chore: Bump ort to v2.0.0-rc.9

Signed-off-by: Anush008 <anushshetty90@gmail.com>

Cargo.toml

@@ -13,7 +13,7 @@ authors = [
   "Luya Wang <luya.wang@qq.com>",
   "Tri <tri@triandco.com>",
   "Denny Wong <denwong47@hotmail.com>",
-  "Alex Rozgo <alex.rozgo@gmail.com>"
+  "Alex Rozgo <alex.rozgo@gmail.com>",
 ]
 documentation = "https://docs.rs/fastembed"
 repository = "https://github.com/Anush008/fastembed-rs"
@@ -26,8 +26,9 @@ anyhow = { version = "1" }
 hf-hub = { version = "0.3", default-features = false }
 image = "0.25.2"
 ndarray = { version = "0.16", default-features = false }
-ort = { version = "=2.0.0-rc.8", default-features = false, features = [
-    "half", "ndarray",
+ort = { version = "=2.0.0-rc.9", default-features = false, features = [
+  "half",
+  "ndarray",
 ] }
 rayon = { version = "1.10", default-features = false }
 serde_json = { version = "1" }

src/image_embedding/impl.rs

@@ -4,7 +4,10 @@ use hf_hub::{
     Cache,
 };
 use ndarray::{Array3, ArrayView3};
-use ort::{GraphOptimizationLevel, Session, Value};
+use ort::{
+    session::{builder::GraphOptimizationLevel, Session},
+    value::Value,
+};
 #[cfg(feature = "online")]
 use std::path::PathBuf;
 use std::{path::Path, thread::available_parallelism};
@@ -32,6 +35,8 @@ impl ImageEmbedding {
     /// Uses the total number of CPUs available as the number of intra-threads
     #[cfg(feature = "online")]
     pub fn try_new(options: ImageInitOptions) -> anyhow::Result<Self> {
+        use ort::session::{builder::GraphOptimizationLevel, Session};
+
         let ImageInitOptions {
             model_name,
             execution_providers,

src/image_embedding/init.rs

@@ -1,6 +1,6 @@
 use std::path::{Path, PathBuf};
 
-use ort::{ExecutionProviderDispatch, Session};
+use ort::{execution_providers::ExecutionProviderDispatch, session::Session};
 
 use crate::{ImageEmbeddingModel, DEFAULT_CACHE_DIR};
 

src/lib.rs

@@ -62,7 +62,7 @@ mod reranking;
 mod sparse_text_embedding;
 mod text_embedding;
 
-pub use ort::ExecutionProviderDispatch;
+pub use ort::execution_providers::ExecutionProviderDispatch;
 
 pub use crate::common::{
     read_file_to_bytes, Embedding, Error, SparseEmbedding, TokenizerFiles, DEFAULT_CACHE_DIR,

src/output/embedding_output.rs

@@ -1,4 +1,5 @@
 use ndarray::{Array2, ArrayView, Dim, IxDynImpl};
+use ort::session::SessionOutputs;
 
 use crate::pooling;
 
@@ -10,7 +11,7 @@ use super::{OutputKey, OutputPrecedence};
 /// pooling etc. This struct should contain all the necessary information for the
 /// post-processing to be performed.
 pub struct SingleBatchOutput<'r, 's> {
-    pub session_outputs: ort::SessionOutputs<'r, 's>,
+    pub session_outputs: SessionOutputs<'r, 's>,
     pub attention_mask_array: Array2<i64>,
 }
 
@@ -23,17 +24,12 @@ impl<'r, 's> SingleBatchOutput<'r, 's> {
         &self,
         precedence: &impl OutputPrecedence,
     ) -> anyhow::Result<ArrayView<f32, Dim<IxDynImpl>>> {
-        let ort_output = precedence
+        let ort_output: &ort::value::Value = precedence
             .key_precedence()
             .find_map(|key| match key {
-                OutputKey::OnlyOne => {
-                    // Only export the value if there is only one output available.
-                    if self.session_outputs.len() == 1 {
-                        self.session_outputs.values().next()
-                    } else {
-                        None
-                    }
-                }
+                OutputKey::OnlyOne => self
+                    .session_outputs
+                    .get(self.session_outputs.keys().nth(0)?),
                 OutputKey::ByOrder(idx) => {
                     let x = self
                         .session_outputs

src/reranking/impl.rs

@@ -1,4 +1,8 @@
 use anyhow::Result;
+use ort::{
+    session::{builder::GraphOptimizationLevel, Session},
+    value::Value,
+};
 use std::thread::available_parallelism;
 
 #[cfg(feature = "online")]
@@ -10,7 +14,6 @@ use crate::{
 #[cfg(feature = "online")]
 use hf_hub::{api::sync::ApiBuilder, Cache};
 use ndarray::{s, Array};
-use ort::{GraphOptimizationLevel, Session, Value};
 use rayon::{iter::ParallelIterator, slice::ParallelSlice};
 use tokenizers::Tokenizer;
 
@@ -47,6 +50,8 @@ impl TextRerank {
 
     #[cfg(feature = "online")]
     pub fn try_new(options: RerankInitOptions) -> Result<TextRerank> {
+        use ort::session::{builder::GraphOptimizationLevel, Session};
+
         use super::RerankInitOptions;
 
         let RerankInitOptions {

src/reranking/init.rs

@@ -1,6 +1,6 @@
 use std::path::{Path, PathBuf};
 
-use ort::{ExecutionProviderDispatch, Session};
+use ort::{execution_providers::ExecutionProviderDispatch, session::Session};
 use tokenizers::Tokenizer;
 
 use crate::{RerankerModel, TokenizerFiles, DEFAULT_CACHE_DIR};

src/sparse_text_embedding/impl.rs

@@ -11,9 +11,8 @@ use hf_hub::{
     Cache,
 };
 use ndarray::{Array, CowArray};
+use ort::{session::Session, value::Value};
 #[cfg_attr(not(feature = "online"), allow(unused_imports))]
-use ort::GraphOptimizationLevel;
-use ort::{Session, Value};
 use rayon::{iter::ParallelIterator, slice::ParallelSlice};
 #[cfg(feature = "online")]
 use std::path::PathBuf;
@@ -34,6 +33,8 @@ impl SparseTextEmbedding {
     /// Uses the total number of CPUs available as the number of intra-threads
     #[cfg(feature = "online")]
     pub fn try_new(options: SparseInitOptions) -> Result<Self> {
+        use ort::session::{builder::GraphOptimizationLevel, Session};
+
         use super::SparseInitOptions;
 
         let SparseInitOptions {

src/sparse_text_embedding/init.rs

@@ -1,6 +1,6 @@
 use std::path::{Path, PathBuf};
 
-use ort::{ExecutionProviderDispatch, Session};
+use ort::{execution_providers::ExecutionProviderDispatch, session::Session};
 use tokenizers::Tokenizer;
 
 use crate::{models::sparse::SparseModel, TokenizerFiles, DEFAULT_CACHE_DIR};

src/text_embedding/impl.rs

@@ -14,7 +14,10 @@ use hf_hub::{
     Cache,
 };
 use ndarray::Array;
-use ort::{GraphOptimizationLevel, Session, Value};
+use ort::{
+    session::{builder::GraphOptimizationLevel, Session},
+    value::Value,
+};
 use rayon::{
     iter::{FromParallelIterator, ParallelIterator},
     slice::ParallelSlice,
@@ -38,6 +41,8 @@ impl TextEmbedding {
     /// Uses the total number of CPUs available as the number of intra-threads
     #[cfg(feature = "online")]
     pub fn try_new(options: InitOptions) -> anyhow::Result<Self> {
+        use ort::session::{builder::GraphOptimizationLevel, Session};
+
         let InitOptions {
             model_name,
             execution_providers,

src/text_embedding/init.rs

@@ -6,7 +6,7 @@ use crate::{
     pooling::Pooling,
     EmbeddingModel, QuantizationMode,
 };
-use ort::{ExecutionProviderDispatch, Session};
+use ort::{execution_providers::ExecutionProviderDispatch, session::Session};
 use std::path::{Path, PathBuf};
 use tokenizers::Tokenizer;