Add a new `/tokenize` route to get the tokenized input (#1471)

# What does this PR do?


Ideally this is done client side, but this is a recurring request,
therefore we implemented it.

- Runs only if rust tokenizer is present (not encumbering the main
inference pipeline is important).
- Returns simple results, ID, text (gotten with offsets from the
original string) and offsets (so users can do things like highlighting
text).

<!--
Congratulations! You've made it this far! You're not quite done yet
though.

Once merged, your PR is going to appear in the release notes with the
title you set, so make sure it's a great title that fully reflects the
extent of your awesome contribution.

Then, please replace this with a description of the change and which
issue is fixed (if applicable). Please also include relevant motivation
and context. List any dependencies (if any) that are required for this
change.

Once you're done, someone will review your PR shortly (see the section
"Who can review?" below to tag some potential reviewers). They may
suggest changes to make the code even better. If no one reviewed your PR
after a week has passed, don't hesitate to post a new comment
@-mentioning the same persons---sometimes notifications get lost.
-->

<!-- Remove if not applicable -->

Fixes # (issue)


## Before submitting
- [ ] This PR fixes a typo or improves the docs (you can dismiss the
other checks if that's the case).
- [ ] Did you read the [contributor
guideline](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md#start-contributing-pull-requests),
      Pull Request section?
- [ ] Was this discussed/approved via a Github issue or the
[forum](https://discuss.huggingface.co/)? Please add a link
      to it if that's the case.
- [ ] Did you make sure to update the documentation with your changes?
Here are the
[documentation
guidelines](https://github.com/huggingface/transformers/tree/main/docs),
and
[here are tips on formatting
docstrings](https://github.com/huggingface/transformers/tree/main/docs#writing-source-documentation).
- [ ] Did you write any new necessary tests?


## Who can review?

Anyone in the community is free to review the PR once the tests have
passed. Feel free to tag
members/contributors who may be interested in your PR.

<!-- Your PR will be replied to more quickly if you can figure out the
right person to tag with @


@OlivierDehaene OR @Narsil

 -->

router/src/infer.rs

@@ -165,6 +165,28 @@ impl Infer {
         ))
     }
 
+    /// Tokenizer the input
+    #[instrument(skip_all)]
+    pub(crate) async fn tokenize(
+        &self,
+        request: GenerateRequest,
+    ) -> Result<Option<tokenizers::Encoding>, InferError> {
+        // Tokenize request
+        let inputs = request.inputs;
+        let truncate = request.parameters.truncate;
+        let encoding = self
+            .validation
+            .tokenize(inputs, truncate)
+            .await
+            .map_err(|err| {
+                tracing::error!("Tokenization {err}");
+                err
+            })?;
+
+        // Return Encoding
+        Ok(encoding.map(|(encoding, _)| encoding))
+    }
+
     /// Apply the chat template to the chat request
     #[instrument(skip_all)]
     pub(crate) fn apply_chat_template(&self, messages: Vec<Message>) -> Result<String, InferError> {

router/src/lib.rs

@@ -444,6 +444,18 @@ pub struct Token {
     special: bool,
 }
 
+#[derive(Debug, Serialize, ToSchema)]
+pub struct SimpleToken {
+    #[schema(example = 0)]
+    id: u32,
+    #[schema(example = "test")]
+    text: String,
+    #[schema(example = 0)]
+    start: usize,
+    #[schema(example = 2)]
+    stop: usize,
+}
+
 #[derive(Serialize, ToSchema)]
 #[serde(rename_all(serialize = "snake_case"))]
 pub(crate) enum FinishReason {

router/src/server.rs

@@ -5,8 +5,8 @@ use crate::validation::ValidationError;
 use crate::{
     BestOfSequence, ChatCompletion, ChatCompletionChunk, ChatRequest, CompatGenerateRequest,
     Details, ErrorResponse, FinishReason, GenerateParameters, GenerateRequest, GenerateResponse,
-    HubModelInfo, HubTokenizerConfig, Infer, Info, PrefillToken, StreamDetails, StreamResponse,
-    Token, Validation,
+    HubModelInfo, HubTokenizerConfig, Infer, Info, PrefillToken, SimpleToken, StreamDetails,
+    StreamResponse, Token, Validation,
 };
 use axum::extract::Extension;
 use axum::http::{HeaderMap, Method, StatusCode};
@@ -532,7 +532,7 @@ async fn generate_stream_internal(
     path = "/v1/chat/completions",
     request_body = ChatRequest,
     responses(
-    (status = 200, description = "Generated Text", body = GenerateResponse),
+    (status = 200, description = "Generated Text", body = ChatCompletionChunk),
     (status = 424, description = "Generation Error", body = ErrorResponse,
     example = json ! ({"error": "Request failed during generation"})),
     (status = 429, description = "Model is overloaded", body = ErrorResponse,
@@ -672,6 +672,52 @@ async fn chat_completions(
     }
 }
 
+/// Tokenize inputs
+#[utoipa::path(
+    post,
+    tag = "Text Generation Inference",
+    path = "/tokenize",
+    request_body = TokenizeRequest,
+    responses(
+    (status = 200, description = "Tokenized ids", body = TokenizeResponse),
+    (status = 404, description = "No tokenizer found", body = ErrorResponse,
+    example = json ! ({"error": "No fast tokenizer available"})),
+    )
+    )]
+#[instrument(skip_all)]
+async fn tokenize(
+    Extension(infer): Extension<Infer>,
+    Json(req): Json<GenerateRequest>,
+) -> Result<Response, (StatusCode, Json<ErrorResponse>)> {
+    let input = req.inputs.clone();
+    let encoding = infer.tokenize(req).await?;
+    if let Some(encoding) = encoding {
+        let tokens: Vec<SimpleToken> = encoding
+            .get_ids()
+            .iter()
+            .zip(encoding.get_offsets())
+            .map(|(&id, &(start, stop))| {
+                let text: String = input.chars().skip(start).take(stop - start).collect();
+                SimpleToken {
+                    id,
+                    text,
+                    start,
+                    stop,
+                }
+            })
+            .collect();
+        Ok(Json(tokens).into_response())
+    } else {
+        Err((
+            StatusCode::NOT_FOUND,
+            Json(ErrorResponse {
+                error: "No fast tokenizer or tokenizer.json for this model".to_string(),
+                error_type: "no fast tokenizer".to_string(),
+            }),
+        ))
+    }
+}
+
 /// Prometheus metrics scrape endpoint
 #[utoipa::path(
 get,
@@ -719,6 +765,8 @@ pub async fn run(
     compat_generate,
     generate,
     generate_stream,
+    chat_completions,
+    tokenize,
     metrics,
     ),
     components(
@@ -867,6 +915,7 @@ pub async fn run(
         .route("/generate", post(generate))
         .route("/generate_stream", post(generate_stream))
         .route("/v1/chat/completions", post(chat_completions))
+        .route("/tokenize", post(tokenize))
         .route("/health", get(health))
         .route("/ping", get(health))
         .route("/metrics", get(metrics));

router/src/validation.rs

@@ -70,12 +70,11 @@ impl Validation {
     }
 
     #[instrument(skip(self, inputs))]
-    async fn validate_input(
+    pub async fn tokenize(
         &self,
         inputs: String,
         truncate: Option<usize>,
-        max_new_tokens: Option<u32>,
-    ) -> Result<(String, usize, u32), ValidationError> {
+    ) -> Result<Option<(tokenizers::Encoding, String)>, ValidationError> {
         // If we have a fast tokenizer
         if let Some(sender) = &self.sender {
             // Create response channel
@@ -88,7 +87,24 @@ impl Validation {
 
             // Await on response channel
             // Unwrap is safe here
-            let (inputs, input_length) = response_receiver.await.unwrap()?;
+            let encoding = response_receiver.await.unwrap()?;
+            Ok(Some(encoding))
+        } else {
+            Ok(None)
+        }
+    }
+
+    #[instrument(skip(self, inputs))]
+    async fn validate_input(
+        &self,
+        inputs: String,
+        truncate: Option<usize>,
+        max_new_tokens: Option<u32>,
+    ) -> Result<(String, usize, u32), ValidationError> {
+        // If we have a fast tokenizer
+        if let Some((encoding, inputs)) = self.tokenize(inputs.clone(), truncate).await? {
+            // Create response channel
+            let input_length = encoding.len();
 
             // Get total tokens
             let max_new_tokens: u32 = if let Some(max_new_tokens) = max_new_tokens {
@@ -343,36 +359,31 @@ fn tokenizer_worker(tokenizer: Tokenizer, mut receiver: mpsc::UnboundedReceiver<
 
 /// Get input length and optionally truncate it
 fn prepare_input(
-    inputs: String,
+    mut inputs: String,
     truncate: Option<usize>,
     tokenizer: &Tokenizer,
-) -> Result<(String, usize), ValidationError> {
+) -> Result<(tokenizers::Encoding, String), ValidationError> {
     // Get the number of tokens in the input
     let mut encoding = tokenizer
         .encode(inputs.clone(), true)
         .map_err(|err| ValidationError::Tokenizer(err.to_string()))?;
 
     // Optionally truncate
-    let (inputs, input_length) = match truncate {
-        // Truncate is some and < encoding length
-        Some(truncate) if truncate < encoding.len() => {
-            // truncate encoding and decode new inputs
+    if let Some(truncate) = truncate {
+        if truncate < encoding.len() {
             encoding.truncate(truncate, 0, TruncationDirection::Left);
-            let inputs = tokenizer
+            inputs = tokenizer
                 .decode(encoding.get_ids(), false)
                 .map_err(|err| ValidationError::Tokenizer(err.to_string()))?;
-            (inputs, encoding.len())
         }
-        // Nothing to do
-        _ => (inputs, encoding.len()),
-    };
+    }
 
-    Ok((inputs, input_length))
+    Ok((encoding, inputs))
 }
 
 type TokenizerRequest = (
     (String, Option<usize>),
-    oneshot::Sender<Result<(String, usize), ValidationError>>,
+    oneshot::Sender<Result<(tokenizers::Encoding, String), ValidationError>>,
     Span,
 );