint.rs

@@ -1,45 +0,0 @@
-//! Adapted from https://github.com/dnanhkhoa/rust-background-removal
-
-#[cfg(not(target_os = "macos"))]
-use ml_models::ml::background_removal::remove_image_background::remove_image_background;
-
-use crate::state::app_dir::AppDataRoot;
-use crate::utils::image::decode_base64_image::decode_base64_image;
-use crate::utils::image::encode_dynamic_image_base64_png::encode_dynamic_image_base64_png;
-use image::{imageops, DynamicImage};
-use log::info;
-use ml_weights_registry::weights_registry::weights::DIS_MEDIUM_ONNX;
-use tauri::State;
-
-/// This handler removes the background from an image.
-#[tauri::command]
-pub async fn remove_background(
-  image: &str,
-  app_data_root: State<'_, AppDataRoot>,
-) -> Result<String, String> {
-  info!("remove_background endpoint called.");
-
-  let mut image = decode_base64_image(image)
-      .map_err(|err| format!("Couldn't hydrate image from base64: {}", err))?;
-
-  #[cfg(not(target_os = "macos"))]
-  {
-    image = remove_background_impl(&app_data_root, image).await?;
-  }
-
-  let image = encode_dynamic_image_base64_png(image)
-    .map_err(|err| format!("failure to encode image: {:?}", err))?;
-
-  Ok(image)
-}
-
-#[cfg(not(target_os = "macos"))]
-async fn remove_background_impl(app_data_root: &AppDataRoot, image: DynamicImage) -> Result<DynamicImage, String> {
-  let model_path = app_data_root.weights_dir().weight_path(&DIS_MEDIUM_ONNX);
-
-  let image = remove_image_background(model_path, image)
-      .await
-      .map_err(|err| format!("Couldn't remove image from image: {}", err))?;
-
-  Ok(image)
-}

crates/desktop/src-tauri/src/endpoints/test_counter.rs

@@ -0,0 +1,24 @@
+use std::sync::{Arc, RwLock};
+use once_cell::sync::Lazy;
+
+static COUNTER : Lazy<Arc<RwLock<u64>>> = Lazy::new(|| Arc::new(RwLock::new(0)));
+
+#[tauri::command]
+pub fn test_counter() -> u64 {
+  let value : u64;
+  {
+    match COUNTER.write() {
+      Ok(mut counter) => {
+        *counter += 1;
+        value = *counter;
+      },
+      Err(_e) => {
+        value = 0;
+      }
+    }
+  }
+
+  println!("I was invoked from JavaScript! {:?}", value);
+
+  value
+}

crates/desktop/src-tauri/src/endpoints/text_to_image_endpoint.rs

@@ -1,115 +0,0 @@
-#[cfg(not(target_os = "macos"))]
-use ml_models::ml::stable_diffusion::lcm_pipeline::{lcm_pipeline, Args};
-
-use crate::state::app_config::AppConfig;
-use crate::state::app_dir::AppDataRoot;
-use crate::stubs::model_cache::ModelCache;
-use crate::stubs::prompt_cache::PromptCache;
-use crate::utils::image::encode_dynamic_image_base64_png::encode_dynamic_image_base64_png;
-use image::imageops::FilterType;
-use image::{DynamicImage, ImageReader, RgbImage};
-use log::{error, info};
-use ml_weights_registry::weights_registry::weights::{CLIP_JSON, LYKON_DEAMSHAPER_7_VAE, SDXL_TURBO_CLIP_TEXT_ENCODER, SIMIANLUO_LCM_DREAMSHAPER_V7_UNET};
-
-use once_cell::sync::Lazy;
-use std::io::Cursor;
-use tauri::{AppHandle, State};
-
-const RANDOM_SEED: u32 = 42;
-
-const STRENGTH: f64 = 100.0;
-const PNG_BYTES : &[u8] = include_bytes!("../../binary_includes/1024.png");
-
-// NB: We're just hacking a temporary "text to image" test endpoint leveraging the existing image to image modality.
-static IMAGE : Lazy<DynamicImage> = Lazy::new(||{
-  ImageReader::new(Cursor::new(PNG_BYTES))
-    .with_guessed_format()
-    .expect("failed to detect png format")
-    .decode()
-    .expect("failed to decode image")
-});
-
-#[tauri::command]
-pub async fn text_to_image(
-  prompt: String,
-  model_config: State<'_, AppConfig>,
-  model_cache: State<'_, ModelCache>,
-  prompt_cache: State<'_, PromptCache>,
-  app_data_root: State<'_, AppDataRoot>,
-  app: AppHandle,
-) -> Result<String, String> {
-  info!("text_to_image endpoint called.");
-
-  let mut image = IMAGE.resize(512, 512, FilterType::CatmullRom);
-
-  #[cfg(not(target_os = "macos"))]
-  {
-    let result = text_to_image_impl(
-      &prompt,
-      image,
-      Some(STRENGTH),
-      &model_config,
-      &model_cache,
-      prompt_cache,
-      app,
-      &app_data_root
-    ).await;
-
-    image = match result {
-      Ok(img) => DynamicImage::ImageRgb8(img),
-      Err(err) => {
-        error!("There was an error generating the image: {:?}", err);
-        return Err(format!("There was an error generating the image: {}", err));
-      }
-    };
-  }
-
-  let bytes = encode_dynamic_image_base64_png(image)
-    .map_err(|err| format!("failure to encode image: {:?}", err))?;
-
-  info!("Inference successful; image converted to base64, serving back to browser...");
-
-  Ok(bytes)
-}
-
-#[cfg(not(target_os = "macos"))]
-async fn text_to_image_impl(
-  prompt: &str,
-  image: DynamicImage,
-  strength: Option<f64>,
-  config: &AppConfig,
-  model_cache: &ModelCache,
-  prompt_cache: State<'_, PromptCache>,
-  app: AppHandle,
-  app_data_root: &AppDataRoot,
-) -> Result<RgbImage, String> {
-  
-  let weights_dir = app_data_root.weights_dir();
-  let vae_path = weights_dir.weight_path(&LYKON_DEAMSHAPER_7_VAE);
-  let unet_path = weights_dir.weight_path(&SIMIANLUO_LCM_DREAMSHAPER_V7_UNET);
-  let clip_json_path = weights_dir.weight_path(&CLIP_JSON);
-  let clip_weights_path= weights_dir.weight_path(&SDXL_TURBO_CLIP_TEXT_ENCODER);
-
-  let args = Args {
-    image: &image,
-    prompt: prompt.to_string(),
-    uncond_prompt: "".to_string(),
-    model_cache,
-    prompt_cache: &prompt_cache,
-    img2img_strength: strength,
-    cfg_scale: config.cfg_scale,
-    use_flash_attn: true,
-    model_config: &config.model_config,
-    maybe_seed: None,
-    scheduler_steps: config.scheduler_steps,
-    vae_path: &vae_path,
-    unet_path: &unet_path,
-    clip_json_path: &clip_json_path,
-    clip_weights_path: &clip_weights_path,
-  };
-
-  let image = lcm_pipeline(args)
-    .map_err(|err| format!("failure to encode image: {:?}", err))?;
-
-  Ok(image)
-}

crates/desktop/src-tauri/src/events/mod.rs

@@ -1 +0,0 @@
-pub mod notification_event;

crates/desktop/src-tauri/src/events/notification_event.rs

@@ -1,30 +0,0 @@
-use serde::Serialize;
-
-#[derive(Clone, Copy, Serialize)]
-#[serde(rename_all = "snake_case")]
-pub enum NotificationModelType {
-  ImageSegmentation,
-  TextEncoder,
-  TextTokenizer,
-  Unet,
-  Vae,
-}
-
-#[derive(Clone, Serialize)]
-#[serde(rename_all = "snake_case")]
-pub enum NotificationEvent<'a> {
-  ModelDownloadStarted {
-    model_name: &'a str, 
-    model_type: NotificationModelType,
-  },
-  ModelDownloadProgress {
-    model_name: &'a str,
-    model_type: NotificationModelType,
-    currently_downloaded_bytes: usize,
-    total_file_bytes: usize,
-  },
-  ModelDownloadComplete {
-    model_name: &'a str,
-    model_type: NotificationModelType,
-  }
-}

crates/desktop/src-tauri/src/lib.rs

@@ -1,22 +1,12 @@
 pub mod endpoints;
-pub mod events;
+pub mod ml;
 pub mod state;
-pub mod stubs;
-pub mod threads;
-pub mod transfer;
-pub mod utils;
 
-use crate::endpoints::download_models::download_models;
-use crate::endpoints::inpaint_image_endpoint::inpaint_image;
-use crate::endpoints::flip_image::flip_image;
-use crate::endpoints::realtime_image_endpoint::infer_image;
-use crate::endpoints::remove_background_endpoint::remove_background;
-use crate::endpoints::text_to_image_endpoint::text_to_image;
+use crate::ml::model_cache::ModelCache;
+use crate::ml::prompt_cache::PromptCache;
 use crate::state::app_config::AppConfig;
-use crate::stubs::model_cache::ModelCache;
-use crate::stubs::prompt_cache::PromptCache;
-use crate::threads::downloader_thread::downloader_thread;
-
+use endpoints::image_endpoint::infer_image;
+use endpoints::test_counter::test_counter;
 use tauri_plugin_log::Target;
 use tauri_plugin_log::TargetKind;
 
@@ -29,16 +19,17 @@ pub fn run() {
   let config = AppConfig::init()
     .expect("config should load");
 
-
   let prompt_cache = PromptCache::with_capacity(8)
     .expect("prompt cache should load");
 
   println!("Creating model cache...");
   
-  let model_cache = ModelCache::new();
-  
-  let app_data_root = config.app_data_root.clone();
-  let app_data_root2 = config.app_data_root.clone();
+  let model_cache = ModelCache::new(
+    config.device.clone(),
+    config.dtype,
+    config.sd_version.clone(),
+    config.sd_config.clone()
+  ).expect("Model cache should create");
 
   println!("Initializing backend runtime...");
 
@@ -48,9 +39,6 @@ pub fn run() {
       .targets(vec![Target::new(TargetKind::Stdout)])
       .build())
     .setup(|app| {
-      // TODO(bt): This is broken on windows
-      // log_environment_details();
-
       //if cfg!(debug_assertions) {
       //  app.handle().plugin(
       //    tauri_plugin_log::Builder::default()
@@ -58,23 +46,14 @@ pub fn run() {
       //      .build(),
       //  )?;
       //}
-      let app = app.handle().clone();
-
-      tauri::async_runtime::spawn(downloader_thread(app_data_root2, app));
-
       Ok(())
     })
     .manage(config)
     .manage(prompt_cache)
     .manage(model_cache)
-    .manage(app_data_root)
     .invoke_handler(tauri::generate_handler![
-      download_models,
-      flip_image,
+      test_counter, 
       infer_image,
-      inpaint_image,
-      remove_background,
-      text_to_image,
     ])
     .run(tauri::generate_context!())
     .expect("error while running tauri application");

crates/desktop/src-tauri/src/main.rs

@@ -1,15 +1,6 @@
 // Prevents additional console window on Windows in release, DO NOT REMOVE!!
 #![cfg_attr(not(debug_assertions), windows_subsystem = "windows")]
 
-use std::fs::File;
-
 fn main() {
-  //// NB: This is a test to see the behavior of the application with dynamic loading modules.
-  //// We should be able to execute code before loading CUDA and other dependencies.
-  //let r = File::create("program_started.txt");
-  //match r {
-  //  Err(err) => println!("Failed to create program_started.txt: {}", err),
-  //  Ok(_) => {}
-  //}
   app_lib::run();
 }

crates/desktop/src-tauri/src/ml/cumprod.rs

@@ -0,0 +1,150 @@
+use candle_core::shape::Dim;
+use candle_core::Tensor;
+
+/// Tensor Cumulative Product
+/// 
+/// Candle doesn't implement cumprod, only cumsum.
+/// This implementation is adapted from cumsum in candle-ext:
+///  https://github.com/mokeyish/candle-ext/blob/main/tests/cumsum.rs
+///
+/// See Also: https://pytorch.org/docs/stable/generated/torch.cumprod.html
+/// See Also: https://numpy.org/doc/stable/reference/generated/numpy.cumprod.html
+/// See Also: https://github.com/huggingface/candle/issues/1646
+///
+pub fn cumprod<D: Dim>(input: &Tensor, dim: D) -> candle_core::Result<Tensor> {
+  let dim = dim.to_index(input.shape(), "cumprod")?;
+  let dim_size = input.dim(dim)?;
+  let rank = input.rank();
+  if rank == 0 {
+    return Ok(input.clone());
+  }
+
+  let mut tensors = Vec::with_capacity(dim_size);
+
+  let mut a = input.clone();
+  for i in 0..dim_size {
+    if i > 0 {
+      a = a.narrow(dim, 1, dim_size - i)?;
+      let b = input.narrow(dim, 0, dim_size - i)?;
+      a = (a * b)?;
+    }
+    tensors.push(a.narrow(dim, 0, 1)?);
+  }
+
+  let cumsum = Tensor::cat(&tensors, dim)?;
+  Ok(cumsum)
+}
+
+#[cfg(test)]
+mod tests {
+  use crate::ml::cumprod::cumprod;
+  use candle_core::{Device, Tensor};
+
+  mod base_cases {
+    use super::*;
+
+    #[test]
+    fn zeroes_vector() -> anyhow::Result<()> {
+      let a = Tensor::new(&[0i64, 0, 0, 0], &Device::Cpu)?;
+      let b = cumprod(&a, 0)?;
+      assert_eq!(b.to_vec1::<i64>()?, &[0, 0, 0, 0]);
+      Ok(())
+    }
+
+    #[test]
+    fn zeroes_tensor() -> anyhow::Result<()> {
+      let a = Tensor::new(&[
+        [0i64, 0, 0],
+        [0i64, 0, 0],
+        [0i64, 0, 0],
+      ], &Device::Cpu)?;
+
+      let b = cumprod(&a, 0)?;
+
+      assert_eq!(b.to_vec2::<i64>()?, &[
+        [0, 0, 0],
+        [0, 0, 0],
+        [0, 0, 0],
+      ]);
+      Ok(())
+    }
+
+    #[test]
+    fn ones_vector() -> anyhow::Result<()> {
+      let a = Tensor::new(&[1i64, 1, 1, 1], &Device::Cpu)?;
+      let b = cumprod(&a, 0)?;
+      assert_eq!(b.to_vec1::<i64>()?, &[1, 1, 1, 1]);
+      Ok(())
+    }
+
+    #[test]
+    fn ones_tensor() -> anyhow::Result<()> {
+      let a = Tensor::new(&[
+        [1i64, 1, 1],
+        [1i64, 1, 1],
+        [1i64, 1, 1],
+      ], &Device::Cpu)?;
+
+      let b = cumprod(&a, 0)?;
+
+      assert_eq!(b.to_vec2::<i64>()?, &[
+        [1, 1, 1],
+        [1, 1, 1],
+        [1, 1, 1],
+      ]);
+      Ok(())
+    }
+  }
+
+  mod test_cases {
+    use super::*;
+
+    #[test]
+    fn vector() -> anyhow::Result<()> {
+      let a = Tensor::new(&[1i64, 2, 3, 4, 5], &Device::Cpu)?;
+      let b = cumprod(&a, 0)?;
+      assert_eq!(b.to_vec1::<i64>()?, &[1, 2, 6, 24, 120]);
+
+      let a = Tensor::new(&[0, 1i64, 2, 3, 4, 5], &Device::Cpu)?;
+      let b = cumprod(&a, 0)?;
+      assert_eq!(b.to_vec1::<i64>()?, &[0, 0, 0, 0, 0, 0]);
+      Ok(())
+    }
+
+    #[test]
+    fn tensor_dim_0() -> anyhow::Result<()> {
+      let a = Tensor::new(&[
+        [1i64, 2, 3],
+        [4i64, 5, 6],
+        [7i64, 8, 9],
+      ], &Device::Cpu)?;
+
+      let b = cumprod(&a, 0)?;
+
+      assert_eq!(b.to_vec2::<i64>()?, &[
+        [1, 2, 3],
+        [4, 10, 18],
+        [28, 80, 162],
+      ]);
+      Ok(())
+    }
+
+    #[test]
+    fn tensor_dim_1() -> anyhow::Result<()> {
+      let a = Tensor::new(&[
+        [1i64, 2, 3],
+        [4i64, 5, 6],
+        [7i64, 8, 9],
+      ], &Device::Cpu)?;
+
+      let b = cumprod(&a, 1)?;
+
+      assert_eq!(b.to_vec2::<i64>()?, &[
+        [1, 2, 6],
+        [4, 20, 120],
+        [7, 56, 504],
+      ]);
+      Ok(())
+    }
+  }
+}

crates/desktop/src-tauri/src/ml/lcm_scheduler.rs

@@ -0,0 +1,474 @@
+use anyhow::anyhow;
+use candle_core::{DType, Device, Shape, Tensor, WithDType};
+use candle_transformers::models::stable_diffusion::schedulers::Scheduler;
+use candle_transformers::models::stable_diffusion::utils::linspace;
+use log::error;
+
+/*
+Notes on LCM implementation:
+
+  - Robust technical comparison of samplers: https://stable-diffusion-art.com/samplers/
+  - Apple Swift LCM Implementation ticket: https://github.com/apple/ml-stable-diffusion/issues/319
+  - Reference Swift Implementation: https://github.com/GuernikaCore/Schedulers/blob/main/Sources/Schedulers/LCMScheduler.swift
+
+Candle is missing several "batteries" we'll have to implement, and some we might want to upstream:
+
+  - t[::-1] (syntax to reverse the tensor)
+
+  - torch/numpy.cumprod :
+      - https://numpy.org/doc/stable/reference/generated/numpy.cumprod.html
+      - https://pytorch.org/docs/stable/generated/torch.cumprod.html
+
+Remaining work:
+
+  - [ ] implement add_noise
+  - [ ] implement cumprod
+  - [ ] initialize step_index / begin_index
+  - [ ] find out how `set_timestamps` is called in diffusers and see if we want to change it to CTOR param in candle
+  - [ ] try to plug this into candle and run
+  - [ ] fix timesteps() method to cache values
+  - [ ] compare with diffusers
+  - [ ] compare with candle schedulers
+  - [ ] compare with swift schedulers
+  - [ ] test + unit test
+*/
+
+struct LcmScheduler {
+  alphas: Tensor,
+  alphas_cumprod: Tensor,
+  final_alpha_cumprod: Tensor,
+
+  betas: Tensor,
+
+  timesteps: Tensor,
+
+  // Standard deviation of the initial noise distribution
+  // NB: Only in diffusers, and seems to only hold the value "1.0". Swift implementation omits it.
+  init_noise_sigma: f64,
+
+  // TODO: This is set by the set_timesteps() method after construction in python.
+  //  We might want to handle this upfront in the constructor ourselves
+  num_inference_steps: Option<usize>,
+
+  _step_index: Option<usize>,
+  _begin_index: Option<usize>,
+}
+
+impl LcmScheduler {
+  pub fn new(
+    num_train_timestamps: i64,
+    beta_start: f64,
+    beta_end: f64,
+    beta_schedule: &str,
+    trained_betas: Option<Tensor>,
+    set_alpha_to_one: bool,
+    rescale_betas_zero_snr: bool,
+    device: &Device,
+  ) -> anyhow::Result<Self> {
+    let betas;
+
+    if let Some(trained) = trained_betas {
+      betas = trained;
+    } else {
+      betas = match beta_schedule {
+        "scaled_linear" => {
+          // this schedule is very specific to the latent diffusion model.
+          let t = linspace(
+            beta_start.powf(0.5),
+            beta_end.powf(0.5),
+            num_train_timestamps as usize
+          )?;
+          t.to_device(device)?
+        }
+        _ => {
+          return Err(anyhow!("schedule {beta_schedule} is not implemented"))
+        }
+      };
+    }
+
+    if rescale_betas_zero_snr {
+      return Err(anyhow!("rescale_betas_zero_snr is not used upstream"));
+    }
+
+    let alphas = (1.0 - &betas)?;
+
+    // TODO: Implement cumprod: https://github.com/huggingface/candle/issues/1646
+    //  https://numpy.org/doc/stable/reference/generated/numpy.cumprod.html
+    error!("THIS IS USING CUMSUM, NOT CUMPROD. WE NEED TO IMPLEMENT CUMPROD.");
+    // self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+    let alphas_cumprod = alphas.cumsum(0)?;
+
+    // TODO: Check implementation
+    // At every step in ddim, we are looking into the previous alphas_cumprod
+    // For the final step, there is no previous alphas_cumprod because we are already at 0
+    // `set_alpha_to_one` decides whether we set this parameter simply to one or
+    // whether we use the final alpha of the "non-previous" one.
+    let final_alpha_cumprod = if set_alpha_to_one {
+      initialize_scalar_tensor(1.0, device)? // NB: `set_alpha_to_one` is false in python
+    } else {
+      alphas_cumprod.get(0)?
+    };
+
+    // standard deviation of the initial noise distribution
+    let init_noise_sigma = 1.0;
+
+    let timesteps = initialize_timesteps(num_train_timestamps, device)?;
+
+    Ok(Self {
+      alphas,
+      alphas_cumprod,
+      final_alpha_cumprod,
+      betas,
+      init_noise_sigma,
+      timesteps,
+      num_inference_steps: None,
+      _step_index: None,
+      _begin_index: None,
+    })
+  }
+
+  pub fn set_begin_index(&mut self, begin_index: usize) {
+    self._begin_index = Some(begin_index);
+  }
+
+  pub fn init_step_index(&mut self, timestep: usize) -> candle_core::Result<()> {
+    match self._begin_index {
+      None => {
+        todo!("implement begin index none branch")
+      }
+      Some(begin_index) => {
+        self._step_index = Some(begin_index);
+      }
+    }
+    Ok(())
+  }
+
+  pub fn set_timesteps(&mut self, num_inference_steps: usize, device: &Device) -> candle_core::Result<()> {
+    // NB: The diffusers version of this method passes `None` for `original_inference_steps` and `timesteps`.
+    // It passes strength = 1.0, and config.original_inference_steps = 50.
+    const STRENGTH : f64 = 1.0;
+    const CONFIG_ORIGINAL_INFERENCE_STEPS : usize = 50;
+    const CONFIG_NUM_TRAIN_TIMESTEPS : usize = 1000;
+
+    let original_steps = CONFIG_ORIGINAL_INFERENCE_STEPS;
+
+    // LCM Timesteps Setting
+    // The skipping step parameter k from the paper.
+    let k = (CONFIG_NUM_TRAIN_TIMESTEPS as f64 / original_steps as f64).floor();
+
+    let end = (original_steps as f64 * STRENGTH) as i64 + 1;
+    let lcm_origin_timesteps = ((Tensor::arange::<i64>(1, end, device)? * k)? - 1.0)?;
+
+    if num_inference_steps > CONFIG_NUM_TRAIN_TIMESTEPS {
+      return Err(candle_core::Error::Msg("num_inference_steps cannot be larger than train timesteps".to_string()));
+    }
+
+    let skipping_step = (lcm_origin_timesteps.dims().len() as f64 / num_inference_steps as f64).floor() as usize;
+
+    self.num_inference_steps = Some(num_inference_steps);
+    
+    // TODO: This is not efficient
+    let lcm_origin_timesteps = lcm_origin_timesteps.to_vec1::<i64>()?
+      .into_iter()
+      .map(i64::from)
+      .rev()
+      .collect::<Vec<i64>>();
+    
+    // TODO: This might not be exact. We should unit test this
+    let inference_indices = linspace(0.0, lcm_origin_timesteps.len() as f64, num_inference_steps)?;
+    
+    let inference_indices = inference_indices.floor()?
+      .to_dtype(DType::I64)?;
+    
+    let timestamps = {
+      // TODO: This code is awful
+      let inference_indices_t: Vec<usize> = inference_indices.to_vec1::<i64>()?
+        .into_iter()
+        .map(|n| n as usize)
+        .collect();
+
+      let mut timesteps: Vec<i64> = Vec::with_capacity(inference_indices.dims().len());
+      
+      for i in inference_indices_t {
+        if let Some(val) = lcm_origin_timesteps.get(i) {
+          timesteps.push(*val);
+        }
+      }
+      
+      let shape = Shape::from_dims(&[timesteps.len()]);
+      
+      Tensor::from_vec(timesteps, shape, device)?
+        .to_dtype(DType::I64)?
+        .to_device(device)?
+    };
+
+    self.timesteps = timestamps;
+
+    Ok(())
+  }
+
+}
+
+impl Scheduler for LcmScheduler {
+  fn timesteps(&self) -> &[usize] {
+    // TODO: We can't use the implementation from DDIM. This method requires
+    //  (1) an infallible response and (2) a reference we can't dynamically allocate. We must
+    //  cache this somewhere.
+    &[]
+  }
+
+  fn add_noise(&self, original_samples: &Tensor, noise: Tensor, timestep: usize) -> candle_core::Result<Tensor> {
+    // TODO: Make sure alphas_cumprod and timestep have same device and dtype as original_samples
+    //  Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement
+    //  for the subsequent add_noise calls
+    let mut alphas_cumprod = self.alphas_cumprod.to_device(original_samples.device())?;
+    alphas_cumprod = alphas_cumprod.to_dtype(original_samples.dtype())?;
+    //let timesteps = timesteps.to_device(original_samples.device())?;
+
+    let sqrt_alpha_prod = alphas_cumprod.get(timestep)?.powf(0.5)?;
+    let mut sqrt_alpha_prod = sqrt_alpha_prod.flatten_all()?; // TODO: Verify that flatten_all works here
+
+
+    while sqrt_alpha_prod.dims().len() < original_samples.dims().len() {
+      // TODO: Verify that this works for `unsqueeze(-1)`
+      sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(sqrt_alpha_prod.dims().len())?;
+    }
+
+    let sqrt_one_minus_alpha_prod = (1.0 - alphas_cumprod.get(timestep)?)?.powf(0.5)?;
+    let mut sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten_all()?; // TODO: Verify flatten all works here.
+
+    while sqrt_one_minus_alpha_prod.dims().len() < original_samples.dims().len() {
+      // TODO: Verify that this works for `unsqueeze(-1)`
+      sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(sqrt_one_minus_alpha_prod.dims().len())?;
+    }
+
+    let noisy_samples = ((sqrt_alpha_prod * original_samples)? + (sqrt_one_minus_alpha_prod * noise)?)?;
+
+    Ok(noisy_samples)
+  }
+
+  fn init_noise_sigma(&self) -> f64 {
+    self.init_noise_sigma
+  }
+
+  /// Ensures interchangeability with schedulers that need to scale the denoising model input
+  /// depending on the current timestep. (For LCM, this is a pass through no-op!)
+  fn scale_model_input(&self, sample: Tensor, _timestep: usize) -> candle_core::Result<Tensor> {
+    Ok(sample) // NB: The diffusers scheduling_lcm.py implementation is a pass through no-op !
+  }
+
+  /// Predict the sample from the previous timestep by reversing the SDE. This function
+  /// propagates the diffusion process from the learned model outputs (most often the
+  /// predicted noise).
+  fn step(&mut self, model_output: &Tensor, timestep: usize, sample: &Tensor) -> candle_core::Result<Tensor> {
+
+    if self.num_inference_steps.is_none() {
+      return Err(candle_core::Error::Msg(
+        "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+          .to_string()));
+    }
+
+    if self._step_index.is_none() {
+      self.init_step_index(timestep)?;
+    }
+
+    let step_index = self._step_index.clone().expect("unsafe unwrap; should exist per above");
+
+    // 1. Get previous step value
+    let prev_step_index = step_index + 1;
+    let prev_timestep;
+
+    if prev_step_index < self.timesteps.elem_count() {
+      prev_timestep = self.timesteps.get(timestep)?.to_scalar::<i64>()? as usize;
+    } else {
+      prev_timestep = timestep;
+    }
+
+    // 2. Compute alphas, betas
+    let alpha_prod_t = self.alphas_cumprod.get(timestep)?;
+    let alpha_prod_t_prev = if prev_timestep >= 0 {
+      self.alphas_cumprod.get(prev_timestep)?
+    } else {
+      self.final_alpha_cumprod.clone()
+    };
+
+    let beta_prod_t = (1. - &alpha_prod_t)?;
+    let beta_prod_t_prev = (1. - &alpha_prod_t_prev)?;
+
+    // 3. Get scaling for boundary conditions
+    // NB: timestep_scaling seems to be always set to 10.0 in diffusers examples, but it is passed
+    // via a config object. Hardcoding this for now until we need to change it.
+    const TIMESTEP_SCALING : f64 = 10.0;
+    let (c_skip, c_out) = get_scalings_for_boundary_condition_discrete(timestep, TIMESTEP_SCALING);
+
+    // 4. Compute the predicted original sample x_0 based on the model parameterization
+    // NB: Prediction type is "epsilon" in the diffusers examples, but that is configurable.
+    // This is the only one we'll implement for now.
+    let predicted_original_sample = ((sample - (beta_prod_t.sqrt()? * model_output)?)? / &alpha_prod_t.sqrt()?)?;
+
+    // 5. (skip clip/threshold - not used in diffusers example)
+
+    // 6. Denoise model output using boundary conditions
+    let denoised = ((c_out * predicted_original_sample)? + (c_skip * sample)?)?;
+
+    // TODO: Step 7. We have to return the previous example somehow.
+    //  Diffusers tolerates it not being sent back, but the quality degrades as if one step is absent.
+    Ok(denoised)
+  }
+}
+
+fn initialize_scalar_tensor<D: WithDType>(value: D, device: &Device) -> anyhow::Result<Tensor> {
+  let scalar = Tensor::from_vec(vec![value], Shape::from(vec![]), device)?;
+  Ok(scalar)
+}
+
+// TODO(bt,2025-02-23): Use native implementations
+fn initialize_timesteps(num_train_timesteps: i64, device: &Device) -> anyhow::Result<Tensor> {
+  // Testing LCM scheduler assumptions:
+  // [::-1] is shorthand for reversing a numpy vector, eg. [0, 1,2,3,4] -> [4, 3, 2, 1, 0]
+  // self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64))
+  //let tensor = Tensor::arange::<i64>(0, 5, &Device::Cpu)?; // This works, but we can't reverse/flip it!
+
+  // https://github.com/tracel-ai/burn/pull/1468 flip impl? Will this work in place?
+  // https://github.com/huggingface/candle/issues/1875
+
+  let data = (0..num_train_timesteps)
+    .map(i64::from)
+    .rev()
+    .collect();
+
+  let shape = Shape::from_dims(&[num_train_timesteps as usize]);
+
+  let tensor = Tensor::from_vec(data, shape, device)?;
+
+  Ok(tensor)
+}
+
+fn get_scalings_for_boundary_condition_discrete(timestep: usize, timestep_scaling: f64) -> (f64, f64) {
+  let sigma_data = 0.5f64; // Default 0.5
+  let scaled_timestep = timestep as f64 * timestep_scaling;
+
+  let c_skip = sigma_data.powf(2.0) / (scaled_timestep.powf(2.0) + sigma_data.powf(2.0));
+  let c_out = scaled_timestep / (scaled_timestep.powf(2.0) + sigma_data.powf(2.0)).powf(0.5);
+
+  (c_skip, c_out)
+}
+
+// The Candle sources, and moreover, the Candle tests, are a great source of information for how the library works.
+//assert_eq!(t.i(..-1)?, 1.0);
+#[cfg(test)]
+mod tests {
+  use crate::ml::lcm_scheduler::{initialize_scalar_tensor, initialize_timesteps};
+  use crate::ml::lcm_scheduler::tests::test_helper_fn::*;
+  use candle_core::Device;
+
+  mod helper_function_unit_tests {
+    use crate::ml::lcm_scheduler::get_scalings_for_boundary_condition_discrete;
+    use super::*;
+
+    #[test]
+    fn test_initialize_scalar_tensor() -> anyhow::Result<()> {
+      let t = initialize_scalar_tensor(1.0, &Device::Cpu)?;
+      let empty: [usize; 0] = [];
+      assert_eq!(t.dims(), &empty);
+      assert_eq!(t.to_vec0::<f64>()?, 1.0);
+      Ok(())
+    }
+
+    #[test]
+    fn test_initialize_timesteps() -> anyhow::Result<()> {
+      let tensor = initialize_timesteps(5, &Device::Cpu)?;
+      assert_eq!(tensor.to_vec1::<i64>()?, &[4, 3, 2, 1,0]);
+      Ok(())
+    }
+
+    #[test]
+    fn test_get_scalings_for_boundary_condition_discrete() -> anyhow::Result<()> {
+      // Base cases
+      let v = get_scalings_for_boundary_condition_discrete(0, 0.0);
+      assert_eq!(v, (1.0, 0.0));
+      let v = get_scalings_for_boundary_condition_discrete(1, 0.0);
+      assert_eq!(v, (1.0, 0.0));
+      let v = get_scalings_for_boundary_condition_discrete(0, 1.0);
+      assert_eq!(v, (1.0, 0.0));
+
+      // Production code has "timestep_scaling" = 10.0
+      let v = get_scalings_for_boundary_condition_discrete(0, 10.0);
+      assert_eq!(v, (1.0, 0.0));
+
+      // Production code has "timestep_scaling" = 10.0
+      let v = get_scalings_for_boundary_condition_discrete(1, 10.0);
+      assert!((v.0 - 0.0024937655860349127).abs() < 0.01);
+      assert!((v.1 - 0.9987523388778445).abs() < 0.01);
+
+      // TODO: More test cases -
+      //let v = get_scalings_for_boundary_condition_discrete(10, 10.0);
+      //let v = get_scalings_for_boundary_condition_discrete(100, 10.0);
+
+      Ok(())
+    }
+  }
+
+  mod confirm_various_pytorch_behaviors {
+    use candle_core::IndexOp;
+    use super::*;
+
+    // NB: Non-unit tests, just to test understanding while porting code
+
+    #[test]
+    fn unsqueeze_zero() -> anyhow::Result<()> {
+      let t = tensor_1234()?;
+      let x = t.unsqueeze(0)?;
+      assert_eq!(x.to_vec2::<i64>()?, &[&[1, 2, 3, 4]]);
+      Ok(())
+    }
+
+    #[test]
+    fn unsqueeze_simulating_negative_one() -> anyhow::Result<()> {
+      let t = tensor_1234()?;
+      // NB: It looks like this matches the behavior of unsqueeze(-1)
+      // Per: https://pytorch.org/docs/stable/generated/torch.unsqueeze.html
+      let x = t.unsqueeze(t.dims().len())?;
+      assert_eq!(x.to_vec2::<i64>()?, &[&[1],&[2],&[3],&[4]]);
+      Ok(())
+    }
+
+    #[test]
+    fn one_minus_vector() -> anyhow::Result<()> {
+      let t = tensor_01234()?;
+      let x = (1. - t)?;
+      assert_eq!(x.rank(), 1); // NB: 1-dimensional output
+      assert_eq!(x.to_vec1::<i64>()?, &[1, 0, -1, -2, -3]);
+      Ok(())
+    }
+    
+    // #[test]
+    // fn reverse_vector() -> anyhow::Result<()> {
+    //   let t = tensor_1234()?;
+    //   let x = t.i(5..0)?;
+    //   assert_eq!(x.rank(), 1); // NB: 1-dimensional output
+    //   assert_eq!(x.to_vec1::<i64>()?, &[1, 0, -1, -2, -3]);
+    //   Ok(())
+    // }
+  }
+
+  pub mod test_helper_fn {
+    use candle_core::{Device, Shape, Tensor};
+
+    pub fn tensor_1234() -> anyhow::Result<Tensor> {
+      // Return an example tensor
+      let data: Vec<i64> = vec![1, 2, 3, 4];
+      let t = Tensor::from_vec(data, Shape::from(vec![4]), &Device::Cpu)?;
+      assert_eq!(t.to_vec1::<i64>()?, &[1, 2, 3, 4]);
+      Ok(t)
+    }
+
+    pub fn tensor_01234() -> anyhow::Result<Tensor> {
+      // Return an example tensor
+      let data: Vec<i64> = vec![0, 1, 2, 3, 4];
+      let t = Tensor::from_vec(data, Shape::from(vec![5]), &Device::Cpu)?;
+      assert_eq!(t.to_vec1::<i64>()?, &[0, 1, 2, 3, 4]);
+      Ok(t)
+    }
+  }
+}

crates/desktop/src-tauri/src/ml/mod.rs

@@ -1,7 +1,8 @@
-pub mod background_removal;
+pub mod cumprod;
 pub mod image;
+pub mod lcm_scheduler;
 pub mod model_cache;
-pub mod model_config;
 pub mod model_file;
+pub mod models;
 pub mod prompt_cache;
 pub mod stable_diffusion;

crates/desktop/src-tauri/src/ml/model_cache.rs

@@ -0,0 +1,118 @@
+use crate::ml::models::unet_model::UNetModel;
+use crate::ml::model_file::StableDiffusionVersion;
+use anyhow::anyhow;
+use candle_core::{DType, Device, Tensor};
+use candle_transformers::models::stable_diffusion::StableDiffusionConfig;
+use hf_hub::api::sync::Api;
+use std::sync::{Arc, RwLock};
+use candle_transformers::models::stable_diffusion::vae::DiagonalGaussianDistribution;
+use crate::ml::models::lazy_load_vae_model::LazyLoadVaeModel;
+// TODO: This data structure is gross. Generalize the locking and loading semantics for reuse.
+
+// Simple registry for now. We can build complex machinery that aids in 
+// VRAM utilization, disk space saving, intelligent scheduling, etc. in the future.
+pub struct ModelCache {
+  // TODO(bt,2025-02-14): Support for more than one device in the future.
+  device: Device,
+  dtype: DType,
+  hf_api: Api,
+  sd_config: StableDiffusionConfig,
+  sd_version: StableDiffusionVersion,
+  
+  unet: Arc<RwLock<Option<UNetModel>>>,
+  lazy_vae: LazyLoadVaeModel,
+}
+
+impl ModelCache {
+  pub fn new(
+    device: Device, 
+    dtype: DType,
+    sd_version: StableDiffusionVersion, 
+    sd_config: StableDiffusionConfig,
+  ) -> anyhow::Result<Self> {
+    let api = Api::new()?;
+    
+    Ok(Self { 
+      device: device.clone(),
+      dtype,
+      hf_api: api.clone(),
+      unet: Arc::new(RwLock::new(None)),
+      lazy_vae: LazyLoadVaeModel::new(
+        &sd_config,
+        &sd_version,
+        &api,
+        &device,
+        dtype,
+      )?,
+      sd_config,
+      sd_version,
+    })
+  }
+
+  pub fn vae_encode(&self, xs: &Tensor) -> anyhow::Result<DiagonalGaussianDistribution> {
+    self.lazy_vae.encode(xs)
+  }
+
+  pub fn vae_decode(&self, xs: &Tensor) -> anyhow::Result<Tensor> {
+    self.lazy_vae.decode(xs)
+  }
+  
+  pub fn unet_inference(&self, latent_model_input: &Tensor, timestep: f64, text_embeddings: &Tensor) -> anyhow::Result<Tensor> {
+    println!("unet_inference");
+    match self.try_unet_inference(latent_model_input, timestep, text_embeddings) {
+      Err(err) => Err(err),
+      Ok(Some(result)) => Ok(result),
+      Ok(None) => {
+        // Model wasn't previously loaded
+        self.load_unet()?;
+        self.try_unet_inference(latent_model_input, timestep, text_embeddings)?
+          .ok_or(anyhow!("model was not previously loaded"))
+      }
+    }
+  }
+  
+  fn load_unet(&self) -> anyhow::Result<()> {
+    println!("load_unet");
+    match self.unet.write() {
+      Err(err) => return Err(anyhow!("lock error: {:?}", err)),
+      Ok(mut maybe_model) => {
+        match &*maybe_model {
+          Some(_model) => {} // Fall through
+          None => {
+            let repo = self.sd_version.repo();
+            
+            println!("Downloading UNET model files from: {} ...", repo);
+
+            let unet_file = self.hf_api.model(repo.to_string())
+              .get("unet/diffusion_pytorch_model.safetensors")
+              .map_err(|err| anyhow!("error fetching model: {:?}", err))?;
+
+            println!("Building UNET model... (3)");
+            let unet = UNetModel::new(&self.sd_config, unet_file, &self.device, self.dtype)
+              .map_err(|err| anyhow!("error initializing unet model: {:?}", err))?;
+
+            println!("Built UNET...");
+            *maybe_model = Some(unet);
+          }
+        }
+      }
+    }
+    Ok(())
+  }
+
+  pub fn try_unet_inference(&self, latent_model_input: &Tensor, timestep: f64, text_embeddings: &Tensor) -> anyhow::Result<Option<Tensor>> {
+    println!("try_unet_inference");
+    match self.unet.read() {
+      Ok(mut model) => {
+        match &*model {
+          None => Ok(None), // Model is not yet loaded
+          Some(model) => {
+            Ok(Some(model.inference(latent_model_input, timestep, text_embeddings)?))
+          }
+        }
+      }
+      Err(err) => Err(anyhow!("lock error: {:?}", err)),
+    }
+  }
+  
+}

crates/desktop/src-tauri/src/ml/model_file.rs

@@ -0,0 +1,83 @@
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+use candle_transformers::models::stable_diffusion;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum ModelFile {
+  Tokenizer,
+  Tokenizer2,
+  Clip,
+  Clip2,
+  Unet,
+  Vae,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum StableDiffusionVersion {
+  V1_5,
+  V1_5Inpaint,
+  V2_1,
+  V2Inpaint,
+  Xl,
+  XlInpaint,
+  Turbo,
+}
+
+impl StableDiffusionVersion {
+  pub fn repo(&self) -> &'static str {
+    match self {
+      Self::XlInpaint => "diffusers/stable-diffusion-xl-1.0-inpainting-0.1",
+      Self::Xl => "stabilityai/stable-diffusion-xl-base-1.0",
+      Self::V2Inpaint => "stabilityai/stable-diffusion-2-inpainting",
+      Self::V2_1 => "stabilityai/stable-diffusion-2-1",
+      Self::V1_5 => "runwayml/stable-diffusion-v1-5",
+      Self::V1_5Inpaint => "stable-diffusion-v1-5/stable-diffusion-inpainting",
+      Self::Turbo => "stabilityai/sdxl-turbo",
+    }
+  }
+}
+
+impl ModelFile {
+  pub fn get(
+    &self,
+    filename: Option<String>,
+    version: StableDiffusionVersion,
+    use_f16: bool,
+  ) -> anyhow::Result<std::path::PathBuf> {
+    use hf_hub::api::sync::Api;
+    match filename {
+      Some(filename) => Ok(std::path::PathBuf::from(filename)),
+      None => {
+        let (repo, path) = match self {
+          Self::Tokenizer => {
+            let tokenizer_repo = match version {
+              StableDiffusionVersion::V1_5
+              | StableDiffusionVersion::V2_1
+              | StableDiffusionVersion::V1_5Inpaint
+              | StableDiffusionVersion::V2Inpaint => "openai/clip-vit-base-patch32",
+              StableDiffusionVersion::Xl
+              | StableDiffusionVersion::XlInpaint
+              | StableDiffusionVersion::Turbo => {
+                "openai/clip-vit-large-patch14"
+              }
+            };
+            (tokenizer_repo, "tokenizer.json")
+          }
+          Self::Tokenizer2 => {
+            ("laion/CLIP-ViT-bigG-14-laion2B-39B-b160k", "tokenizer.json")
+          }
+          Self::Clip => (version.repo(), "text_encoder/model.safetensors"),
+          Self::Clip2 => (version.repo(), "text_encoder_2/model.safetensors"),
+          Self::Unet => (version.repo(), "unet/diffusion_pytorch_model.safetensors"),
+          Self::Vae => (version.repo(), "vae/diffusion_pytorch_model.safetensors"),
+        };
+        let filename = Api::new()?.model(repo.to_string()).get(path)?;
+        Ok(filename)
+      }
+    }
+  }
+}
+

crates/desktop/src-tauri/src/ml/models/lazy_load_vae_model.rs

@@ -0,0 +1,130 @@
+use crate::ml::model_file::StableDiffusionVersion;
+use anyhow::anyhow;
+use candle_core::{DType, Device, Tensor};
+use candle_transformers::models::stable_diffusion::vae::{AutoEncoderKL, DiagonalGaussianDistribution};
+use candle_transformers::models::stable_diffusion::StableDiffusionConfig;
+use hf_hub::api::sync::Api;
+use std::sync::{Arc, RwLock};
+
+pub struct LazyLoadVaeModel {
+  lazy_model: Arc<RwLock<Option<AutoEncoderKL>>>,
+  model_configs: VaeConfigs,
+}
+
+// TODO: Maybe we pass these via a central config read at startup, passed as an Arc<ConfigT>
+pub struct VaeConfigs {
+  sd_config: StableDiffusionConfig,
+  sd_version: StableDiffusionVersion,
+  device: Device,
+  dtype: DType,
+  hf_api: Api,
+}
+
+impl LazyLoadVaeModel {
+  pub fn new(
+    sd_config: &StableDiffusionConfig,
+    sd_version: &StableDiffusionVersion,
+    hf_api: &Api,
+    device: &Device,
+    dtype: DType,
+  ) -> anyhow::Result<Self> {
+    println!("building vae model... (1)");
+    
+    Ok(Self {
+      lazy_model: Arc::new(RwLock::new(None)),
+      model_configs: VaeConfigs {
+        sd_config: sd_config.clone(),
+        sd_version: sd_version.clone(),
+        device: device.clone(),
+        hf_api: hf_api.clone(),
+        dtype
+      }
+    })
+  }
+
+  pub fn encode(&self, xs: &Tensor) -> anyhow::Result<DiagonalGaussianDistribution> {
+    println!("vae encode...");
+    match self.try_encode(xs) {
+      Err(err) => Err(err),
+      Ok(Some(result)) => Ok(result),
+      Ok(None) => {
+        // Model wasn't previously loaded
+        self.load_model()?;
+        self.try_encode(xs)?
+          .ok_or(anyhow!("model was not previously loaded"))
+      }
+    }
+  }
+  
+  pub fn decode(&self, xs: &Tensor) -> anyhow::Result<Tensor> {
+    println!("vae decode...");
+    match self.try_decode(xs) {
+      Err(err) => Err(err),
+      Ok(Some(result)) => Ok(result),
+      Ok(None) => {
+        // Model wasn't previously loaded
+        self.load_model()?;
+        self.try_decode(xs)?
+          .ok_or(anyhow!("model was not previously loaded"))
+      }
+    }
+  }
+
+  fn try_encode(&self, xs: &Tensor) -> anyhow::Result<Option<DiagonalGaussianDistribution>> {
+    match self.lazy_model.read() {
+      Err(err) => Err(anyhow!("lock error: {:?}", err)),
+      Ok(model) => {
+        match &*model {
+          None => Ok(None), // Model is not yet loaded
+          Some(model) => {
+            Ok(Some(model.encode(xs)?))
+          }
+        }
+      }
+    }
+  }
+
+  fn try_decode(&self, xs: &Tensor) -> anyhow::Result<Option<Tensor>> {
+    match self.lazy_model.read() {
+      Err(err) => Err(anyhow!("lock error: {:?}", err)),
+      Ok(model) => {
+        match &*model {
+          None => Ok(None), // Model is not yet loaded
+          Some(model) => {
+            Ok(Some(model.decode(xs)?))
+          }
+        }
+      }
+    }
+  }
+
+  fn load_model(&self) -> anyhow::Result<()> {
+    println!("load vae");
+    match self.lazy_model.write() {
+      Err(err) => return Err(anyhow!("lock error: {:?}", err)),
+      Ok(mut maybe_model) => {
+        match &*maybe_model {
+          Some(_model) => {} // Already loaded; fall through
+          None => {
+            let repo = self.model_configs.sd_version.repo();
+
+            println!("Building VAE model from : {:?} ... (3)", repo);
+            
+            let vae_file = self.model_configs.hf_api.model(repo.to_string())
+              .get("vae/diffusion_pytorch_model.safetensors")?;
+
+            println!("Building VAE model from file {:?}...", &vae_file);
+            let vae = self.model_configs
+              .sd_config
+              .build_vae(vae_file, &self.model_configs.device, self.model_configs.dtype)?;
+
+            println!("Built VAE...");
+            *maybe_model = Some(vae);
+          }
+        }
+      }
+    }
+    Ok(())
+  }
+ 
+}

crates/desktop/src-tauri/src/ml/models/mod.rs

@@ -0,0 +1,2 @@
+pub mod unet_model;
+pub mod lazy_load_vae_model;

crates/desktop/src-tauri/src/ml/models/unet_model.rs

@@ -0,0 +1,35 @@
+use candle_core::{DType, Device, Tensor};
+use candle_transformers::models::stable_diffusion::unet_2d::UNet2DConditionModel;
+use candle_transformers::models::stable_diffusion::StableDiffusionConfig;
+use std::path::Path;
+
+pub struct UNetModel {
+  model: UNet2DConditionModel,
+}
+
+impl UNetModel {
+  pub fn new<P: AsRef<Path>>(
+    sd_config: &StableDiffusionConfig,
+    unet_weights_file: P,
+    device: &Device,
+    dtype: DType,
+  ) -> anyhow::Result<Self> {
+    println!("building unet model... (1)");
+    
+    let model = sd_config.build_unet(
+      unet_weights_file, 
+      &device, 
+      4, 
+      false, 
+      dtype
+    )?;
+    
+    Ok(Self {
+      model,
+    })
+  }
+  
+  pub fn inference(&self, latent_model_input: &Tensor, timestep: f64, text_embeddings: &Tensor) -> anyhow::Result<Tensor> {
+    Ok(self.model.forward(&latent_model_input, timestep, &text_embeddings)?)
+  }
+}

crates/desktop/src-tauri/src/ml/stable_diffusion/infer_clip_text_embeddings.rs

@@ -0,0 +1,153 @@
+#[cfg(feature = "accelerate")]
+extern crate accelerate_src;
+#[cfg(feature = "mkl")]
+extern crate intel_mkl_src;
+
+use candle_transformers::models::stable_diffusion;
+
+use crate::ml::model_file::{ModelFile, StableDiffusionVersion};
+use anyhow::Error as E;
+use candle_core::{DType, Device, Module, Tensor, D};
+use log::info;
+use tokenizers::Tokenizer;
+
+#[allow(clippy::too_many_arguments)]
+pub fn infer_clip_text_embeddings(
+  prompt: &str,
+  uncond_prompt: &str,
+  tokenizer: Option<String>,
+  clip_weights: Option<String>,
+  clip2_weights: Option<String>,
+  sd_version: StableDiffusionVersion,
+  sd_config: &stable_diffusion::StableDiffusionConfig,
+  use_f16: bool,
+  device: &Device,
+  dtype: DType,
+  use_guide_scale: bool,
+) -> anyhow::Result<Tensor> {
+  info!("Preparing text embeddings...");
+
+  let which = match sd_version {
+    StableDiffusionVersion::Xl
+    | StableDiffusionVersion::XlInpaint
+    | StableDiffusionVersion::Turbo => vec![true, false],
+    _ => vec![true],
+  };
+
+  let text_embeddings = which
+    .iter()
+    .map(|first| {
+      do_infer_clip_text_embeddings(
+        &prompt,
+        &uncond_prompt,
+        None, // tokenizer
+        None, // clip_weights
+        None, // clip2_weights
+        sd_version,
+        &sd_config,
+        false, // use_f16
+        &device,
+        dtype,
+        use_guide_scale,
+        *first,
+      )
+    })
+    .collect::<anyhow::Result<Vec<_>>>()?;
+
+  let text_embeddings = Tensor::cat(&text_embeddings, D::Minus1)?;
+  
+  Ok(text_embeddings)
+}
+
+#[allow(clippy::too_many_arguments)]
+fn do_infer_clip_text_embeddings(
+    prompt: &str,
+    uncond_prompt: &str,
+    tokenizer: Option<String>,
+    clip_weights: Option<String>,
+    clip2_weights: Option<String>,
+    sd_version: StableDiffusionVersion,
+    sd_config: &stable_diffusion::StableDiffusionConfig,
+    use_f16: bool,
+    device: &Device,
+    dtype: DType,
+    use_guide_scale: bool,
+    first: bool,
+  ) -> anyhow::Result<Tensor> {
+  
+  let tokenizer_file = if first {
+    ModelFile::Tokenizer
+  } else {
+    ModelFile::Tokenizer2
+  };
+  let tokenizer = tokenizer_file.get(tokenizer, sd_version, use_f16)?;
+  let tokenizer = Tokenizer::from_file(tokenizer).map_err(E::msg)?;
+  let pad_id = match &sd_config.clip.pad_with {
+    Some(padding) => *tokenizer.get_vocab(true).get(padding.as_str()).unwrap(),
+    None => *tokenizer.get_vocab(true).get("<|endoftext|>").unwrap(),
+  };
+
+  println!("Running with prompt \"{prompt}\".");
+  let mut tokens = tokenizer
+    .encode(prompt, true)
+    .map_err(E::msg)?
+    .get_ids()
+    .to_vec();
+  if tokens.len() > sd_config.clip.max_position_embeddings {
+    anyhow::bail!(
+            "the prompt is too long, {} > max-tokens ({})",
+            tokens.len(),
+            sd_config.clip.max_position_embeddings
+        )
+  }
+  while tokens.len() < sd_config.clip.max_position_embeddings {
+    tokens.push(pad_id)
+  }
+  let tokens = Tensor::new(tokens.as_slice(), device)?.unsqueeze(0)?;
+
+  println!("Building the Clip transformer.");
+  let clip_weights_file = if first {
+    ModelFile::Clip
+  } else {
+    ModelFile::Clip2
+  };
+  let clip_weights = if first {
+    clip_weights_file.get(clip_weights, sd_version, use_f16)?
+  } else {
+    clip_weights_file.get(clip2_weights, sd_version, use_f16)?
+  };
+  let clip_config = if first {
+    &sd_config.clip
+  } else {
+    sd_config.clip2.as_ref().unwrap()
+  };
+  let text_model = stable_diffusion::build_clip_transformer(clip_config, clip_weights, device, DType::F32)?;
+
+  let text_embeddings = text_model.forward(&tokens)?;
+
+  let text_embeddings = if use_guide_scale {
+    let mut uncond_tokens = tokenizer
+      .encode(uncond_prompt, true)
+      .map_err(E::msg)?
+      .get_ids()
+      .to_vec();
+    if uncond_tokens.len() > sd_config.clip.max_position_embeddings {
+      anyhow::bail!(
+                "the negative prompt is too long, {} > max-tokens ({})",
+                uncond_tokens.len(),
+                sd_config.clip.max_position_embeddings
+            )
+    }
+    while uncond_tokens.len() < sd_config.clip.max_position_embeddings {
+      uncond_tokens.push(pad_id)
+    }
+
+    let uncond_tokens = Tensor::new(uncond_tokens.as_slice(), device)?.unsqueeze(0)?;
+    let uncond_embeddings = text_model.forward(&uncond_tokens)?;
+
+    Tensor::cat(&[uncond_embeddings, text_embeddings], 0)?.to_dtype(dtype)?
+  } else {
+    text_embeddings.to_dtype(dtype)?
+  };
+  Ok(text_embeddings)
+}

crates/desktop/src-tauri/src/ml/stable_diffusion/mod.rs

@@ -0,0 +1,3 @@
+pub mod get_vae_scale;
+pub mod infer_clip_text_embeddings;
+pub mod stable_diffusion_pipeline;

crates/desktop/src-tauri/src/ml/stable_diffusion/stable_diffusion_pipeline.rs

@@ -1,5 +1,4 @@
-use std::path::Path;
-use std::sync::{Arc, RwLock, RwLockReadGuard, RwLockWriteGuard};
+use std::sync::{RwLock, RwLockReadGuard, RwLockWriteGuard};
 
 use crate::ml::image::dynamic_image_to_tensor::dynamic_image_to_tensor;
 use crate::ml::image::tensor_to_image_buffer::{tensor_to_image_buffer, RgbImage};
@@ -7,80 +6,49 @@ use crate::ml::model_cache::ModelCache;
 use crate::ml::model_file::StableDiffusionVersion;
 use crate::ml::prompt_cache::PromptCache;
 use crate::ml::stable_diffusion::get_vae_scale::get_vae_scale;
-use crate::ml::stable_diffusion::infer_clip_text_embeddings::{infer_clip_text_embeddings, ClipArgs};
+use crate::ml::stable_diffusion::infer_clip_text_embeddings::infer_clip_text_embeddings;
+use crate::state::app_config::AppConfig;
 use anyhow::{anyhow, Error as E, Result};
-use candle_core::{DType, Device, IndexOp, Tensor, D};
-use candle_transformers::models::stable_diffusion::vae::{AutoEncoderKL, DiagonalGaussianDistribution};
-use candle_transformers::models::stable_diffusion::StableDiffusionConfig;
+use candle_core::{DType, IndexOp, Tensor, D};
+use candle_transformers::models::stable_diffusion::vae::DiagonalGaussianDistribution;
 use image::DynamicImage;
 use log::info;
 use rand::Rng;
 
-pub struct Args<'a, P: AsRef<Path>, Q: AsRef<Path>, R: AsRef<Path>, S: AsRef<Path>> {
+pub struct Args<'a> {
     pub image: &'a DynamicImage,
     pub prompt: String,
     pub uncond_prompt: String,
     pub cfg_scale: Option<f64>,
     pub i2i_strength: Option<u8>,
+    pub configs: &'a AppConfig,
     pub model_cache: &'a ModelCache,
     pub prompt_cache: &'a PromptCache,
-    pub sd_version: StableDiffusionVersion,
-    pub sd_config: StableDiffusionConfig,
-    pub device: &'a Device,
-    pub dtype: DType,
-    pub maybe_seed: Option<u64>,
-    pub scheduler_steps: usize,
-    pub sdxl_turbo_vae_path: P,
-    pub sdxl_turbo_unet_path: Q,
-    pub clip_json_path: R,
-    pub clip_weights_path: S,
 }
 
-pub fn stable_diffusion_pipeline<P1, P2, P3, P4>(args: Args<'_, P1, P2, P3, P4>) -> Result<RgbImage> 
-  where P1: AsRef<Path>, P2: AsRef<Path>, P3: AsRef<Path>, P4: AsRef<Path>
-{
-    let Args { 
-        prompt, 
-        uncond_prompt, 
-        cfg_scale, 
-        i2i_strength, 
-        model_cache, 
-        prompt_cache, 
-        image,
-        sd_version,
-        sd_config,
-        device,
-        dtype,
-        maybe_seed,
-        scheduler_steps,
-        sdxl_turbo_vae_path,
-        sdxl_turbo_unet_path,
-        clip_json_path,
-        clip_weights_path,
-    } = args;
-    
+pub fn stable_diffusion_pipeline(args: Args<'_>) -> Result<RgbImage> {
     println!("Starting image generation with the following configuration:");
-    println!("  Model: {:?}", sd_version);
-    println!("  Prompt: {}", prompt);
-    println!("  Steps: {}", scheduler_steps);
-    println!("  Device: {:?}", device);
+    println!("  Model: {:?}", args.configs.sd_version);
+    println!("  Prompt: {}", args.prompt);
+    println!("  Steps: {}", args.configs.scheduler_steps);
+    println!("  Device: {:?}", args.configs.device);
 
-    println!("Model dimensions: {}x{}", sd_config.width, sd_config.height);
+    println!("Model dimensions: {}x{}", args.configs.sd_config.width, args.configs.sd_config.height);
 
     // TODO(bt,2025-02-18): The scheduler is `EulerAncestralDiscreteScheduler`, but we may want to port an LCM scheduler.
     //  This is a target for performance improvement
     //  See: https://github.com/huggingface/candle/issues/1331
-    let mut scheduler = sd_config.build_scheduler(scheduler_steps)?;
+    let mut scheduler = args.configs.sd_config.build_scheduler(
+        args.configs.scheduler_steps)?;
 
-    let seed = maybe_seed.unwrap_or_else(|| rand::thread_rng().gen());
-    
-    device.set_seed(seed)?;
+    let seed = args.configs.seed.unwrap_or_else(|| rand::thread_rng().gen());
+    args.configs.device.set_seed(seed)?;
 
-    info!("Using seed: {}", seed);
+    println!("Using seed: {}", seed);
 
-    let guidance_scale = match cfg_scale {
+    let guidance_scale = match args.cfg_scale {
         Some(guidance_scale) => guidance_scale,
-        None => match sd_version {
+        None => match args.configs.sd_version {
             StableDiffusionVersion::V1_5
             | StableDiffusionVersion::V2_1
             | StableDiffusionVersion::Xl => 7.5,
@@ -91,32 +59,29 @@ pub fn stable_diffusion_pipeline<P1, P2, P3, P4>(args: Args<'_, P1, P2, P3, P4>)
 
     let use_guide_scale = guidance_scale > 1.0;
 
-    info!("Using guide scale: {}", use_guide_scale);
+    println!("Using guide scale: {}", use_guide_scale);
 
     info!("Checking if prompt is cached");
-    let maybe_cached = prompt_cache.get_copy(&prompt)?;
+    let maybe_cached = args.prompt_cache.get_copy(&args.prompt)?;
 
     let mut text_embeddings = if let Some(tensor) = maybe_cached {
         tensor
     } else {
-        
         info!("Prompt is NOT cached! Calculating embedding...");
-        let tensor = infer_clip_text_embeddings(ClipArgs {
-            prompt: &prompt,
-            uncond_prompt: &uncond_prompt,
-            tokenizer: None, // tokenizer
-            clip_weights: None, // clip_weights
-            clip2_weights: None, // clip2_weights
-            sd_version: sd_version,
-            sd_config: &sd_config,
-            use_f16: false, // use_f16
-            device: device,
-            dtype: dtype,
+        let tensor = infer_clip_text_embeddings(
+            &args.prompt,
+            &args.uncond_prompt,
+            None, // tokenizer
+            None, // clip_weights
+            None, // clip2_weights
+            args.configs.sd_version,
+            &args.configs.sd_config,
+            false, // use_f16
+            &args.configs.device,
+            args.configs.dtype,
             use_guide_scale,
-            clip_json_path: clip_json_path.as_ref(),
-            clip_weights_path: clip_weights_path.as_ref(),
-        })?;
-        prompt_cache.store_copy(&prompt, &tensor)?;
+        )?;
+        args.prompt_cache.store_copy(&args.prompt, &tensor)?;
         tensor
     };
 
@@ -124,73 +89,37 @@ pub fn stable_diffusion_pipeline<P1, P2, P3, P4>(args: Args<'_, P1, P2, P3, P4>)
 
     println!("Loading input image into tensor...");
 
-    let input_image = dynamic_image_to_tensor(image, device, dtype)?;
+    let input_image = dynamic_image_to_tensor(
+        args.image,
+        &args.configs.device,
+        args.configs.dtype)?;
 
     println!("Reference image shape: {:?}", input_image.shape());
 
-    let vae_scale = get_vae_scale(sd_version);
+    let vae_scale = get_vae_scale(args.configs.sd_version);
 
-    let maybe_vae = model_cache.get_vae()?;
-    
-    let vae = match maybe_vae {
-        Some(vae) => vae,
-        None => {
-            info!("No vae found in cache; loading...");
-            
-            let vae_file = sdxl_turbo_vae_path.as_ref();
-            
-            let vae = sd_config.build_vae(vae_file, &device, dtype)?;
-
-            let vae = Arc::new(vae);
-            
-            model_cache.set_vae(vae.clone())?;
-
-            vae
-        }
-    };
-
-    let maybe_unet = model_cache.get_unet()?;
-
-    let unet = match maybe_unet {
-        Some(unet) => unet,
-        None => {
-            info!("No unet found in cache; loading...");
-
-            let unet_weights_file = sdxl_turbo_unet_path.as_ref();
-
-            let unet = sd_config.build_unet(unet_weights_file, &device, 4, true, dtype)
-              .map_err(|err| anyhow!("error initializing unet model: {:?}", err))?;
-
-            let unet = Arc::new(unet);
-
-            model_cache.set_unet(unet.clone())?;
-
-            unet
-        }
-    };
-
-    let init_latent_dist : DiagonalGaussianDistribution = vae.encode(&input_image)?;
+    let init_latent_dist : DiagonalGaussianDistribution = args.model_cache.vae_encode(&input_image)?;
 
     // TODO(bt,2025-02-18): This takes a little bit to generate the sample.
     //  This is a target for performance improvement
     println!("Generating latents from input image...");
     let latents = (init_latent_dist.sample()? * vae_scale)?;
     
-    //.to_device(&configs.device)?;
+    //.to_device(&args.configs.device)?;
     
     println!("Initial latents shape: {:?}", latents.shape());
 
     println!("Calculating start step for diffusion process...");
 
-    let img2img_strength = match i2i_strength {
+    let img2img_strength = match args.i2i_strength {
         None => 0.75f64,
         Some(strength) => (strength as f64) / 100.0f64,
     };
 
     let t_start = {
-        let start = scheduler_steps - (scheduler_steps as f64 * img2img_strength) as usize;
+        let start = args.configs.scheduler_steps - (args.configs.scheduler_steps as f64 * img2img_strength) as usize;
 
-        println!("Starting from step {} of {} (strength: {})", start, scheduler_steps, img2img_strength);
+        println!("Starting from step {} of {} (strength: {})", start, args.configs.scheduler_steps, img2img_strength);
         start
     };
 
@@ -229,7 +158,7 @@ pub fn stable_diffusion_pipeline<P1, P2, P3, P4>(args: Args<'_, P1, P2, P3, P4>)
 
         let latent_model_input = scheduler.scale_model_input(latent_model_input, timestep)?;
 
-        let mut noise_pred = match unet.forward(&latent_model_input, timestep as f64, &text_embeddings) {
+        let mut noise_pred = match args.model_cache.unet_inference(&latent_model_input, timestep as f64, &text_embeddings) {
             Ok(pred) => pred,
             Err(e) => {
                 println!("UNet inference failed with error: {}", e);
@@ -247,7 +176,7 @@ pub fn stable_diffusion_pipeline<P1, P2, P3, P4>(args: Args<'_, P1, P2, P3, P4>)
     }
 
     println!("Diffusion process completed, decoding image...");
-    let image = vae.decode(&(latents / vae_scale)?)?;
+    let image = args.model_cache.vae_decode(&(latents / vae_scale)?)?;
     
     println!("VAE decode completed");
 

crates/desktop/src-tauri/src/state/app_config.rs

@@ -1,14 +1,31 @@
-use crate::state::app_dir::AppDataRoot;
-use crate::state::os_platform::OsPlatform;
+use crate::ml::model_file::StableDiffusionVersion;
 use crate::state::yaml_config::YamlConfig;
-use crate::stubs::model_config::ModelConfig;
+use candle_core::{DType, Device};
+use candle_transformers::models::stable_diffusion::StableDiffusionConfig;
+use filesys::recursively_find_file_by_name::recursively_find_file_by_name;
+use hf_hub::api::sync::Api;
+use log::info;
 
-const DEFAULT_SD_IMAGE_WIDTH: usize = 1024;
-const DEFAULT_SD_IMAGE_HEIGHT: usize = 1024;
+const CONFIG_FILENAME : &str = "app_config.yaml";
 
+const DEFAULT_SD_IMAGE_WIDTH: usize = 512;
+const DEFAULT_SD_IMAGE_HEIGHT: usize = 512;
+
+const DEFAULT_SD_NUMERIC_DATATYPE: DType = DType::F32;
+const SD_VERSION: StableDiffusionVersion = StableDiffusionVersion::Turbo;
 
 /// A central place to configure app-wide details.
 pub struct AppConfig {
+  /// We only support one device type currently, and we can't multiplex 
+  /// over several devices yet.
+  pub device: Device,
+  
+  /// The core numeric type to use for models.
+  pub dtype: DType,
+  
+  pub sd_version: StableDiffusionVersion,
+  pub sd_config: StableDiffusionConfig,
+
   pub image_height: usize,
   pub image_width: usize,
   pub scheduler_steps: usize,
@@ -18,45 +35,62 @@ pub struct AppConfig {
 
   pub cfg_scale: Option<f64>,
   
-  /// Root location for application data.
-  pub app_data_root: AppDataRoot,
-  
-  pub model_config: ModelConfig,
+  /// Probably shouldn't be here.
+  pub hf_api: Api,
 }
 
 impl AppConfig {
   
   pub fn init() -> anyhow::Result<Self> {
-    let yaml_configs = YamlConfig::load_from_config_file_recursive();
+    println!("Configuration filename: {}", CONFIG_FILENAME);
+    let yaml_configs = load_yaml_configs();
     
     println!("Configs: {:?}", yaml_configs);
     
-    let os_platform = OsPlatform::get();
-
-    let default_data_root = match os_platform {
-      OsPlatform::Linux => yaml_configs.linux_default_data_path.as_deref(),
-      OsPlatform::MacOs => yaml_configs.mac_default_data_path.as_deref(),
-      OsPlatform::Windows => yaml_configs.windows_default_data_path.as_deref(),
+    let device = Device::new_cuda(0)
+      .unwrap_or_else(|e| {
+          info!("CUDA not available ({}), falling back to CPU", e);
+          Device::Cpu
+        });
+    
+    let sd_config = match SD_VERSION {
+      StableDiffusionVersion::Turbo => {
+        StableDiffusionConfig::sdxl_turbo(None, Some(DEFAULT_SD_IMAGE_HEIGHT), Some(DEFAULT_SD_IMAGE_WIDTH))
+      }
+      _ => {
+        StableDiffusionConfig::v2_1(None, Some(DEFAULT_SD_IMAGE_HEIGHT), Some(DEFAULT_SD_IMAGE_WIDTH))
+      }
     };
     
-    println!("Possible user-defined data root: {:?}", default_data_root);
-
-    let app_data_root = match default_data_root {
-      None => AppDataRoot::create_default()?,
-      Some(path) => AppDataRoot::create_existing(path)?,
-    };
-
-    println!("App data root: {:?}", app_data_root.path());
-
+    let hf_api = Api::new()?;
+    
     Ok(Self {
+      device,
+      dtype: DEFAULT_SD_NUMERIC_DATATYPE,
+      sd_version: SD_VERSION,
+      sd_config,
       image_height: yaml_configs.image_height.unwrap_or(DEFAULT_SD_IMAGE_HEIGHT),
       image_width: yaml_configs.image_width.unwrap_or(DEFAULT_SD_IMAGE_WIDTH),
-      scheduler_steps: yaml_configs.scheduler_steps.unwrap_or(4),
+      scheduler_steps: yaml_configs.scheduler_steps.unwrap_or(1),
       scheduler_samples: yaml_configs.scheduler_samples.unwrap_or(15),
       seed: yaml_configs.seed,
       cfg_scale: yaml_configs.cfg_scale,
-      app_data_root,
-      model_config: ModelConfig::init()?,
+      hf_api,
     })
   }
 }
+
+fn load_yaml_configs() -> YamlConfig {
+  let maybe_path = recursively_find_file_by_name(CONFIG_FILENAME, "../../../", 5)
+    .ok()
+    .flatten();
+  if let Some(path) = maybe_path {
+    let path = path.canonicalize().unwrap_or(path);
+    println!("Attempting to load configs from: {:?}", &path);
+    if let Ok(config) = YamlConfig::read_from_file(path) {
+      return config;
+    }
+  }
+  println!("Loading default configs.");
+  YamlConfig::default()
+}

crates/desktop/src-tauri/src/state/app_dir.rs

@@ -1,187 +0,0 @@
-use crate::state::expanduser::expanduser;
-use crate::state::os_platform::OsPlatform;
-use anyhow::anyhow;
-use directories::UserDirs;
-use ml_weights_registry::weights_registry::weight_descriptor::WeightDescriptor;
-use std::path::{Path, PathBuf};
-use tempdir::TempDir;
-use tempfile::{Builder, NamedTempFile};
-
-const DEFAULT_DATA_DIR : &str = "artcraft";
-const ASSETS_SUBDIRECTORY : &str = "assets";
-const WEIGHTS_SUBDIRECTORY : &str = "weights";
-
-const TEMPORARY_SUBDIRECTORY : &str = "temp";
-
-/// The path to the application data directory, which includes "asset" and "weights" data.
-#[derive(Clone)]
-pub struct AppDataRoot {
-  path: PathBuf,
-  assets_dir: AppAssetsDir,
-  weights_dir: AppWeightsDir,
-  temp_dir: TemporaryDir,
-}
-
-#[derive(Clone)]
-pub struct AppAssetsDir {
-  path: PathBuf,
-}
-
-#[derive(Clone)]
-pub struct AppWeightsDir {
-  path: PathBuf,
-}
-
-#[derive(Clone)]
-pub struct TemporaryDir {
-  path: PathBuf,
-}
-
-impl AppDataRoot {
-  pub fn create_default() -> anyhow::Result<Self> {
-    let directory = get_default_data_dir()?;
-    Self::create_existing(directory)
-  }
-
-  pub fn create_existing<P: AsRef<Path>>(dir: P) -> anyhow::Result<Self> {
-    let mut dir = dir.as_ref().to_path_buf();
-    
-    match OsPlatform::get() {
-      OsPlatform::Linux | OsPlatform::MacOs => {
-        if let Some(d) = dir.as_os_str().to_str() {
-          dir = expanduser(d)?;
-        }
-      },
-      OsPlatform::Windows => {}
-    }
-    
-    if !dir.is_dir() {
-      println!("Creating directory {:?}", dir);
-      std::fs::create_dir_all(&dir)?;
-    }
-
-    match dir.canonicalize() {
-      Ok(d) => dir = d,
-      Err(err) => {
-        println!("Error canonicalizing {:?}: {}", dir, err);
-      }
-    }
-    
-    let assets_dir = AppAssetsDir::create_existing(dir.join(ASSETS_SUBDIRECTORY))?;
-    let weights_dir = AppWeightsDir::create_existing(dir.join(WEIGHTS_SUBDIRECTORY))?;
-    let temp_dir = TemporaryDir::create_existing(dir.join(TEMPORARY_SUBDIRECTORY))?;
-
-    Ok(Self {
-      path: dir,
-      assets_dir,
-      weights_dir,
-      temp_dir,
-    })
-  }
-  
-  pub fn assets_dir(&self) -> &AppAssetsDir {
-    &self.assets_dir
-  }
-  
-  pub fn weights_dir(&self) -> &AppWeightsDir {
-    &self.weights_dir
-  }
-
-  pub fn temp_dir(&self) -> &TemporaryDir {
-    &self.temp_dir
-  }
-
-  pub fn path(&self) -> &Path {
-    &self.path
-  }
-}
-
-impl AppAssetsDir {
-  pub fn create_existing<P: AsRef<Path>>(dir: P) -> anyhow::Result<Self> {
-    let mut dir = dir.as_ref().to_path_buf();
-    match dir.canonicalize() {
-      Ok(d) => dir = d,
-      Err(err) => {
-        println!("Error canonicalizing {:?}: {}", dir, err);
-      }
-    }
-    if !dir.exists() {
-      println!("Creating directory {:?}", dir);
-      std::fs::create_dir(&dir)?;
-    }
-    Ok(Self {
-      path: dir,
-    })
-  }
-
-  pub fn path(&self) -> &Path {
-    &self.path
-  }
-}
-
-impl AppWeightsDir {
-  pub fn create_existing<P: AsRef<Path>>(dir: P) -> anyhow::Result<Self> {
-    let mut dir = dir.as_ref().to_path_buf();
-    match dir.canonicalize() {
-      Ok(d) => dir = d,
-      Err(err) => {
-        println!("Error canonicalizing {:?}: {}", dir, err);
-      }
-    }
-    if !dir.exists() {
-      println!("Creating directory {:?}", dir);
-      std::fs::create_dir(&dir)?;
-    }
-    Ok(Self {
-      path: dir,
-    })
-  }
-
-  pub fn path(&self) -> &Path {
-    &self.path
-  }
-
-  pub fn weight_path(&self, descriptor: &WeightDescriptor) -> PathBuf {
-    self.path.join(descriptor.filename)
-  }
-}
-
-impl TemporaryDir {
-  pub fn create_existing<P: AsRef<Path>>(dir: P) -> anyhow::Result<Self> {
-    let mut dir = dir.as_ref().to_path_buf();
-    match dir.canonicalize() {
-      Ok(d) => dir = d,
-      Err(err) => {
-        println!("Error canonicalizing {:?}: {}", dir, err);
-      }
-    }
-    if !dir.exists() {
-      println!("Creating directory {:?}", dir);
-      std::fs::create_dir(&dir)?;
-    }
-    Ok(Self {
-      path: dir,
-    })
-  }
-
-  pub fn with_prefix(&self, prefix: &str) -> anyhow::Result<TempDir> {
-    let tempdir = TempDir::new_in(&self.path, prefix)?;
-    Ok(tempdir)
-  }
-
-  pub fn new_named_temp_file(&self) -> anyhow::Result<NamedTempFile> {
-    let tempfile = Builder::new()
-      .prefix("temp_")
-      .suffix(".bin")
-      .tempfile_in(&self.path)?;
-    Ok(tempfile)
-  }
-}
-
-// eg. /home/bob/artcraft, /Users/bob/artcraft, or C:\Users\Alice\artcraft
-fn get_default_data_dir() -> anyhow::Result<PathBuf> {
-  Ok(UserDirs::new()
-      .ok_or_else(|| anyhow!("could not determine user home directory"))?
-      .home_dir()
-      .join(DEFAULT_DATA_DIR))
-}

crates/desktop/src-tauri/src/state/expanduser.rs

@@ -1,73 +0,0 @@
-use once_cell::sync::Lazy;
-use std::io;
-use std::path::{PathBuf, MAIN_SEPARATOR};
-
-static PREFIX: Lazy<String> = Lazy::new(|| format!("~{}", MAIN_SEPARATOR));
-
-/// The "expanduser" crate doesn't compile on Windows, so we replace its functionality slightly
-pub fn expanduser<P: AsRef<str>>(path: P) -> io::Result<PathBuf> {
-
-   Ok(match path.as_ref() {
-       // matches an exact "~"
-       s if s == "~" => {
-           home_dir()?
-       },
-       // matches paths that start with `~/`
-       s if s.starts_with(&*PREFIX) => {
-           let home = home_dir()?;
-           home.join(&s[2..])
-       },
-       // // matches paths that start with `~` but not `~/`, might be a `~username/` path
-       // s if s.starts_with("~") => {
-       //     let mut parts = s[1..].splitn(2, MAIN_SEPARATOR);
-       //     let user = parts.next()
-       //         .ok_or_else(|| io::Error::new(io::ErrorKind::Other, "malformed path"))?;
-       //     let user = Passwd::from_name(&user)
-       //         .map_err(|_| io::Error::new(io::ErrorKind::Other, "error searching for user"))?
-       //         .ok_or_else(|| io::Error::new(io::ErrorKind::Other, format!("user '{}', does not exist", &user)))?;
-       //     if let Some(ref path) = parts.next() {
-       //         PathBuf::from(user.dir).join(&path)
-       //     } else {
-       //         PathBuf::from(user.dir)
-       //     }
-       // },
-       // nothing to expand, just make a PathBuf
-       s => PathBuf::from(s)
-   })
-}
-
-pub fn home_dir() -> io::Result<PathBuf> {
-    dirs::home_dir()
-        .ok_or_else(|| io::Error::new(io::ErrorKind::Other, "no home directory is set"))
-}
-
-#[cfg(test)]
-mod tests {
-    use crate::state::expanduser::expanduser;
-    use std::env;
-    use std::path::PathBuf;
-
-    #[test]
-    #[serial_test::serial]
-    fn test_success() {
-        let old_home = env::var("HOME").expect("no home dir set");
-        let new_home = "/home/foo";
-        env::set_var("HOME", new_home);
-        let path = expanduser("~/path/to/directory");
-        env::set_var("HOME", old_home);
-        assert_eq!(path.expect("io error"), PathBuf::from("/home/foo/path/to/directory"));
-    }
-
-    #[test]
-    #[serial_test::serial]
-    fn test_only_tilde() {
-        let old_home = env::var("HOME").expect("no home dir set");
-        let new_home = "/home/foo";
-        env::set_var("HOME", new_home);
-        let pathstr = "~";
-        let path = expanduser(pathstr);
-        env::set_var("HOME", old_home);
-        assert_eq!(path.expect("io error"), PathBuf::from("/home/foo"));
-    }
-
-}

crates/desktop/src-tauri/src/state/mod.rs

@@ -1,5 +1,2 @@
-pub mod app_config;
-pub mod app_dir;
-pub mod os_platform;
 pub mod yaml_config;
-pub mod expanduser;
+pub mod app_config;

crates/desktop/src-tauri/src/state/os_platform.rs

@@ -1,20 +0,0 @@
-
-#[derive(Copy,Clone,Eq,PartialEq,Debug)]
-pub enum OsPlatform {
-  Linux,
-  MacOs,
-  Windows,
-}
-
-impl OsPlatform {
-  pub fn get() -> OsPlatform {
-    // Totally scientific.
-    // https://doc.rust-lang.org/std/env/consts/constant.OS.html
-    match std::env::consts::OS {
-      "linux" => OsPlatform::Linux,
-      "windows" => OsPlatform::Windows,
-      "macos" | "apple" => OsPlatform::MacOs,
-      _ => OsPlatform::Linux,
-    }
-  }
-}

crates/desktop/src-tauri/src/state/yaml_config.rs

@@ -1,10 +1,7 @@
-use filesys::recursively_find_file_by_name::recursively_find_file_by_name;
 use std::fs::File;
 use std::io::BufReader;
 use std::path::Path;
 
-const CONFIG_FILENAME : &str = "app_config.yaml";
-
 /// Configuration read or derived from a YAML config file at startup
 /// This is not meant to contain any runtime config values.
 /// This may not be exhaustive.
@@ -16,15 +13,6 @@ pub struct YamlConfig {
   pub scheduler_samples: Option<usize>,
   pub seed: Option<u64>,
   pub cfg_scale: Option<f64>,
-
-  /// Override the default location to store data on Linux
-  pub linux_default_data_path: Option<String>,
-
-  /// Override the default location to store data on Windows
-  pub windows_default_data_path: Option<String>,
-  
-  /// Override the default location to store data on Mac
-  pub mac_default_data_path: Option<String>,
 }
 
 impl YamlConfig {
@@ -34,19 +22,4 @@ impl YamlConfig {
     let config = serde_yaml::from_reader(reader)?;
     Ok(config)
   }
-
-  pub fn load_from_config_file_recursive() -> YamlConfig {
-    let maybe_path = recursively_find_file_by_name(CONFIG_FILENAME, "../../../", 5)
-      .ok()
-      .flatten();
-    if let Some(path) = maybe_path {
-      let path = path.canonicalize().unwrap_or(path);
-      println!("Attempting to load configs from: {:?}", &path);
-      if let Ok(config) = YamlConfig::read_from_file(path) {
-        return config;
-      }
-    }
-    println!("Couldn't load config file. Using default configs.");
-    YamlConfig::default()
-  }
 }

crates/desktop/src-tauri/src/stubs/mod.rs

@@ -1,3 +0,0 @@
-pub mod model_cache;
-pub mod model_config;
-pub mod prompt_cache;

crates/desktop/src-tauri/src/stubs/model_cache.rs

@@ -1,14 +0,0 @@
-#[cfg(not(target_os = "macos"))]
-pub use ml_models::ml::model_cache::ModelCache;
-
-#[cfg(not(not(target_os = "macos")))]
-pub struct ModelCache {
-  // Intentionally left blank
-}
-
-#[cfg(not(not(target_os = "macos")))]
-impl ModelCache {
-  pub fn new() -> Self {
-    Self {}
-  }
-}

crates/desktop/src-tauri/src/stubs/model_config.rs

@@ -1,14 +0,0 @@
-#[cfg(not(target_os = "macos"))]
-pub use ml_models::ml::model_config::ModelConfig;
-
-#[cfg(not(not(target_os = "macos")))]
-pub struct ModelConfig {
-  // Intentionally left blank
-}
-
-#[cfg(not(not(target_os = "macos")))]
-impl ModelConfig {
-  pub fn init() -> anyhow::Result<Self> {
-    Ok(Self {})
-  }
-}

crates/desktop/src-tauri/src/stubs/prompt_cache.rs

@@ -1,15 +0,0 @@
-
-#[cfg(not(target_os = "macos"))]
-pub use ml_models::ml::prompt_cache::PromptCache;
-
-#[cfg(not(not(target_os = "macos")))]
-pub struct PromptCache {
-  // Intentionally left blank
-}
-
-#[cfg(not(not(target_os = "macos")))]
-impl PromptCache {
-  pub fn with_capacity(capacity: usize) -> anyhow::Result<Self> {
-    Ok(Self {})
-  }
-}

crates/desktop/src-tauri/src/threads/downloader_thread.rs

@@ -1,181 +0,0 @@
-use crate::events::notification_event::{NotificationEvent, NotificationModelType};
-use crate::state::app_dir::AppDataRoot;
-use crate::transfer::download::{download_async, ProgressUpdate};
-use futures::StreamExt;
-use futures::stream::FuturesUnordered;
-use log::{debug, error, info};
-use ml_weights_registry::weights_registry::weight_descriptor::WeightFunction;
-use ml_weights_registry::weights_registry::weights::{CLIP_JSON, DIS_MEDIUM_ONNX, LYKON_DEAMSHAPER_7_TEXT_ENCODER_FP16, LYKON_DEAMSHAPER_7_VAE, SDXL_TURBO_CLIP_TEXT_ENCODER, SDXL_TURBO_CLIP_TEXT_ENCODER_2, SDXL_TURBO_UNET, SDXL_TURBO_VAE, SIMIANLUO_LCM_DREAMSHAPER_V7_UNET};
-use std::collections::VecDeque;
-use std::path::PathBuf;
-use std::sync::mpsc::Receiver;
-use std::sync::{Arc, Mutex};
-use tauri::{AppHandle, Emitter};
-use tempfile::NamedTempFile;
-use tokio::task::JoinHandle;
-
-const NOTIFICATION_CHANNEL_NAME : &str = "notification";
-const MAX_FILES : usize = 8;
-const CHUNK_SIZE : usize = 1024 * 1024;
-const PARALLEL_FAILURES : usize = 8;
-const MAX_RETRIES : usize = 30;
-
-struct ModelDownloadStatusChannel {
-  pub model_name: String,
-  pub model_type: NotificationModelType,
-  pub final_file_path: PathBuf,
-  pub receiver: Receiver<ProgressUpdate>,
-}
-
-pub async fn downloader_thread(app_data_root: AppDataRoot, app: AppHandle) -> ! {
-
-  let mut status_check_queue : Arc<Mutex<Vec<ModelDownloadStatusChannel>>> = Arc::new(Mutex::new(Vec::new()));
-
-  {
-    let status_check_queue2 = status_check_queue.clone();
-    let app2 = app.clone();
-
-    tokio::spawn(download_notify_thread(status_check_queue2, app2));
-  }
-
-  let mut download_queue = VecDeque::new();
-
-  download_queue.push_back(CLIP_JSON);
-  download_queue.push_back(LYKON_DEAMSHAPER_7_TEXT_ENCODER_FP16);
-  download_queue.push_back(LYKON_DEAMSHAPER_7_VAE);
-  download_queue.push_back(SIMIANLUO_LCM_DREAMSHAPER_V7_UNET);
-  download_queue.push_back(SDXL_TURBO_CLIP_TEXT_ENCODER); // TODO(bt): Why is this still needed?
-  download_queue.push_back(DIS_MEDIUM_ONNX);
-  //download_queue.push_back(SDXL_TURBO_CLIP_TEXT_ENCODER_2); // TODO(bt): REMOVE
-  //download_queue.push_back(SDXL_TURBO_VAE); // TODO(bt): REMOVE
-  //download_queue.push_back(SDXL_TURBO_UNET); // TODO(bt): REMOVE
-
-  let mut handles = FuturesUnordered::new();
-
-  while let Some(model) = download_queue.pop_front() {
-    let url = model.r2_download_url.to_string();
-
-    let filename = app_data_root.weights_dir().weight_path(&model);
-
-    if filename.exists() {
-      info!("File already exists: {:?}", &filename);
-      continue;
-    }
-
-    let headers = None;
-    
-    let (tx, rx) = std::sync::mpsc::channel();
-
-    info!("Download: {:?}", filename);
-
-    let temp_file = match app_data_root.temp_dir().new_named_temp_file() {
-      Ok(file) => file,
-      Err(err) => {
-        error!("Couldn't create a temporary file: {:?}", err);
-        continue;
-      }
-    };
-
-    let model_type = match model.function {
-      WeightFunction::TextEncoder => NotificationModelType::TextEncoder,
-      WeightFunction::TextTokenizer => NotificationModelType::TextTokenizer,
-      WeightFunction::Unet => NotificationModelType::Unet,
-      WeightFunction::Vae => NotificationModelType::Vae,
-      WeightFunction::ImageSegmentation => NotificationModelType::ImageSegmentation,
-    };
-
-    let result = app.emit(NOTIFICATION_CHANNEL_NAME, NotificationEvent::ModelDownloadStarted {
-      model_name: model.name,
-      model_type,
-    });
-
-    if let Err(err) = result {
-      error!("Could not emit event: {}", err);
-    }
-
-    let task = tokio::spawn(download_async(
-      url,
-      temp_file,
-      filename.clone(),
-      MAX_FILES,
-      CHUNK_SIZE,
-      PARALLEL_FAILURES,
-      MAX_RETRIES,
-      headers,
-      Some(tx),
-    ));
-
-    match status_check_queue.lock() {
-      Err(err) => error!("Could not unlock: {:?}", err),
-      Ok(mut lock) => {
-        lock.push(ModelDownloadStatusChannel {
-          receiver: rx,
-          final_file_path: filename.clone(),
-          model_name: model.to_descriptive_name(),
-          model_type,
-        })
-      }
-    }
-
-    handles.push(task);
-  }
-
-  while let Some(result) = handles.next().await {
-    match result {
-      Err(err) => error!("Error downloading: {:?}", err),
-      Ok(_) => {},
-    }
-  }
-
-  loop {
-    // TODO(bt,2025-03-10): Queue management for new files entering the queue.
-    tokio::time::sleep(std::time::Duration::from_millis(60_000)).await;
-  }
-}
-
-pub async fn download_notify_thread(queue: Arc<Mutex<Vec<ModelDownloadStatusChannel>>>, app: AppHandle) -> ! {
-  loop {
-    match queue.lock() {
-      Err(err) => error!("Could not unlock: {:?}", err),
-      Ok(mut lock) => {
-        debug!(">>>>> Notifier has {} items", lock.len());
-
-        let mut reenqueue = Vec::with_capacity(lock.len());
-
-        for item in (*lock).drain(..) {
-          if let Ok(update) = item.receiver.try_recv() {
-            let result = app.emit(NOTIFICATION_CHANNEL_NAME, NotificationEvent::ModelDownloadProgress {
-              model_name: &item.model_name,
-              model_type: item.model_type,
-              currently_downloaded_bytes: update.complete,
-              total_file_bytes: update.total_length,
-            });
-
-            if let Err(err) = result {
-              error!("Could not emit event: {}", err);
-            }
-          }
-
-          if item.final_file_path.exists() {
-            debug!("No longer sending notifications for {:?}", &item.final_file_path);
-
-            let result = app.emit(NOTIFICATION_CHANNEL_NAME, NotificationEvent::ModelDownloadComplete {
-              model_name: &item.model_name,
-              model_type: item.model_type,
-            });
-
-            if let Err(err) = result {
-              error!("Could not emit event: {}", err);
-            }
-          } else {
-            reenqueue.push(item);
-          }
-        }
-
-        lock.extend(reenqueue);
-      }
-    }
-
-    tokio::time::sleep(std::time::Duration::from_millis(500)).await;
-  }
-}

crates/desktop/src-tauri/src/threads/mod.rs

@@ -1 +0,0 @@
-pub mod downloader_thread;

crates/desktop/src-tauri/src/transfer/download.rs

@@ -1,241 +0,0 @@
-//! Implemented from https://github.com/huggingface/hf_transfer/blob/main/src/lib.rs
-
-use crate::transfer::error::Error;
-use crate::transfer::util::{exponential_backoff, BASE_WAIT_TIME, MAX_WAIT_TIME};
-use futures::stream::FuturesUnordered;
-use futures::StreamExt;
-use reqwest::header::{
-  HeaderMap, HeaderName, HeaderValue, AUTHORIZATION, CONTENT_RANGE,
-  RANGE,
-};
-use reqwest::Url;
-use std::collections::HashMap;
-use std::fmt::Display;
-use std::fs::remove_file;
-use std::io::SeekFrom;
-use std::path::Path;
-use std::sync::Arc;
-use std::sync::atomic::{AtomicUsize, Ordering};
-use std::sync::mpsc::Sender;
-use std::time::Duration;
-use log::info;
-use tempfile::NamedTempFile;
-use tokio::fs::OpenOptions;
-use tokio::io::AsyncWriteExt;
-use tokio::io::AsyncSeekExt;
-use tokio::sync::Semaphore;
-use tokio::time::sleep;
-
-pub struct ProgressUpdate {
-  pub complete: usize,
-  pub total_length: usize,
-}
-
-/// max_files: Number of open file handles, which determines the maximum number of parallel downloads
-/// parallel_failures:  Number of maximum failures of different chunks in parallel (cannot exceed max_files)
-/// max_retries: Number of maximum attempts per chunk. (Retries are exponentially backed off + jitter)
-///
-/// The number of threads can be tuned by the environment variable `TOKIO_WORKER_THREADS` as documented in
-/// https://docs.rs/tokio/latest/tokio/runtime/struct.Builder.html#method.worker_threads
-#[allow(clippy::too_many_arguments)]
-pub async fn download_async<P: AsRef<Path>>(
-  url: String,
-  filename: NamedTempFile,
-  final_filename: P,
-  max_files: usize,
-  chunk_size: usize,
-  parallel_failures: usize,
-  max_retries: usize,
-  input_headers: Option<HashMap<String, String>>,
-  maybe_progress_sender: Option<Sender<ProgressUpdate>>,
-) -> Result<(), Error> {
-  info!(">>> Downloading: {}", filename.as_ref().display());
-
-  let client = reqwest::Client::builder()
-    // https://github.com/hyperium/hyper/issues/2136#issuecomment-589488526
-    .http2_keep_alive_timeout(Duration::from_secs(15))
-    .build()?;
-
-  let mut headers = HeaderMap::new();
-  let mut auth_token = None;
-  if let Some(input_headers) = input_headers {
-    headers.reserve(input_headers.len());
-    for (k, v) in input_headers {
-      let name: HeaderName = k
-        .try_into()
-        .map_err(|err| Error::new_err(format!("Invalid header: {err}")))?;
-      let value: HeaderValue = AsRef::<str>::as_ref(&v)
-        .try_into()
-        .map_err(|err| Error::new_err(format!("Invalid header value: {err}")))?;
-      if name == AUTHORIZATION {
-        auth_token = Some(value);
-      } else {
-        headers.insert(name, value);
-      }
-    }
-  };
-
-  let response = if let Some(token) = auth_token.as_ref() {
-    client.get(&url).header(AUTHORIZATION, token)
-  } else {
-    client.get(&url)
-  }
-    .headers(headers.clone())
-    .header(RANGE, "bytes=0-0")
-    .send()
-    .await
-    .map_err(|err| Error::new_err(format!("Error while downloading: {err}")))?
-    .error_for_status()
-    .map_err(|err| Error::new_err(err.to_string()))?;
-
-  // Only call the final redirect URL to avoid overloading the Hub with requests and also
-  // altering the download count
-  let redirected_url = response.url();
-  if Url::parse(&url)
-    .map_err(|err| Error::new_err(format!("failed to parse url: {err}")))?
-    .host()
-    == redirected_url.host()
-  {
-    if let Some(token) = auth_token {
-      headers.insert(AUTHORIZATION, token);
-    }
-  }
-
-  let content_range = response
-    .headers()
-    .get(CONTENT_RANGE)
-    .ok_or(Error::new_err("No content length".to_string()))?
-    .to_str()
-    .map_err(|err| Error::new_err(format!("Error while downloading: {err}")))?;
-
-  let size: Vec<&str> = content_range.split('/').collect();
-  // Content-Range: bytes 0-0/702517648
-  // https://developer.mozilla.org/en-US/docs/Web/HTTP/Headers/Content-Range
-  let length: usize = size
-    .last()
-    .ok_or(Error::new_err(
-      "Error while downloading: No size was detected".to_string(),
-    ))?
-    .parse()
-    .map_err(|err| Error::new_err(format!("Error while downloading: {err}")))?;
-
-
-  info!("File {:?} is {:?} length", filename.as_ref(), length);
-
-  let mut handles = FuturesUnordered::new();
-  let semaphore = Arc::new(Semaphore::new(max_files));
-  let parallel_failures_semaphore = Arc::new(Semaphore::new(parallel_failures));
-  let max_chunk_complete = Arc::new(AtomicUsize::new(0));
-
-  for start in (0..length).step_by(chunk_size) {
-    let url = redirected_url.to_string();
-    let filename = filename.as_ref().to_owned();
-    let client = client.clone();
-    let headers = headers.clone();
-
-    let stop = std::cmp::min(start + chunk_size - 1, length);
-    let semaphore = semaphore.clone();
-    let parallel_failures_semaphore = parallel_failures_semaphore.clone();
-    let maybe_progress = maybe_progress_sender.clone();
-    let max_chunk_complete = max_chunk_complete.clone();
-    handles.push(tokio::spawn(async move {
-      let permit = semaphore
-        .acquire_owned()
-        .await
-        .map_err(|err| Error::new_err(format!("Error while downloading: {err}")))?;
-      let mut chunk = download_chunk(&client, &url, &filename, start, stop, headers.clone()).await;
-      let mut i = 0;
-      if parallel_failures > 0 {
-        while let Err(dlerr) = chunk {
-          if i >= max_retries {
-            return Err(Error::new_err(format!(
-              "Failed after too many retries ({max_retries}): {dlerr}"
-            )));
-          }
-          let parallel_failure_permit = parallel_failures_semaphore.clone().try_acquire_owned().map_err(|err| {
-            Error::new_err(format!(
-              "Failed too many failures in parallel ({parallel_failures}): {dlerr} ({err})"
-            ))
-          })?;
-
-          let wait_time = exponential_backoff(BASE_WAIT_TIME, i, MAX_WAIT_TIME);
-          sleep(Duration::from_millis(wait_time as u64)).await;
-
-          chunk = download_chunk(&client, &url, &filename, start, stop, headers.clone()).await;
-          i += 1;
-          drop(parallel_failure_permit);
-          
-          if let Some(sender) = maybe_progress.as_ref() {
-            // NB: Calling fetch_max twice. First call will return the lower value.
-            let _old = max_chunk_complete.fetch_max(stop, Ordering::SeqCst);
-            let max_chunk = max_chunk_complete.fetch_max(stop, Ordering::SeqCst);
-            let progress_update = ProgressUpdate { 
-              complete: max_chunk,
-              total_length: length,
-            };
-            if let Err(_err) = sender.send(progress_update) {
-              // Fail silently, not essential to issue progress updates
-            }
-          }
-        }
-      }
-      drop(permit);
-      chunk.map_err(|e| Error::new_err(format!("Downloading error {e}"))).and(Ok(stop - start))
-    }));
-  }
-
-  // Output the chained result
-  while let Some(result) = handles.next().await {
-    match result {
-      Ok(Ok(size)) => {
-        //if let Some(ref callback) = callback {
-        //  callback.call((size,), None)?;
-        //}
-      }
-      Ok(Err(py_err)) => {
-        return Err(py_err);
-      }
-      Err(err) => {
-        return Err(Error::new_err(format!(
-          "Error while downloading: {err}"
-        )));
-      }
-    }
-  }
-
-  let file = filename.persist(final_filename)
-    .map_err(|e| Error::new_err(format!("Could not rename to final filename: {:?}", e)))?;
-
-  file.sync_all()?;
-
-  Ok(())
-}
-
-async fn download_chunk<P: AsRef<Path>>(
-  client: &reqwest::Client,
-  url: &str,
-  filename: P,
-  start: usize,
-  stop: usize,
-  headers: HeaderMap,
-) -> Result<(), Error> {
-  // Process each socket concurrently.
-  let range = format!("bytes={start}-{stop}");
-  let mut file = OpenOptions::new()
-    .write(true)
-    .truncate(false)
-    .create(true)
-    .open(filename)
-    .await?;
-  file.seek(SeekFrom::Start(start as u64)).await?;
-  let response = client
-    .get(url)
-    .headers(headers)
-    .header(RANGE, range)
-    .send()
-    .await?
-    .error_for_status()?;
-  let content = response.bytes().await?;
-  file.write_all(&content).await?;
-  Ok(())
-}

crates/desktop/src-tauri/src/transfer/error.rs

@@ -1,49 +0,0 @@
-//! Implemented from https://github.com/huggingface/hf_transfer/blob/main/src/lib.rs
-
-use std::fmt::Display;
-use reqwest::header::ToStrError;
-
-#[derive(Debug)]
-pub enum Error {
-  Io(std::io::Error),
-  Request(reqwest::Error),
-  ToStrError(ToStrError),
-  Misc(String),
-}
-
-impl From<std::io::Error> for Error {
-  fn from(value: std::io::Error) -> Self {
-    Self::Io(value)
-  }
-}
-
-impl From<reqwest::Error> for Error {
-  fn from(value: reqwest::Error) -> Self {
-    Self::Request(value)
-  }
-}
-
-impl From<ToStrError> for Error {
-  fn from(value: ToStrError) -> Self {
-    Self::ToStrError(value)
-  }
-}
-
-impl Display for Error {
-  fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-    match self {
-      Self::Io(io) => write!(f, "Io: {io}"),
-      Self::Request(req) => write!(f, "Request: {req}"),
-      Self::ToStrError(req) => write!(f, "Response non ascii: {req}"),
-      Self::Misc(err) => write!(f, "Misc error: {err}"),
-    }
-  }
-}
-
-impl std::error::Error for Error {}
-
-impl Error {
-  pub fn new_err(reason: String) -> Self {
-    Self::Misc(reason)
-  }
-}

crates/desktop/src-tauri/src/transfer/mod.rs

@@ -1,4 +0,0 @@
-pub mod download;
-pub mod upload;
-pub mod util;
-pub mod error;

crates/desktop/src-tauri/src/transfer/upload.rs

@@ -1,189 +0,0 @@
-//! Implemented from https://github.com/huggingface/hf_transfer/blob/main/src/lib.rs
-
-use crate::transfer::error::Error;
-use crate::transfer::util::{exponential_backoff, BASE_WAIT_TIME, MAX_WAIT_TIME};
-use futures::stream::FuturesUnordered;
-use futures::StreamExt;
-use reqwest::header::CONTENT_LENGTH;
-use std::collections::HashMap;
-use std::fmt::Display;
-use std::io::SeekFrom;
-use std::sync::Arc;
-use std::time::Duration;
-use tokio::fs::OpenOptions;
-use tokio::io::{AsyncReadExt, AsyncSeekExt};
-use tokio::sync::Semaphore;
-use tokio::time::sleep;
-use tokio_util::codec::{BytesCodec, FramedRead};
-
-/// parts_urls: Dictionary consisting of part numbers as keys and the associated url as values
-/// completion_url: The url that should be called when the upload is finished
-/// max_files: Number of open file handles, which determines the maximum number of parallel uploads
-/// parallel_failures:  Number of maximum failures of different chunks in parallel (cannot exceed max_files)
-/// max_retries: Number of maximum attempts per chunk. (Retries are exponentially backed off + jitter)
-///
-/// The number of threads can be tuned by the environment variable `TOKIO_WORKER_THREADS` as documented in
-/// https://docs.rs/tokio/latest/tokio/runtime/struct.Builder.html#method.worker_threads
-///
-/// See https://docs.aws.amazon.com/AmazonS3/latest/userguide/mpuoverview.html for more information
-/// on the multipart upload
-#[allow(clippy::too_many_arguments)]
-fn multipart_upload(
-  file_path: String,
-  parts_urls: Vec<String>,
-  chunk_size: u64,
-  max_files: usize,
-  parallel_failures: usize,
-  max_retries: usize,
-) -> Result<Vec<HashMap<String, String>>, Error> {
-  if parallel_failures > max_files {
-    return Err(Error::new_err(
-      "Error parallel_failures cannot be > max_files".to_string(),
-    ));
-  }
-  if (parallel_failures == 0) != (max_retries == 0) {
-    return Err(Error::new_err(
-      "For retry mechanism you need to set both `parallel_failures` and `max_retries`"
-        .to_string(),
-    ));
-  }
-
-  tokio::runtime::Builder::new_multi_thread()
-    .enable_all()
-    .build()?
-    .block_on(async {
-      upload_async(
-        file_path,
-        parts_urls,
-        chunk_size,
-        max_files,
-        parallel_failures,
-        max_retries,
-      )
-        .await
-    })
-}
-
-
-
-#[allow(clippy::too_many_arguments)]
-async fn upload_async(
-  file_path: String,
-  parts_urls: Vec<String>,
-  chunk_size: u64,
-  max_files: usize,
-  parallel_failures: usize,
-  max_retries: usize,
-) -> Result<Vec<HashMap<String, String>>, Error> {
-  let client = reqwest::Client::new();
-
-  let mut handles = FuturesUnordered::new();
-  let semaphore = Arc::new(Semaphore::new(max_files));
-  let parallel_failures_semaphore = Arc::new(Semaphore::new(parallel_failures));
-
-  for (part_number, part_url) in parts_urls.iter().enumerate() {
-    let url = part_url.to_string();
-    let path = file_path.to_owned();
-    let client = client.clone();
-
-    let start = (part_number as u64) * chunk_size;
-    let semaphore = semaphore.clone();
-    let parallel_failures_semaphore = parallel_failures_semaphore.clone();
-    handles.push(tokio::spawn(async move {
-      let permit = semaphore
-        .clone()
-        .acquire_owned()
-        .await
-        .map_err(|err| Error::new_err(format!("Error acquiring semaphore: {err}")))?;
-      let mut chunk = upload_chunk(&client, &url, &path, start, chunk_size).await;
-      let mut i = 0;
-      if parallel_failures > 0 {
-        while let Err(ul_err) = chunk {
-          if i >= max_retries {
-            return Err(Error::new_err(format!(
-              "Failed after too many retries ({max_retries}): {ul_err}"
-            )));
-          }
-
-          let parallel_failure_permit = parallel_failures_semaphore.clone().try_acquire_owned().map_err(|err| {
-            Error::new_err(format!(
-              "Failed too many failures in parallel ({parallel_failures}): {ul_err} ({err})"
-            ))
-          })?;
-
-          let wait_time = exponential_backoff(BASE_WAIT_TIME, i, MAX_WAIT_TIME);
-          sleep(Duration::from_millis(wait_time as u64)).await;
-
-          chunk = upload_chunk(&client, &url, &path, start, chunk_size).await;
-          i += 1;
-          drop(parallel_failure_permit);
-        }
-      }
-      drop(permit);
-      chunk.map_err(|e|{
-        //match e {
-        //  Error::Io(io) => Error::new_err(format!("Io error {io}")),
-        //  Error::Request(req) => Error::new_err(format!("Error while sending chunk {req}")),
-        //  Error::ToStrError(req) => Error::new_err(format!("Response header contains non ASCII chars: {req}")),
-        //}
-        e
-      }).map(|chunk| (part_number, chunk, chunk_size))
-    }));
-  }
-
-  let mut results: Vec<HashMap<String, String>> = vec![HashMap::default(); parts_urls.len()];
-
-  while let Some(result) = handles.next().await {
-    match result {
-      Ok(Ok((part_number, headers, size))) => {
-        //if let Some(ref callback) = callback {
-        //  callback.call((size,), None)?;
-        //}
-        results[part_number] = headers;
-      }
-      Ok(Err(py_err)) => {
-        return Err(py_err);
-      }
-      Err(err) => {
-        return Err(Error::new_err(format!(
-          "Error occurred while uploading: {err}"
-        )));
-      }
-    }
-  }
-
-  Ok(results)
-}
-
-async fn upload_chunk(
-  client: &reqwest::Client,
-  url: &str,
-  path: &str,
-  start: u64,
-  chunk_size: u64,
-) -> Result<HashMap<String, String>, Error> {
-  let mut options = OpenOptions::new();
-  let mut file = options.read(true).open(path).await?;
-  let file_size = file.metadata().await?.len();
-  let bytes_transferred = std::cmp::min(file_size - start, chunk_size);
-
-  file.seek(SeekFrom::Start(start)).await?;
-  let chunk = file.take(chunk_size);
-
-  let response = client
-    .put(url)
-    .header(CONTENT_LENGTH, bytes_transferred)
-    .body(reqwest::Body::wrap_stream(FramedRead::new(
-      chunk,
-      BytesCodec::new(),
-    )))
-    .send()
-    .await?;
-  let response = response.error_for_status()?;
-  let mut headers = HashMap::new();
-  for (name, value) in response.headers().into_iter() {
-    headers.insert(name.to_string(), value.to_str()?.to_owned());
-  }
-  Ok(headers)
-}
-

crates/desktop/src-tauri/src/transfer/util.rs

@@ -1,15 +0,0 @@
-//! Implemented from https://github.com/huggingface/hf_transfer/blob/main/src/lib.rs
-
-use rand::{thread_rng, Rng};
-use std::fmt::Display;
-
-pub const BASE_WAIT_TIME: usize = 300;
-pub const MAX_WAIT_TIME: usize = 10_000;
-
-fn jitter() -> usize {
-  thread_rng().gen_range(0..=500)
-}
-
-pub fn exponential_backoff(base_wait_time: usize, n: usize, max: usize) -> usize {
-  (base_wait_time + n.pow(2) + jitter()).min(max)
-}

crates/desktop/src-tauri/src/utils/image/decode_base64_image.rs

@@ -1,14 +0,0 @@
-use base64::prelude::BASE64_STANDARD;
-use base64::Engine;
-use image::{DynamicImage, ImageReader};
-use std::io::Cursor;
-
-pub fn decode_base64_image(base64_image: &str) -> anyhow::Result<DynamicImage> {
-  let bytes = BASE64_STANDARD.decode(base64_image)?;
-
-  let image = ImageReader::new(Cursor::new(bytes))
-    .with_guessed_format()?
-    .decode()?;
-
-  Ok(image)
-}

crates/desktop/src-tauri/src/utils/image/encode_dynamic_image_base64_png.rs

@@ -1,14 +0,0 @@
-use base64::prelude::BASE64_STANDARD;
-use base64::Engine;
-use image::{DynamicImage, ImageFormat};
-use std::io::Cursor;
-
-pub fn encode_dynamic_image_base64_png(image: DynamicImage) -> anyhow::Result<String> {
-  let mut bytes = Vec::with_capacity(1024*1024);
-
-  image.write_to(&mut Cursor::new(&mut bytes), ImageFormat::Png)?;
-
-  let bytes = BASE64_STANDARD.encode(bytes);
-
-  Ok(bytes)
-}

crates/desktop/src-tauri/src/utils/image/encode_rgb_image_base64_png.rs

@@ -1,16 +0,0 @@
-use base64::prelude::BASE64_STANDARD;
-use base64::Engine;
-use image::{ImageFormat, RgbImage};
-use std::io::Cursor;
-
-pub fn encode_rgb_image_base64_png(
-  image: RgbImage
-) -> anyhow::Result<String> {
-  let mut bytes = Vec::with_capacity(1024*1024);
-
-  image.write_to(&mut Cursor::new(&mut bytes), ImageFormat::Png)?;
-
-  let bytes = BASE64_STANDARD.encode(bytes);
-
-  Ok(bytes)
-}

crates/desktop/src-tauri/src/utils/image/mod.rs

@@ -1,3 +0,0 @@
-pub mod decode_base64_image;
-pub mod encode_dynamic_image_base64_png;
-pub mod encode_rgb_image_base64_png;

crates/desktop/src-tauri/src/utils/log_environment_details.rs

@@ -1,11 +0,0 @@
-use log::info;
-//use nvidia_checker::{get_cuda_version, get_cudnn_version, get_kernel_version, get_nvidia_driver_version, get_os_version, get_tensorrt_version};
-
-pub fn log_environment_details() {
-  //info!("NVIDIA driver version: {}", get_nvidia_driver_version(false));
-  //info!("CUDA version: {}", get_cuda_version(false));
-  //info!("cuDNN version: {}", get_cudnn_version(false));
-  //info!("TensorRT version: {}", get_tensorrt_version(false));
-  //info!("OS version: {}", get_os_version(false));
-  //info!("Kernel version: {}", get_kernel_version(false));
-}

crates/desktop/src-tauri/src/utils/mod.rs

@@ -1,2 +0,0 @@
-pub mod image;
-pub mod log_environment_details;

crates/desktop/src-tauri/tauri.conf.json

@@ -1,6 +1,6 @@
 {
   "$schema": "https://schema.tauri.app/config/2.0.0-rc",
-  "productName": "ArtCraft",
+  "productName": "Realtime Desktop App",
   "version": "0.1.0",
   "identifier": "com.unique.value.thing.why",
   "build": {
@@ -12,9 +12,9 @@
   "app": {
     "windows": [
       {
-        "title": "ArtCraft",
-        "width": 2400,
-        "height": 1300,
+        "title": "Realtime Desktop App Title Goes Here",
+        "width": 1200,
+        "height": 800,
         "resizable": true,
         "fullscreen": false
       }

crates/desktop/src-tauri/tests/test_dynamic_library_loading.rs

@@ -1,7 +0,0 @@
-
-#[test]
-fn test_dynamic_library_loading() {
-  let name = libloading::library_filename("cublas");
-
-  assert_eq!(name, "libcublas.so");
-}

crates/lib/data/primitives/Cargo.toml

@@ -32,4 +32,3 @@ my-workspace-hack = { version = "0.1", path = "../../../my-workspace-hack" }
 
 [dev-dependencies]
 testing = { path = "../../testing" }
-assertor = "0.0.3"

crates/lib/data/primitives/src/numerics/mod.rs

@@ -1,6 +1,5 @@
 pub mod i32_to_u32_zero_clamped;
 pub mod i64_to_u64_zero_clamped;
-pub mod remap_range_f64;
 pub mod u64_to_i32_saturating;
 pub mod u64_to_i64_saturating;
 pub mod u64_to_u32_saturating;

crates/lib/data/primitives/src/numerics/remap_range_f64.rs

@@ -1,131 +0,0 @@
-
-pub struct Args {
-  pub old_min: f64,
-  pub old_max: f64,
-  pub new_min: f64,
-  pub new_max: f64,
-  
-  /// The value to remap to the new range.
-  pub value: f64,
-}
-
-pub fn remap_range_f64(args: Args) -> f64 {
-  let old_range = args.old_max - args.old_min;
-  let new_range = args.new_max -  args.new_min;
-
-  let new_value = (args.value -  args.old_min) * new_range / old_range + args.new_min;
-  
-  if new_value > args.new_max {
-    args.new_max
-  } else if new_value < args.new_min { 
-    args.new_min
-  } else {
-    new_value
-  }
-}
-
-#[cfg(test)]
-mod tests {
-  use crate::numerics::remap_range_f64::{remap_range_f64, Args};
-  use assertor::{assert_that, FloatAssertion};
-
-  #[test]
-  fn test_base_cases() {
-    assert_that!(remap_range_f64(Args {
-      old_min: 0.0,
-      old_max: 1.0,
-      new_min: 0.0,
-      new_max: 1.0,
-      value: 0.0,
-    })).with_abs_tol(0.01).is_approx_equal_to(0.0);
-    
-    assert_that!(remap_range_f64(Args {
-      old_min: 0.0,
-      old_max: 1.0,
-      new_min: 0.0,
-      new_max: 1.0,
-      value: 1.0,
-    })).with_abs_tol(0.01).is_approx_equal_to(1.0);
-  }
-
-  #[test]
-  fn test_maxima() {
-    assert_that!(remap_range_f64(Args {
-      old_min: 0.0,
-      old_max: 100.0,
-      new_min: 0.0,
-      new_max: 1.0,
-      value: 100.0,
-    })).with_abs_tol(0.01).is_approx_equal_to(1.0);
-
-    assert_that!(remap_range_f64(Args {
-      old_min: 0.0,
-      old_max: 1.0,
-      new_min: 0.0,
-      new_max: 50.0,
-      value: 1.0,
-    })).with_abs_tol(0.01).is_approx_equal_to(50.0);
-
-    assert_that!(remap_range_f64(Args {
-      old_min: 0.0,
-      old_max: 1.0,
-      new_min: -100.0,
-      new_max: 100.0,
-      value: 1.0,
-    })).with_abs_tol(0.01).is_approx_equal_to(100.0);
-  }
-
-  #[test]
-  fn test_minima() {
-    assert_that!(remap_range_f64(Args {
-      old_min: 0.0,
-      old_max: 100.0,
-      new_min: 0.0,
-      new_max: 1.0,
-      value: 0.0,
-    })).with_abs_tol(0.01).is_approx_equal_to(0.0);
-
-    assert_that!(remap_range_f64(Args {
-      old_min: 0.0,
-      old_max: 1.0,
-      new_min: 0.0,
-      new_max: 50.0,
-      value: 0.0,
-    })).with_abs_tol(0.01).is_approx_equal_to(0.0);
-
-    assert_that!(remap_range_f64(Args {
-      old_min: 0.0,
-      old_max: 1.0,
-      new_min: -100.0,
-      new_max: 100.0,
-      value: 0.0,
-    })).with_abs_tol(0.01).is_approx_equal_to(-100.0);
-  }
-
-  #[test]
-  fn test_other_values() {
-    assert_that!(remap_range_f64(Args {
-      old_min: 0.0,
-      old_max: 100.0,
-      new_min: 0.0,
-      new_max: 1.0,
-      value: 50.0,
-    })).with_abs_tol(0.01).is_approx_equal_to(0.50);
-
-    assert_that!(remap_range_f64(Args {
-      old_min: 0.0,
-      old_max: 1.0,
-      new_min: 0.0,
-      new_max: 50.0,
-      value: 0.10,
-    })).with_abs_tol(0.01).is_approx_equal_to(5.0);
-
-    assert_that!(remap_range_f64(Args {
-      old_min: 0.0,
-      old_max: 1.0,
-      new_min: -100.0,
-      new_max: 100.0,
-      value: 0.25,
-    })).with_abs_tol(0.01).is_approx_equal_to(-50.0);
-  }
-}

crates/service/web/storyteller_web/src/configs/plans/plan_list.rs

@@ -118,40 +118,24 @@ pub static DEVELOPMENT_PREMIUM_PLANS: Lazy<HashSet<Plan>> = Lazy::new(|| {
 
 /// All paid-for production premium plans in the system.
 pub static PRODUCTION_PREMIUM_PLANS: Lazy<HashSet<Plan>> = Lazy::new(|| {
-  let mut plans = HashSet::new();
+    let mut plans = HashSet::new();
 
-  // ========== All-locale plans ==========
+    // ========== All-locale plans ==========
 
-  // Old "plus" price
-  plans.insert(PlanBuilder::new("fakeyou_plus")
-      .plan_category(PlanCategory::Paid)
-      .stripe_product_id("prod_MoLv23HrxPiY7D")
-      .stripe_price_id("price_1M4jDCEU5se17MekOaJ92HYX")
-      .cost_per_month_dollars(7)
-      .tts_base_priority_level(20)
-      .tts_max_character_length(2048)
-      .tts_max_duration_seconds(30)
-      .web_vc_base_priority_level(20)
-      .web_vc_requires_frontend_keepalive(false)
-      .lipsync_requires_frontend_keepalive(false)
-      .build());
+    plans.insert(PlanBuilder::new("fakeyou_plus")
+        .plan_category(PlanCategory::Paid)
+        .stripe_product_id("prod_MoLv23HrxPiY7D")
+        .stripe_price_id("price_1M4jDCEU5se17MekOaJ92HYX")
+        .cost_per_month_dollars(7)
+        .tts_base_priority_level(20)
+        .tts_max_character_length(2048)
+        .tts_max_duration_seconds(30)
+        .web_vc_base_priority_level(20)
+        .web_vc_requires_frontend_keepalive(false)
+        .lipsync_requires_frontend_keepalive(false)
+        .build());
 
-  // New "plus" price
-  plans.insert(PlanBuilder::new("fakeyou_plus_plan")
-      .plan_category(PlanCategory::Paid)
-      .stripe_product_id("prod_RvC5qUgLTSLGSD")
-      .stripe_price_id("price_1R1Lh8EU5se17MekaUNtWc59")
-      .cost_per_month_dollars(12)
-      .tts_base_priority_level(20)
-      .tts_max_character_length(2048)
-      .tts_max_duration_seconds(30)
-      .web_vc_base_priority_level(20)
-      .web_vc_requires_frontend_keepalive(false)
-      .lipsync_requires_frontend_keepalive(false)
-      .build());
-
-  // Old "pro" price
-  plans.insert(PlanBuilder::new("fakeyou_pro")
+    plans.insert(PlanBuilder::new("fakeyou_pro")
         .plan_category(PlanCategory::Paid)
         .stripe_product_id("prod_MoLw8nA6eFxHzc")
         .stripe_price_id("price_1M4jEQEU5se17MeksNfA0EKm")
@@ -165,23 +149,7 @@ pub static PRODUCTION_PREMIUM_PLANS: Lazy<HashSet<Plan>> = Lazy::new(|| {
         .can_remove_visual_watermarks(true)
         .build());
 
-  // New "pro" price
-  plans.insert(PlanBuilder::new("fakeyou_pro_plan")
-      .plan_category(PlanCategory::Paid)
-      .stripe_product_id("prod_RvC6nB9rw6yAsI")
-      .stripe_price_id("price_1R1LhzEU5se17Mek3JLwN3Gg")
-      .cost_per_month_dollars(25)
-      .tts_base_priority_level(30)
-      .tts_max_character_length(3072)
-      .tts_max_duration_seconds(60 * 2)
-      .web_vc_base_priority_level(30)
-      .web_vc_requires_frontend_keepalive(false)
-      .lipsync_requires_frontend_keepalive(false)
-      .can_remove_visual_watermarks(true)
-      .build());
-
-  // Old "elite" price
-  plans.insert(PlanBuilder::new("fakeyou_elite")
+    plans.insert(PlanBuilder::new("fakeyou_elite")
         .plan_category(PlanCategory::Paid)
         .stripe_product_id("prod_MoLxQmLA4R24fv")
         .stripe_price_id("price_1M4jFREU5se17Mekc8pQaSKB")
@@ -195,21 +163,6 @@ pub static PRODUCTION_PREMIUM_PLANS: Lazy<HashSet<Plan>> = Lazy::new(|| {
         .can_remove_visual_watermarks(true)
         .build());
 
-  // New "elite" price
-  plans.insert(PlanBuilder::new("fakeyou_elite_plan")
-      .plan_category(PlanCategory::Paid)
-      .stripe_product_id("prod_RvC6i84YiAYStd")
-      .stripe_price_id("price_1R1LiQEU5se17MekiJdBcsHQ")
-      .cost_per_month_dollars(40)
-      .tts_base_priority_level(40)
-      .tts_max_character_length(4096)
-      .tts_max_duration_seconds(60 * 5)
-      .web_vc_base_priority_level(40)
-      .web_vc_requires_frontend_keepalive(false)
-      .lipsync_requires_frontend_keepalive(false)
-      .can_remove_visual_watermarks(true)
-      .build());
-
     // ========== Spanish plans ==========
 
     plans.insert(PlanBuilder::new("fakeyou_basic")
@@ -340,7 +293,7 @@ mod test {
     // making ourselves check twice when we make additions, removals, or other changes.
     #[test]
     fn test_number_of_plans_is_expected() {
-        assert_eq!(14, ALL_PLANS_BY_SLUG.len());
+        assert_eq!(11, ALL_PLANS_BY_SLUG.len());
     }
 
     #[test]
@@ -350,7 +303,7 @@ mod test {
             .filter(|plan| !plan.is_synthetic_plan()) // NB: Synthetic plans are production!
             .collect::<Vec<&Plan>>();
 
-        let expected = 7;
+        let expected = 4;
 
         assert_eq!(expected, production_plans.len());
         assert_eq!(expected, PRODUCTION_PREMIUM_PLANS.len());
@@ -393,9 +346,6 @@ mod test {
             "prod_MoLw8nA6eFxHzc".to_string(), // NB: Production
             "prod_MoLxQmLA4R24fv".to_string(), // NB: Production
             "prod_MoLyt8qMDmscjr".to_string(), // NB: Production
-            "prod_RvC5qUgLTSLGSD".to_string(), // NB: Production - new "plus" product
-            "prod_RvC6nB9rw6yAsI".to_string(), // NB: Production - new "pro" product
-            "prod_RvC6i84YiAYStd".to_string(), // NB: Production - new "elite" product
         ];
 
         assert_that(&product_ids).contains_all_of(&expected_product_ids.iter());
@@ -416,9 +366,6 @@ mod test {
             "price_1M4jEQEU5se17MeksNfA0EKm".to_string(), // NB: Production
             "price_1M4jFREU5se17Mekc8pQaSKB".to_string(), // NB: Production
             "price_1M4jGPEU5se17Mek2FiztNE5".to_string(), // NB: Production
-            "price_1R1Lh8EU5se17MekaUNtWc59".to_string(), // NB: Production - new "plus" price
-            "price_1R1LhzEU5se17Mek3JLwN3Gg".to_string(), // NB: Production - new "pro" price
-            "price_1R1LiQEU5se17MekiJdBcsHQ".to_string(), // NB: Production - new "elite" price
         ];
 
         assert_that(&price_ids).contains_all_of(&expected_price_ids.iter());

crates/vendor/candle/.cargo/config.toml

@@ -1,8 +0,0 @@
-[build]
-rustflags = ["-C", "target-cpu=native"]
-
-[target.wasm32-unknown-unknown]
-rustflags = ["-C", "target-feature=+simd128"]
-
-[target.x86_64-apple-darwin]
-rustflags = ["-C", "target-feature=-avx,-avx2"]

crates/vendor/candle/.github/dependabot.yml

@@ -1,7 +0,0 @@
-version: 2
-updates:
-  - package-ecosystem: "cargo"
-    directory: "/"
-    schedule:
-      interval: "weekly"
-    open-pull-requests-limit: 5

crates/vendor/candle/.github/workflows/book-cd.yml

@@ -1,40 +0,0 @@
-name: Deploy Rust book
-on:
-  push:
-    branches:
-      - main
-
-jobs:
-  deploy:
-    runs-on: ubuntu-latest
-    permissions:
-      contents: write  # To push a branch 
-      pull-requests: write  # To create a PR from that branch
-    steps:
-    - uses: actions/checkout@v3
-      with:
-        fetch-depth: 0
-    - name: Install latest mdbook
-      run: |
-        tag=$(curl 'https://api.github.com/repos/rust-lang/mdbook/releases/latest' | jq -r '.tag_name')
-        url="https://github.com/rust-lang/mdbook/releases/download/${tag}/mdbook-${tag}-x86_64-unknown-linux-gnu.tar.gz"
-        mkdir mdbook
-        curl -sSL $url | tar -xz --directory=./mdbook
-        echo `pwd`/mdbook >> $GITHUB_PATH
-    - name: Deploy GitHub Pages
-      run: |
-        # This assumes your book is in the root of your repository.
-        # Just add a `cd` here if you need to change to another directory.
-        cd candle-book
-        mdbook build
-        git worktree add gh-pages
-        git config user.name "Deploy from CI"
-        git config user.email ""
-        cd gh-pages
-        # Delete the ref to avoid keeping history.
-        git update-ref -d refs/heads/gh-pages
-        rm -rf *
-        mv ../book/* .
-        git add .
-        git commit -m "Deploy $GITHUB_SHA to gh-pages"
-        git push --force --set-upstream origin gh-pages

crates/vendor/candle/.github/workflows/book.yml

@@ -1,29 +0,0 @@
-name: CI
-on: 
-  pull_request:
-
-jobs:
-  test:
-    name: Test candle-book
-    runs-on: ubuntu-latest
-    permissions:
-      contents: write  # To push a branch 
-      pull-requests: write  # To create a PR from that branch
-    steps:
-    - uses: actions/checkout@master
-    - name: Install Rust
-      run: |
-        rustup set profile minimal
-        rustup toolchain install stable
-        rustup default stable
-    - name: Install latest mdbook
-      run: |
-        tag=$(curl 'https://api.github.com/repos/rust-lang/mdbook/releases/latest' | jq -r '.tag_name')
-        url="https://github.com/rust-lang/mdbook/releases/download/${tag}/mdbook-${tag}-x86_64-unknown-linux-gnu.tar.gz"
-        mkdir bin
-        curl -sSL $url | tar -xz --directory=bin
-        echo "$(pwd)/bin" >> $GITHUB_PATH
-    - name: Run tests
-      run: cd candle-book && cargo build && mdbook test -L ../target/debug/deps/
-
-

crates/vendor/candle/.github/workflows/ci_cuda.yaml

@@ -1,34 +0,0 @@
-name: CI / cuda
-
-on:
-  workflow_dispatch:
-  pull_request:
-
-jobs:
-  test-cuda:
-    concurrency:
-      group: ${{ github.workflow }}-${{ github.job }}-${{ github.head_ref || github.run_id }}
-      cancel-in-progress: true
-    runs-on:
-      group: aws-g4dn-2xlarge
-    container:
-      image: nvidia/cuda:12.3.1-devel-ubuntu22.04
-      options: --gpus 0 
-    if: ${{ github.event.pull_request.head.repo.full_name == github.event.pull_request.base.repo.full_name }}
-    permissions:
-      contents: write
-      packages: write
-      # This is used to complete the identity challenge
-      # with sigstore/fulcio when running outside of PRs.
-      id-token: write
-      security-events: write
-    steps:
-      - name: Checkout repository
-        uses: actions/checkout@v3
-      - name: Install dependencies
-        run: apt-get update && apt install curl build-essential libssl-dev protobuf-compiler pkg-config -y
-      - name: Install Rust Stable
-        uses: actions-rust-lang/setup-rust-toolchain@v1
-      - uses: Swatinem/rust-cache@v2
-      - name: Test (cuda)
-        run: cargo test --features cuda

crates/vendor/candle/.github/workflows/python.yml

@@ -1,68 +0,0 @@
-name: PyO3-CI
-
-on:
-  workflow_dispatch:
-  push:
-    branches:
-      - main
-    paths:
-      - candle-pyo3/**
-  pull_request:
-    paths:
-      - candle-pyo3/**
-
-jobs:
-  build_and_test:
-    name: Check everything builds & tests
-    runs-on: ${{ matrix.os }}
-    strategy:
-      matrix:
-        os: [ubuntu-latest] # For now, only test on Linux
-    steps:
-      - name: Checkout repository
-        uses: actions/checkout@v4
-
-      - name: Install Rust
-        uses: actions-rs/toolchain@v1
-        with:
-          toolchain: stable
-
-      - name: Install Python
-        uses: actions/setup-python@v4
-        with:
-          python-version: 3.11
-          architecture: "x64"
-
-      - name: Cache Cargo Registry
-        uses: actions/cache@v1
-        with:
-          path: ~/.cargo/registry
-          key: ${{ runner.os }}-cargo-registry-${{ hashFiles('**/Cargo.lock') }}
-
-      - name: Install Protoc
-        uses: arduino/setup-protoc@v2
-        with:
-            version: "25.0"
-            repo-token: ${{ secrets.GITHUB_TOKEN }}
-
-      - name: Install
-        working-directory: ./candle-pyo3
-        run: |
-          python -m venv .env
-          source .env/bin/activate
-          pip install -U pip
-          pip install pytest maturin black
-          python -m maturin develop -r --features onnx
-
-      - name: Check style
-        working-directory: ./candle-pyo3
-        run: |
-          source .env/bin/activate
-          python stub.py --check
-          black --check .
-
-      - name: Run tests
-        working-directory: ./candle-pyo3
-        run: |
-          source .env/bin/activate
-          python -m pytest -s -v tests

crates/vendor/candle/.github/workflows/rust-ci.yml

@@ -1,87 +0,0 @@
-on:
-  push:
-    branches:
-      - main
-  pull_request:
-
-name: Continuous integration
-
-jobs:
-  check:
-    name: Check
-    runs-on: ${{ matrix.os }}
-    strategy:
-      matrix:
-        os: [ubuntu-latest, windows-latest, macOS-latest]
-        rust: [stable]
-    steps:
-      - uses: actions/checkout@v4
-      - uses: actions/setup-python@v5
-        with:
-          python-version: "3.11"
-      - uses: actions-rs/toolchain@v1
-        with:
-          profile: minimal
-          toolchain: ${{ matrix.rust }}
-          override: true
-      - uses: actions-rs/cargo@v1
-        with:
-          command: check
-          args: --workspace
-
-  test:
-    name: Test Suite
-    runs-on: ${{ matrix.os }}
-    strategy:
-      matrix:
-        os: [ubuntu-latest, windows-latest, macOS-latest]
-        rust: [stable]
-    steps:
-      - name: Delete huge unnecessary tools folder
-        if: runner.os == 'Linux'
-        run: rm -rf /opt/hostedtoolcache
-      - uses: actions/checkout@v4
-      - uses: actions/setup-python@v5
-        with:
-          python-version: "3.11"
-      - uses: actions-rs/toolchain@v1
-        with:
-          profile: minimal
-          toolchain: ${{ matrix.rust }}
-          override: true
-      - uses: actions-rs/cargo@v1
-        with:
-          command: test
-          args: --workspace
-
-  fmt:
-    name: Rustfmt
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/checkout@v4
-      - uses: actions-rs/toolchain@v1
-        with:
-          profile: minimal
-          toolchain: stable
-          override: true
-      - run: rustup component add rustfmt
-      - uses: actions-rs/cargo@v1
-        with:
-          command: fmt
-          args: --all -- --check
-
-  clippy:
-    name: Clippy
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/checkout@v4
-      - uses: actions-rs/toolchain@v1
-        with:
-          profile: minimal
-          toolchain: stable
-          override: true
-      - run: rustup component add clippy
-      - uses: actions-rs/cargo@v1
-        with:
-          command: clippy
-          args: --workspace --tests --examples -- -D warnings

crates/vendor/candle/.github/workflows/trufflehog.yml

@@ -1,15 +0,0 @@
-on:
-  push:
-
-name: Secret Leaks
-
-jobs:
-  trufflehog:
-    runs-on: ubuntu-latest
-    steps:
-    - name: Checkout code
-      uses: actions/checkout@v4
-      with:
-        fetch-depth: 0
-    - name: Secret Scanning
-      uses: trufflesecurity/trufflehog@main

crates/vendor/candle/.gitignore

@@ -1,48 +0,0 @@
-# Generated by Cargo
-# will have compiled files and executables
-debug/
-data/
-dist/
-target/
-
-# Remove Cargo.lock from gitignore if creating an executable, leave it for libraries
-# More information here https://doc.rust-lang.org/cargo/guide/cargo-toml-vs-cargo-lock.html
-Cargo.lock
-
-# editor config
-.helix
-.vscode
-
-# These are backup files generated by rustfmt
-**/*.rs.bk
-
-# MSVC Windows builds of rustc generate these, which store debugging information
-*.pdb
-
-*tokenizer*.json
-*.npz
-
-perf.data
-flamegraph.svg
-*.dylib
-*.so
-*.swp
-*.swo
-trace-*.json
-
-candle-wasm-examples/*/build
-candle-wasm-examples/*/*.bin
-candle-wasm-examples/*/*.jpeg
-candle-wasm-examples/*/audios/*.wav
-candle-wasm-examples/**/*.safetensors
-candle-wasm-examples/**/*.gguf
-candle-wasm-examples/*/package-lock.json
-candle-wasm-examples/**/config*.json
-.DS_Store
-.idea/*
-__pycache__
-out.safetensors
-out.wav
-bria.mp3
-bria.safetensors
-bria.wav

crates/vendor/candle/.gitmodules

@@ -1,3 +0,0 @@
-[submodule "candle-examples/examples/flash-attn/cutlass"]
-	path = candle-flash-attn/cutlass
-	url = https://github.com/NVIDIA/cutlass.git

crates/vendor/candle/.pre-commit-config.yaml

@@ -1,15 +0,0 @@
-repos:
-  - repo: https://github.com/Narsil/pre-commit-rust
-    rev: 2eed6366172ef2a5186e8785ec0e67243d7d73d0
-    hooks:
-      - id: fmt
-        name: "Rust (fmt)"
-      - id: clippy
-        name: "Rust (clippy)"
-        args:
-          [
-            "--tests",
-            "--examples",
-            "--",
-            "-Dwarnings",
-          ]

crates/vendor/candle/CHANGELOG.md

@@ -1,113 +0,0 @@
-# Changelog
-This documents the main changes to the `candle` crate.
-
-## v0.3.1 - Unreleased
-
-### Added
-
-### Modified
-
-## v0.3.0 - 2023-10-01
-
-### Added
-
-- Added the Mistral 7b v0.1 model
-  [983](https://github.com/huggingface/candle/pull/983).
-- Quantized version of the Mistral model
-  [1009](https://github.com/huggingface/candle/pull/1009).
-- Add the gelu-erf op and activation function
-  [969](https://github.com/huggingface/candle/pull/969).
-- Add the mixformer/phi-v1.5 model
-  [930](https://github.com/huggingface/candle/pull/930).
-- Add the sclice-scatter op
-  [927](https://github.com/huggingface/candle/pull/927).
-- Add the Wuerstchen diffusion model
-  [911](https://github.com/huggingface/candle/pull/911).
-
-### Modified
-
-- Support for simd128 intrinsics in some quantized vecdots
-  [982](https://github.com/huggingface/candle/pull/982).
-- Optimize the index-select cuda kernel
-  [976](https://github.com/huggingface/candle/pull/976).
-- Self-contained safetensor wrappers
-  [946](https://github.com/huggingface/candle/pull/946).
-
-## v0.2.2 - 2023-09-18
-
-### Added
-- Support for `top_p` sampling
-  [819](https://github.com/huggingface/candle/pull/819).
-- T5 model including decoding
-  [864](https://github.com/huggingface/candle/pull/864).
-- 1-d upsampling
-  [839](https://github.com/huggingface/candle/pull/839).
-
-### Modified
-- Bugfix for conv2d
-  [820](https://github.com/huggingface/candle/pull/820).
-- Support tensor based indexing using `.i`
-  [842](https://github.com/huggingface/candle/pull/842).
-
-## v0.2.1 - 2023-09-11
-
-### Added
-- Add some RNNs (GRU and LSTM) in `candle-nn`
-  [674](https://github.com/huggingface/candle/pull/674),
-  [688](https://github.com/huggingface/candle/pull/688).
-- gguf v2 support
-  [725](https://github.com/huggingface/candle/pull/725).
-- Quantized llama example in Python using the pyo3 api
-  [716](https://github.com/huggingface/candle/pull/716).
-- `candle-nn` layer for conv2d-transposed
-  [760](https://github.com/huggingface/candle/pull/760).
-- Add the Segment-Anything Model (SAM) as an example
-  [773](https://github.com/huggingface/candle/pull/773).
-- TinyViT backbone for the segment anything example
-  [787](https://github.com/huggingface/candle/pull/787).
-- Shape with holes support
-  [770](https://github.com/huggingface/candle/pull/770).
-
-### Modified
-- Dilations are now supported in conv-transpose2d.
-  [671](https://github.com/huggingface/candle/pull/671).
-- Interactive mode for the quantized model
-  [690](https://github.com/huggingface/candle/pull/690).
-- Faster softmax operation
-  [747](https://github.com/huggingface/candle/pull/747).
-- Faster convolution operations on CPU and CUDA via im2col
-  [802](https://github.com/huggingface/candle/pull/802).
-- Moving some models to a more central location
-  [796](https://github.com/huggingface/candle/pull/796).
-
-## v0.2.0 - 2023-08-30
-
-### Added
-- Add the powf op
-  [664](https://github.com/huggingface/candle/pull/664).
-- Stable Diffusion XL support
-  [647](https://github.com/huggingface/candle/pull/647).
-- Add the conv-transpose2d op
-  [635](https://github.com/huggingface/candle/pull/635).
-- Refactor the VarBuilder api
-  [627](https://github.com/huggingface/candle/pull/627).
-- Add some quantization command
-  [625](https://github.com/huggingface/candle/pull/625).
-- Support more quantized types, e.g. Q2K, Q4K, Q5K...
-  [586](https://github.com/huggingface/candle/pull/586).
-- Add pose estimation to the yolo example
-  [589](https://github.com/huggingface/candle/pull/589).
-- Api to write GGUF files
-  [585](https://github.com/huggingface/candle/pull/585).
-- Support more quantization types
-  [580](https://github.com/huggingface/candle/pull/580).
-- Add EfficientNet as an example Computer Vision model
-  [572](https://github.com/huggingface/candle/pull/572).
-- Add a group parameter to convolutions
-  [566](https://github.com/huggingface/candle/pull/566).
-- New dtype: int64
-  [563](https://github.com/huggingface/candle/pull/563).
-- Handling of the GGUF file format.
-  [559](https://github.com/huggingface/candle/pull/559).
-
-## v0.1.2 - 2023-08-21

crates/vendor/candle/Cargo.toml

@@ -1,82 +0,0 @@
-[workspace]
-members = [
-    "candle-core",
-    "candle-datasets",
-    "candle-examples",
-    "candle-book",
-    "candle-nn",
-    "candle-pyo3",
-    "candle-transformers",
-    "candle-wasm-examples/*",
-    "candle-wasm-tests",
-    "tensor-tools",
-]
-exclude = [
-   "candle-flash-attn",
-   "candle-kernels",
-   "candle-metal-kernels",
-   "candle-onnx",
-]
-resolver = "2"
-
-[workspace.package]
-version = "0.8.3"
-edition = "2021"
-description = "Minimalist ML framework."
-repository = "https://github.com/huggingface/candle"
-keywords = ["blas", "tensor", "machine-learning"]
-categories = ["science"]
-license = "MIT OR Apache-2.0"
-
-[workspace.dependencies]
-ab_glyph = "0.2.23"
-accelerate-src = { version = "0.3.2" }
-anyhow = { version = "1", features = ["backtrace"] }
-byteorder = "1.4.3"
-candle = { path = "./candle-core", package = "candle-core", version = "0.8.3" }
-candle-datasets = { path = "./candle-datasets", version = "0.8.3" }
-candle-flash-attn = { path = "./candle-flash-attn", version = "0.8.3" }
-candle-kernels = { path = "./candle-kernels", version = "0.8.3" }
-candle-metal-kernels = { path = "./candle-metal-kernels", version = "0.8.3" }
-candle-nn = { path = "./candle-nn", version = "0.8.3" }
-candle-onnx = { path = "./candle-onnx", version = "0.8.3" }
-candle-transformers = { path = "./candle-transformers", version = "0.8.3" }
-clap = { version = "4.2.4", features = ["derive"] }
-criterion = { version = "0.5.1", default-features=false }
-cudarc = { version = "0.13.5", features = ["std", "cublas", "cublaslt", "curand", "driver", "nvrtc", "f16", "cuda-version-from-build-system", "dynamic-linking"], default-features=false }
-fancy-regex = "0.13.0"
-gemm = { version = "0.17.0", features = ["wasm-simd128-enable"] }
-hf-hub = "0.4.1"
-half = { version = "2.3.1", features = ["num-traits", "use-intrinsics", "rand_distr"] }
-hound = "3.5.1"
-image = { version = "0.25.2", default-features = false, features = ["jpeg", "png"] }
-imageproc = { version = "0.24.0", default-features = false }
-intel-mkl-src = { version = "0.8.1", features = ["mkl-static-lp64-iomp"] }
-libc = { version = "0.2.147" }
-log = "0.4"
-memmap2 = { version = "0.9.3", features = ["stable_deref_trait"] }
-num_cpus = "1.15.0"
-num-traits = "0.2.15"
-parquet = { version = "51.0.0" }
-rand = "0.8.5"
-rand_distr = "0.4.3"
-rayon = "1.7.0"
-safetensors = "0.4.1"
-serde = { version = "1.0.171", features = ["derive"] }
-serde_plain = "1.0.2"
-serde_json = "1.0.99"
-thiserror = "1"
-tokenizers = { version = "0.19.1", default-features = false }
-tracing = "0.1.37"
-tracing-chrome = "0.7.1"
-tracing-subscriber = "0.3.7"
-ug = "0.1.0"
-ug-cuda = "0.1.0"
-ug-metal = "0.1.0"
-yoke = { version = "0.7.2", features = ["derive"] }
-zip = { version = "1.1.1", default-features = false }
-metal = { version = "0.27.0", features = ["mps"]}
-
-[profile.release-with-debug]
-inherits = "release"
-debug = true