|
| 1 | +use anyhow::{bail, Error}; |
| 2 | +use deno_core::error::AnyError; |
| 3 | +use deno_core::op2; |
| 4 | +use deno_core::OpState; |
| 5 | +use ndarray::{Array1, Array2, Axis, Ix2}; |
| 6 | +use ndarray_linalg::norm::{normalize, NormalizeAxis}; |
| 7 | +use ort::{inputs, GraphOptimizationLevel, Session, Tensor}; |
| 8 | +use std::path::Path; |
| 9 | +use tokenizers::normalizers::bert::BertNormalizer; |
| 10 | +use tokenizers::Tokenizer; |
| 11 | + |
| 12 | +deno_core::extension!( |
| 13 | + sb_ai, |
| 14 | + ops = [op_sb_ai_run_model], |
| 15 | + esm_entry_point = "ext:sb_ai/ai.js", |
| 16 | + esm = ["ai.js",] |
| 17 | +); |
| 18 | + |
| 19 | +fn run_gte(state: &mut OpState, prompt: String) -> Result<Vec<f32>, Error> { |
| 20 | + // Create the ONNX Runtime environment, for all sessions created in this process. |
| 21 | + ort::init().with_name("GTE").commit()?; |
| 22 | + |
| 23 | + let models_dir = std::env::var("SB_AI_MODELS_DIR").unwrap_or("/etc/sb_ai/models".to_string()); |
| 24 | + |
| 25 | + let mut session = state.try_take::<Session>(); |
| 26 | + if session.is_none() { |
| 27 | + session = Some( |
| 28 | + Session::builder()? |
| 29 | + .with_optimization_level(GraphOptimizationLevel::Disable)? |
| 30 | + .with_intra_threads(1)? |
| 31 | + .with_model_from_file( |
| 32 | + Path::new(&models_dir) |
| 33 | + .join("gte") |
| 34 | + .join("gte_small_quantized.onnx"), |
| 35 | + )?, |
| 36 | + ); |
| 37 | + } |
| 38 | + let session = session.unwrap(); |
| 39 | + |
| 40 | + // Load the tokenizer and encode the prompt into a sequence of tokens. |
| 41 | + let mut tokenizer = state.try_take::<Tokenizer>(); |
| 42 | + if tokenizer.is_none() { |
| 43 | + tokenizer = Some( |
| 44 | + Tokenizer::from_file( |
| 45 | + Path::new(&models_dir) |
| 46 | + .join("gte") |
| 47 | + .join("gte_small_tokenizer.json"), |
| 48 | + ) |
| 49 | + .map_err(anyhow::Error::msg)?, |
| 50 | + ) |
| 51 | + } |
| 52 | + let mut tokenizer = tokenizer.unwrap(); |
| 53 | + |
| 54 | + let tokenizer_impl = tokenizer |
| 55 | + .with_normalizer(BertNormalizer::default()) |
| 56 | + .with_padding(None) |
| 57 | + .with_truncation(None) |
| 58 | + .map_err(anyhow::Error::msg)?; |
| 59 | + |
| 60 | + let tokens = tokenizer_impl |
| 61 | + .encode(prompt, true) |
| 62 | + .map_err(anyhow::Error::msg)? |
| 63 | + .get_ids() |
| 64 | + .iter() |
| 65 | + .map(|i| *i as i64) |
| 66 | + .collect::<Vec<_>>(); |
| 67 | + |
| 68 | + let tokens = Array1::from_iter(tokens.iter().cloned()); |
| 69 | + |
| 70 | + let array = tokens.view().insert_axis(Axis(0)); |
| 71 | + let dims = array.raw_dim(); |
| 72 | + let token_type_ids = Array2::<i64>::zeros(dims); |
| 73 | + let attention_mask = Array2::<i64>::ones(dims); |
| 74 | + let outputs = session.run(inputs! { |
| 75 | + "input_ids" => array, |
| 76 | + "token_type_ids" => token_type_ids, |
| 77 | + "attention_mask" => attention_mask, |
| 78 | + }?)?; |
| 79 | + |
| 80 | + let embeddings: Tensor<f32> = outputs["last_hidden_state"].extract_tensor()?; |
| 81 | + |
| 82 | + let embeddings_view = embeddings.view(); |
| 83 | + let mean_pool = embeddings_view.mean_axis(Axis(1)).unwrap(); |
| 84 | + let (normalized, _) = normalize( |
| 85 | + mean_pool.into_dimensionality::<Ix2>().unwrap(), |
| 86 | + NormalizeAxis::Row, |
| 87 | + ); |
| 88 | + |
| 89 | + let slice = normalized.view().to_slice().unwrap().to_vec(); |
| 90 | + |
| 91 | + drop(outputs); |
| 92 | + |
| 93 | + state.put::<Session>(session); |
| 94 | + state.put::<Tokenizer>(tokenizer); |
| 95 | + |
| 96 | + Ok(slice) |
| 97 | +} |
| 98 | + |
| 99 | +#[op2] |
| 100 | +#[serde] |
| 101 | +pub fn op_sb_ai_run_model( |
| 102 | + state: &mut OpState, |
| 103 | + #[string] name: String, |
| 104 | + #[string] prompt: String, |
| 105 | +) -> Result<Vec<f32>, AnyError> { |
| 106 | + if name == "gte" { |
| 107 | + run_gte(state, prompt) |
| 108 | + } else { |
| 109 | + bail!("model not supported") |
| 110 | + } |
| 111 | +} |
0 commit comments