gte 0.0.12 → 0.0.13

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: fd947255138b34b53e9ddf15a5c0e69295df78d95b87e87fa0eb618ae18f1ece
-  data.tar.gz: 87ace8f9622fa68ef2f4d614fba5a18da11891c825f9b3b751896a366c7bf130
+  metadata.gz: 278028df09fbcdd14fd583f0af5e1a8c9553adb28fe7aa0bc67b67666dbbdccd
+  data.tar.gz: ce994e3f505200ed4654ca8f87f585ff88919201fe82dd79007622f07a3d1ea0
 SHA512:
-  metadata.gz: 1b1dbf36073caeced7595047f53cc4c8dc4ae425b1bf06024f10d6835a61cc7270bf3a1768a9e3ba9e5d4e39115d3266db28767f353aaf968633a8c0a450f7f4
-  data.tar.gz: a18206611d3f7c0ddc175514502c471dabb22e0ff225081f73b3f4c267d47f0b85364d99e781f9df7f5184ada30bc31c17351ae1824f05f906bd041f2a0c9e45
+  metadata.gz: 742f1830ff2b83f89726be527c4323a81649b04f341b7adc0544a9000373f6a097c0b4b4ba211ead5912ba45d876565fbaab6d723ef8f06c488ab7827323f827
+  data.tar.gz: 75e91b3d4c3980b166268c6468b96bebe4b74db999e0cee433a295e57d89bec95c7614b004c61e8b3ed88cff30f02f3b6aff74de710d3dd3bb34552f36fb3422

data/README.md

@@ -59,6 +59,43 @@ Notes:
 - Return a `Config::Text` from the block (for example, `config.with(...)`).
 - Model instances are cached by full config key; different config values create different cached instances.
 
+Common model presets:
+
+```ruby
+e5 = GTE.config(ENV.fetch("GTE_MODEL_DIR")) do |config|
+  config.with(
+    model_name: "model.onnx",
+    output_tensor: "last_hidden_state",
+    max_length: 512,
+    execution_providers: "cpu"
+  )
+end
+
+siglip2 = GTE.config(ENV.fetch("GTE_SIGLIP2_DIR")) do |config|
+  config.with(
+    model_name: "text_model_int8.onnx",
+    output_tensor: "pooler_output",
+    max_length: 64,
+    execution_providers: "cpu"
+  )
+end
+
+clip = GTE.config(ENV.fetch("GTE_CLIP_DIR")) do |config|
+  config.with(
+    output_tensor: "sentence_embedding",
+    max_length: 512,
+    execution_providers: "cpu"
+  )
+end
+```
+
+Picking a specific layer:
+
+- Use `output_tensor:` to request a named model output.
+- `last_hidden_state` gives token-level hidden states and is mean-pooled by `gte` when the tensor is rank 3.
+- `pooler_output`, `sentence_embedding`, and similar 2D tensors are returned directly and then L2-normalized by default.
+- If the requested tensor is not present in the model, `gte` raises an error instead of silently falling back.
+
 Low-level embedder setup (without model cache):
 
 ```ruby

data/VERSION

@@ -1 +1 @@
-0.0.12
+0.0.13

data/ext/gte/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "gte"
-version = "0.0.12"
+version = "0.0.13"
 edition = "2021"
 authors = ["elcuervo <elcuervo@elcuervo.net>"]
 license = "MIT"

data/ext/gte/src/session.rs

@@ -7,7 +7,7 @@ use ndarray::{Array2, ArrayView2, ArrayViewD, Ix2};
 use ort::execution_providers::{
     CoreMLExecutionProvider, ExecutionProviderDispatch, XNNPACKExecutionProvider,
 };
-use ort::session::Session;
+use ort::session::{OutputSelector, RunOptions, Session};
 use std::path::{Path, PathBuf};
 use std::sync::atomic::{AtomicUsize, Ordering};
 use std::sync::{Condvar, Mutex};
@@ -216,8 +216,11 @@ pub fn run_session(
     config: &ModelConfig,
 ) -> Result<Array2<f32>> {
     let input_tensors = InputTensors::from_tokenized(tokenized, config.with_attention_mask)?;
+    let run_opts = RunOptions::new()
+        .map_err(|e| GteError::Ort(e.to_string()))?
+        .with_outputs(OutputSelector::no_default().with(config.output_tensor.as_str()));
     let outputs = session
-        .run(input_tensors.inputs)
+        .run_with_options(input_tensors.inputs, &run_opts)
         .map_err(|e| GteError::Ort(e.to_string()))?;
     let array = extract_output_tensor(&outputs, config.output_tensor.as_str())?;
 

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: gte
 version: !ruby/object:Gem::Version
-  version: 0.0.12
+  version: 0.0.13
 platform: ruby
 authors:
 - elcuervo