feat: GTE classification head (huggingface#441)

Co-authored-by: Hyeongchan Kim <[email protected]>
kaixuanliu · Nov 27, 2024 · 0bfeb7e · 0bfeb7e
1 parent 7c4f67e
commit 0bfeb7e
Show file tree

Hide file tree

Showing 18 changed files with 145 additions and 32 deletions.
diff --git a/Cargo.lock b/Cargo.lock
diff --git a/README.md b/README.md
@@ -92,11 +92,12 @@ Text Embeddings Inference currently supports CamemBERT, and XLM-RoBERTa Sequence
 
 Below are some examples of the currently supported models:
 
-| Task               | Model Type  | Model ID                                                                                    |
-|--------------------|-------------|---------------------------------------------------------------------------------------------|
-| Re-Ranking         | XLM-RoBERTa | [BAAI/bge-reranker-large](https://huggingface.co/BAAI/bge-reranker-large)                   |
-| Re-Ranking         | XLM-RoBERTa | [BAAI/bge-reranker-base](https://huggingface.co/BAAI/bge-reranker-base)                     |
-| Sentiment Analysis | RoBERTa     | [SamLowe/roberta-base-go_emotions](https://huggingface.co/SamLowe/roberta-base-go_emotions) |
+| Task               | Model Type  | Model ID                                                                                                        |
+|--------------------|-------------|-----------------------------------------------------------------------------------------------------------------|
+| Re-Ranking         | XLM-RoBERTa | [BAAI/bge-reranker-large](https://huggingface.co/BAAI/bge-reranker-large)                                       |
+| Re-Ranking         | XLM-RoBERTa | [BAAI/bge-reranker-base](https://huggingface.co/BAAI/bge-reranker-base)                                         |
+| Re-Ranking         | GTE         | [Alibaba-NLP/gte-multilingual-reranker-base](https://huggingface.co/Alibaba-NLP/gte-multilingual-reranker-base) |
+| Sentiment Analysis | RoBERTa     | [SamLowe/roberta-base-go_emotions](https://huggingface.co/SamLowe/roberta-base-go_emotions)                     |
 
 ### Docker
 
@@ -372,7 +373,7 @@ docker run --gpus all -p 8080:80 -v $volume:/data --pull always ghcr.io/huggingf
 
 ### Using Re-rankers models
 
-`text-embeddings-inference` v0.4.0 added support for CamemBERT, RoBERTa and XLM-RoBERTa Sequence Classification models.
+`text-embeddings-inference` v0.4.0 added support for CamemBERT, RoBERTa, XLM-RoBERTa, and GTE Sequence Classification models.
 Re-rankers models are Sequence Classification cross-encoders models with a single class that scores the similarity
 between a query and a text.
 
@@ -392,7 +393,7 @@ And then you can rank the similarity between a query and a list of texts with:
 ```bash
 curl 127.0.0.1:8080/rerank \
     -X POST \
-    -d '{"query":"What is Deep Learning?", "texts": ["Deep Learning is not...", "Deep learning is..."]}' \
+    -d '{"query": "What is Deep Learning?", "texts": ["Deep Learning is not...", "Deep learning is..."]}' \
     -H 'Content-Type: application/json'
 ```
 

diff --git a/backends/candle/src/models/flash_bert.rs b/backends/candle/src/models/flash_bert.rs
@@ -529,6 +529,7 @@ impl Model for FlashBertModel {
     fn is_padded(&self) -> bool {
         false
     }
+
     fn embed(&self, batch: Batch) -> Result<(Option<Tensor>, Option<Tensor>)> {
         self.forward(batch)
     }

diff --git a/backends/candle/src/models/flash_gte.rs b/backends/candle/src/models/flash_gte.rs
@@ -198,13 +198,66 @@ impl GTELayer {
     }
 }
 
+pub struct GTEClassificationHead {
+    pooler: Option<Linear>,
+    classifier: Linear,
+    span: tracing::Span,
+}
+
+impl GTEClassificationHead {
+    #[allow(dead_code)]
+    pub(crate) fn load(vb: VarBuilder, config: &GTEConfig) -> Result<Self> {
+        let n_classes = match &config.id2label {
+            None => candle::bail!("`id2label` must be set for classifier models"),
+            Some(id2label) => id2label.len(),
+        };
+
+        let pooler = if let Ok(pooler_weight) = vb
+            .pp("pooler.dense")
+            .get((config.hidden_size, config.hidden_size), "weight")
+        {
+            let pooler_bias = vb.pp("pooler.dense").get(config.hidden_size, "bias")?;
+            Some(Linear::new(pooler_weight, Some(pooler_bias), None))
+        } else {
+            None
+        };
+
+        let classifier_weight = vb
+            .pp("classifier")
+            .get((n_classes, config.hidden_size), "weight")?;
+        let classifier_bias = vb.pp("classifier").get(n_classes, "bias")?;
+        let classifier = Linear::new(classifier_weight, Some(classifier_bias), None);
+
+        Ok(Self {
+            classifier,
+            pooler,
+            span: tracing::span!(tracing::Level::TRACE, "classifier"),
+        })
+    }
+
+    pub(crate) fn forward(&self, hidden_states: &Tensor) -> Result<Tensor> {
+        let _enter = self.span.enter();
+
+        let mut hidden_states = hidden_states.unsqueeze(1)?;
+        if let Some(pooler) = self.pooler.as_ref() {
+            hidden_states = pooler.forward(&hidden_states)?;
+            hidden_states = hidden_states.tanh()?;
+        }
+
+        let hidden_states = self.classifier.forward(&hidden_states)?;
+        let hidden_states = hidden_states.squeeze(1)?;
+        Ok(hidden_states)
+    }
+}
+
 pub struct FlashGTEModel {
     word_embeddings: Embedding,
     token_type_embeddings: Option<Embedding>,
     layers: Vec<GTELayer>,
     embeddings_norm: LayerNorm,
     cos_cache: Tensor,
     sin_cache: Tensor,
+    classifier: Option<GTEClassificationHead>,
     pool: Pool,
     pub device: Device,
 
@@ -233,11 +286,14 @@ impl FlashGTEModel {
             candle::bail!("Only `PositionEmbeddingType::Rope` is supported");
         }
 
-        let pool = match model_type {
+        let (pool, classifier) = match model_type {
             ModelType::Classifier => {
-                candle::bail!("`classifier` model type is not supported for GTE")
+                let pool = Pool::Cls;
+
+                let classifier = GTEClassificationHead::load(vb.clone(), config)?;
+                (pool, Some(classifier))
             }
-            ModelType::Embedding(pool) => pool,
+            ModelType::Embedding(pool) => (pool, None),
         };
 
         let word_embeddings = Embedding::new(
@@ -292,6 +348,7 @@ impl FlashGTEModel {
             embeddings_norm,
             cos_cache,
             sin_cache,
+            classifier,
             pool,
             device: vb.device().clone(),
             span: tracing::span!(tracing::Level::TRACE, "model"),
@@ -457,7 +514,20 @@ impl Model for FlashGTEModel {
     fn is_padded(&self) -> bool {
         false
     }
+
     fn embed(&self, batch: Batch) -> Result<(Option<Tensor>, Option<Tensor>)> {
         self.forward(batch)
     }
+
+    fn predict(&self, batch: Batch) -> Result<Tensor> {
+        match &self.classifier {
+            None => candle::bail!("`predict` is not implemented for this model"),
+            Some(classifier) => {
+                let (pooled_embeddings, _raw_embeddings) = self.forward(batch)?;
+                let pooled_embeddings =
+                    pooled_embeddings.expect("pooled_embeddings is empty. This is a bug.");
+                classifier.forward(&pooled_embeddings)
+            }
+        }
+    }
 }
diff --git a/backends/candle/src/models/gte.rs b/backends/candle/src/models/gte.rs
@@ -1,6 +1,7 @@
 use crate::layers::HiddenAct;
 use crate::models::PositionEmbeddingType;
 use serde::Deserialize;
+use std::collections::HashMap;
 
 #[derive(Debug, Clone, PartialEq, Deserialize)]
 pub struct NTKScaling {
@@ -32,4 +33,5 @@ pub struct GTEConfig {
     pub logn_attention_scale: bool,
     #[serde(default)]
     pub logn_attention_clip1: bool,
+    pub id2label: Option<HashMap<String, String>>,
 }
diff --git a/backends/candle/tests/snapshots/test_flash_gte__gte_classification_single.snap b/backends/candle/tests/snapshots/test_flash_gte__gte_classification_single.snap
@@ -0,0 +1,6 @@
+---
+source: backends/candle/tests/test_flash_gte.rs
+assertion_line: 86
+expression: predictions_single
+---
+- - -0.74365234
diff --git a/backends/candle/tests/test_bert.rs b/backends/candle/tests/test_bert.rs
@@ -13,7 +13,7 @@ fn test_mini() -> Result<()> {
     let tokenizer = load_tokenizer(&model_root)?;
 
     let backend = CandleBackend::new(
-        model_root,
+        &model_root,
         "float32".to_string(),
         ModelType::Embedding(Pool::Mean),
     )?;
@@ -73,7 +73,7 @@ fn test_mini_pooled_raw() -> Result<()> {
     let tokenizer = load_tokenizer(&model_root)?;
 
     let backend = CandleBackend::new(
-        model_root,
+        &model_root,
         "float32".to_string(),
         ModelType::Embedding(Pool::Cls),
     )?;
@@ -142,7 +142,7 @@ fn test_emotions() -> Result<()> {
     let model_root = download_artifacts("SamLowe/roberta-base-go_emotions", None)?;
     let tokenizer = load_tokenizer(&model_root)?;
 
-    let backend = CandleBackend::new(model_root, "float32".to_string(), ModelType::Classifier)?;
+    let backend = CandleBackend::new(&model_root, "float32".to_string(), ModelType::Classifier)?;
 
     let input_batch = batch(
         vec![
@@ -192,7 +192,7 @@ fn test_bert_classification() -> Result<()> {
     let model_root = download_artifacts("ibm/re2g-reranker-nq", Some("refs/pr/3"))?;
     let tokenizer = load_tokenizer(&model_root)?;
 
-    let backend = CandleBackend::new(model_root, "float32".to_string(), ModelType::Classifier)?;
+    let backend = CandleBackend::new(&model_root, "float32".to_string(), ModelType::Classifier)?;
 
     let input_single = batch(
         vec![tokenizer

diff --git a/backends/candle/tests/test_flash_bert.rs b/backends/candle/tests/test_flash_bert.rs
@@ -19,7 +19,7 @@ fn test_flash_mini() -> Result<()> {
     let tokenizer = load_tokenizer(&model_root)?;
 
     let backend = CandleBackend::new(
-        model_root,
+        &model_root,
         "float16".to_string(),
         ModelType::Embedding(Pool::Mean),
     )?;
@@ -83,7 +83,7 @@ fn test_flash_mini_pooled_raw() -> Result<()> {
     let tokenizer = load_tokenizer(&model_root)?;
 
     let backend = CandleBackend::new(
-        model_root,
+        &model_root,
         "float16".to_string(),
         ModelType::Embedding(Pool::Cls),
     )?;
@@ -156,7 +156,7 @@ fn test_flash_emotions() -> Result<()> {
     let model_root = download_artifacts("SamLowe/roberta-base-go_emotions", None)?;
     let tokenizer = load_tokenizer(&model_root)?;
 
-    let backend = CandleBackend::new(model_root, "float16".to_string(), ModelType::Classifier)?;
+    let backend = CandleBackend::new(&model_root, "float16".to_string(), ModelType::Classifier)?;
 
     let input_batch = batch(
         vec![
@@ -210,7 +210,7 @@ fn test_flash_bert_classification() -> Result<()> {
     let model_root = download_artifacts("ibm/re2g-reranker-nq", Some("refs/pr/3"))?;
     let tokenizer = load_tokenizer(&model_root)?;
 
-    let backend = CandleBackend::new(model_root, "float16".to_string(), ModelType::Classifier)?;
+    let backend = CandleBackend::new(&model_root, "float16".to_string(), ModelType::Classifier)?;
 
     let input_single = batch(
         vec![tokenizer

diff --git a/backends/candle/tests/test_flash_gte.rs b/backends/candle/tests/test_flash_gte.rs
@@ -1,9 +1,9 @@
 #![allow(dead_code, unused_imports)]
 mod common;
 
-use crate::common::{sort_embeddings, SnapshotEmbeddings};
+use crate::common::{sort_embeddings, SnapshotEmbeddings, SnapshotScores};
 use anyhow::Result;
-use common::{batch, cosine_matcher, download_artifacts, load_tokenizer};
+use common::{batch, cosine_matcher, download_artifacts, load_tokenizer, relative_matcher};
 use text_embeddings_backend_candle::CandleBackend;
 use text_embeddings_backend_core::{Backend, ModelType, Pool};
 
@@ -15,7 +15,7 @@ fn test_flash_gte() -> Result<()> {
     let tokenizer = load_tokenizer(&model_root)?;
 
     let backend = CandleBackend::new(
-        model_root,
+        &model_root,
         "float16".to_string(),
         ModelType::Embedding(Pool::Cls),
     )?;
@@ -51,3 +51,36 @@ fn test_flash_gte() -> Result<()> {
 
     Ok(())
 }
+
+#[test]
+#[serial_test::serial]
+#[cfg(all(
+    feature = "cuda",
+    any(feature = "flash-attn", feature = "flash-attn-v1")
+))]
+fn test_flash_gte_classification() -> Result<()> {
+    let model_root = download_artifacts("Alibaba-NLP/gte-multilingual-reranker-base", None)?;
+    let tokenizer = load_tokenizer(&model_root)?;
+
+    let backend = CandleBackend::new(&model_root, "float16".to_string(), ModelType::Classifier)?;
+
+    let input_single = batch(
+        vec![tokenizer
+            .encode(("What is Deep Learning?", "Deep Learning is not..."), true)
+            .unwrap()],
+        [0].to_vec(),
+        vec![],
+    );
+
+    let predictions: Vec<Vec<f32>> = backend
+        .predict(input_single)?
+        .into_iter()
+        .map(|(_, v)| v)
+        .collect();
+    let predictions_single = SnapshotScores::from(predictions);
+
+    let matcher = relative_matcher();
+    insta::assert_yaml_snapshot!("gte_classification_single", predictions_single, &matcher);
+
+    Ok(())
+}
diff --git a/backends/candle/tests/test_flash_jina.rs b/backends/candle/tests/test_flash_jina.rs
@@ -15,7 +15,7 @@ fn test_flash_jina_small() -> Result<()> {
     let tokenizer = load_tokenizer(&model_root)?;
 
     let backend = CandleBackend::new(
-        model_root,
+        &model_root,
         "float16".to_string(),
         ModelType::Embedding(Pool::Mean),
     )?;

diff --git a/backends/candle/tests/test_flash_jina_code.rs b/backends/candle/tests/test_flash_jina_code.rs
@@ -15,7 +15,7 @@ fn test_flash_jina_code_base() -> Result<()> {
     let tokenizer = load_tokenizer(&model_root)?;
 
     let backend = CandleBackend::new(
-        model_root,
+        &model_root,
         "float16".to_string(),
         ModelType::Embedding(Pool::Mean),
     )?;

diff --git a/backends/candle/tests/test_flash_mistral.rs b/backends/candle/tests/test_flash_mistral.rs
@@ -15,7 +15,7 @@ fn test_flash_mistral() -> Result<()> {
     let tokenizer = load_tokenizer(&model_root)?;
 
     let backend = CandleBackend::new(
-        model_root,
+        &model_root,
         "float16".to_string(),
         ModelType::Embedding(Pool::Mean),
     )?;

diff --git a/backends/candle/tests/test_flash_nomic.rs b/backends/candle/tests/test_flash_nomic.rs
@@ -15,7 +15,7 @@ fn test_flash_nomic_small() -> Result<()> {
     let tokenizer = load_tokenizer(&model_root)?;
 
     let backend = CandleBackend::new(
-        model_root,
+        &model_root,
         "float16".to_string(),
         ModelType::Embedding(Pool::Mean),
     )?;

diff --git a/backends/candle/tests/test_flash_qwen2.rs b/backends/candle/tests/test_flash_qwen2.rs
@@ -39,7 +39,7 @@ fn test_flash_qwen2() -> Result<()> {
     };
 
     let backend = CandleBackend::new(
-        model_root,
+        &model_root,
         "float16".to_string(),
         ModelType::Embedding(Pool::LastToken),
     )?;

diff --git a/backends/candle/tests/test_jina.rs b/backends/candle/tests/test_jina.rs
@@ -12,7 +12,7 @@ fn test_jina_small() -> Result<()> {
     let tokenizer = load_tokenizer(&model_root)?;
 
     let backend = CandleBackend::new(
-        model_root,
+        &model_root,
         "float32".to_string(),
         ModelType::Embedding(Pool::Mean),
     )?;

diff --git a/backends/candle/tests/test_jina_code.rs b/backends/candle/tests/test_jina_code.rs
@@ -12,7 +12,7 @@ fn test_jina_code_base() -> Result<()> {
     let tokenizer = load_tokenizer(&model_root)?;
 
     let backend = CandleBackend::new(
-        model_root,
+        &model_root,
         "float32".to_string(),
         ModelType::Embedding(Pool::Mean),
     )?;

diff --git a/backends/candle/tests/test_nomic.rs b/backends/candle/tests/test_nomic.rs
@@ -12,7 +12,7 @@ fn test_nomic_small() -> Result<()> {
     let tokenizer = load_tokenizer(&model_root)?;
 
     let backend = CandleBackend::new(
-        model_root,
+        &model_root,
         "float32".to_string(),
         ModelType::Embedding(Pool::Mean),
     )?;
-Original file line number
+Diff line change
@@ Expand Up / @@ -529,6 +529,7 @@ impl Model for FlashBertModel { @@
         fn is_padded(&self) -> bool {
             false
         }
         fn embed(&self, batch: Batch) -> Result<(Option<Tensor>, Option<Tensor>)> {
             self.forward(batch)
         }
@@ Expand Down @@