feat: Implement frontend tokenization for embedding requests (#1494)

launch/dynamo-run/src/subprocess/sglang_inc.py

@@ -13,6 +13,7 @@ from typing import Optional
 
 import sglang
 import uvloop
+from sglang.srt.entrypoints.engine import EmbeddingReqInput
 from sglang.srt.server_args import ServerArgs
 
 from dynamo.llm import ModelType, register_llm
@@ -126,38 +127,21 @@ class RequestHandler:
 
             yield out
 
+    async def encode(self, request):
+        obj = EmbeddingReqInput(input_ids=request["token_ids"])
+        generator = self.engine_client.tokenizer_manager.generate_request(obj, None)
+        engine_results = await anext(generator)
 
-class EmbeddingRequestHandler(RequestHandler):
-    """
-    Request handler for the embedding endpoint
-    """
-
-    def __init__(self, engine: sglang.Engine, model_name: str):
-        super().__init__(engine)
-        self._model_name = model_name
-
-    async def generate(self, request):
-        gen = await self.engine_client.async_encode(prompt=request["input"])
         tokens = 0
         embeddings = []
-        for idx, res in enumerate(gen):
-            embeddings.append(
-                {
-                    "index": idx,
-                    "object": "embedding",
-                    "embedding": res["embedding"],
-                }
-            )
-            tokens += res["meta_info"]["prompt_tokens"]
+        for result in engine_results:
+            embeddings.append(result["embedding"])
+            tokens += result["meta_info"]["prompt_tokens"]
 
         out = {
-            "object": "list",
-            "model": self._model_name,
-            "data": embeddings,
-            "usage": {
+            "embeddings": embeddings,
             "prompt_tokens": tokens,
             "total_tokens": tokens,
-            },
         }
 
         yield out
@@ -222,13 +206,11 @@ async def init(runtime: DistributedRuntime, config: Config):
 
     # the server will gracefully shutdown (i.e., keep opened TCP streams finishes)
     # after the lease is revoked
-    await endpoint.serve_endpoint(
-        RequestHandler(engine_client).generate
-        if not engine_args.is_embedding
-        else EmbeddingRequestHandler(
-            engine_client, model_name=config.model_name or config.model_path
-        ).generate
-    )
+    handler = RequestHandler(engine_client)
+    if engine_args.is_embedding:
+        await endpoint.serve_endpoint(handler.encode)
+    else:
+        await endpoint.serve_endpoint(handler.generate)
 
 
 def cmd_line_args():

lib/llm/src/backend.rs

@@ -44,7 +44,11 @@ use dynamo_runtime::{
 
 use crate::protocols::{
     common::{
-        llm_backend::{BackendOutput, FinishReason, LLMEngineOutput, PreprocessedRequest},
+        llm_backend::{
+            BackendOutput, EmbeddingsEngineOutput, FinishReason, LLMEngineOutput,
+            PreprocessedRequest,
+        },
+        preprocessor::PreprocessedEmbeddingRequest,
         StopConditions,
     },
     TokenIdType,
@@ -233,6 +237,36 @@ impl
     }
 }
 
+#[async_trait]
+impl
+    Operator<
+        SingleIn<PreprocessedEmbeddingRequest>,
+        ManyOut<Annotated<EmbeddingsEngineOutput>>,
+        SingleIn<PreprocessedEmbeddingRequest>,
+        ManyOut<Annotated<EmbeddingsEngineOutput>>,
+    > for Backend
+{
+    async fn generate(
+        &self,
+        request: SingleIn<PreprocessedEmbeddingRequest>,
+        next: ServerStreamingEngine<
+            PreprocessedEmbeddingRequest,
+            Annotated<EmbeddingsEngineOutput>,
+        >,
+    ) -> Result<ManyOut<Annotated<EmbeddingsEngineOutput>>> {
+        // For embeddings, we mostly pass through since no detokenization is needed
+        // But we could add validation, logging, or other post-processing here
+        let response_stream = next.generate(request).await?;
+
+        // Could add embedding-specific post-processing here:
+        // - Validation of embedding dimensions
+        // - Normalization if requested
+        // - Usage statistics validation
+
+        Ok(response_stream)
+    }
+}
+
 // todo - add visible stop conditions
 // visible_stop_ids: HashSet<TokenIdType>,
 // visible_stop_sequences: Vec<String>,

lib/llm/src/discovery/watcher.rs

@@ -20,8 +20,8 @@ use crate::{
     backend::Backend,
     kv_router::{KvPushRouter, KvRouterConfig},
     model_type::ModelType,
-    preprocessor::{OpenAIPreprocessor, PreprocessedRequest},
-    protocols::common::llm_backend::LLMEngineOutput,
+    preprocessor::{OpenAIPreprocessor, PreprocessedEmbeddingRequest, PreprocessedRequest},
+    protocols::common::llm_backend::{EmbeddingsEngineOutput, LLMEngineOutput},
     protocols::openai::chat_completions::{
         NvCreateChatCompletionRequest, NvCreateChatCompletionStreamResponse,
     },
@@ -300,14 +300,42 @@ impl ModelWatcher {
                     .add_completions_model(&model_entry.name, engine)?;
             }
             ModelType::Embedding => {
-                let push_router = PushRouter::<
-                    NvCreateEmbeddingRequest,
-                    Annotated<NvCreateEmbeddingResponse>,
-                >::from_client(client, Default::default())
+                let Some(mut card) = card else {
+                    anyhow::bail!("Missing model deployment card for embedding model");
+                };
+
+                // Download tokenizer files to local disk
+                let _cache_dir = Some(card.move_from_nats(self.drt.nats_client()).await?);
+
+                // Create preprocessing pipeline similar to Backend
+                let frontend = SegmentSource::<
+                    SingleIn<NvCreateEmbeddingRequest>,
+                    ManyOut<Annotated<NvCreateEmbeddingResponse>>,
+                >::new();
+
+                let preprocessor = OpenAIPreprocessor::new(card.clone()).await?.into_operator();
+                let backend = Backend::from_mdc(card.clone()).await?.into_operator();
+
+                let router = PushRouter::<
+                    PreprocessedEmbeddingRequest,
+                    Annotated<EmbeddingsEngineOutput>,
+                >::from_client(client, self.router_mode)
                 .await?;
-                let engine = Arc::new(push_router);
+
+                // Note: Embeddings don't need KV routing complexity
+                let service_backend = ServiceBackend::from_engine(Arc::new(router));
+
+                // Link the pipeline: frontend -> preprocessor -> backend -> service_backend -> backend -> preprocessor -> frontend
+                let embedding_engine = frontend
+                    .link(preprocessor.forward_edge())?
+                    .link(backend.forward_edge())?
+                    .link(service_backend)?
+                    .link(backend.backward_edge())?
+                    .link(preprocessor.backward_edge())?
+                    .link(frontend)?;
+
                 self.manager
-                    .add_embeddings_model(&model_entry.name, engine)?;
+                    .add_embeddings_model(&model_entry.name, embedding_engine)?;
             }
         }
 

lib/llm/src/preprocessor.rs

@@ -27,6 +27,7 @@ pub mod prompt;
 pub mod tools;
 
 use anyhow::Result;
+use async_openai::types::EncodingFormat;
 use futures::stream::{self, StreamExt};
 use prompt::OAIPromptFormatter;
 use rayon::iter::{IntoParallelRefIterator, ParallelIterator};
@@ -48,6 +49,7 @@ use crate::protocols::{
     openai::{
         chat_completions::{NvCreateChatCompletionRequest, NvCreateChatCompletionStreamResponse},
         completions::{NvCreateCompletionRequest, NvCreateCompletionResponse},
+        embeddings::{NvCreateEmbeddingRequest, NvCreateEmbeddingResponse},
         nvext::NvExtProvider,
         DeltaGeneratorExt,
     },
@@ -57,6 +59,9 @@ use crate::tokenizers::{traits::Tokenizer, HuggingFaceTokenizer};
 use crate::preprocessor::prompt::{PromptFormatter, PromptInput, TextInput, TokenInput};
 
 pub use crate::protocols::common::llm_backend::{BackendOutput, PreprocessedRequest};
+pub use crate::protocols::common::preprocessor::PreprocessedEmbeddingRequest;
+
+use crate::protocols::common::llm_backend::EmbeddingsEngineOutput;
 
 pub const ANNOTATION_FORMATTED_PROMPT: &str = "formatted_prompt";
 pub const ANNOTATION_TOKEN_IDS: &str = "token_ids";
@@ -265,6 +270,68 @@ impl OpenAIPreprocessor {
         Ok((builder.build()?, annotations))
     }
 
+    /// Preprocess an embedding request, handling both text and token ID inputs.
+    ///
+    /// For text inputs, tokenizes the text using the configured tokenizer.
+    /// For token ID inputs, uses the provided token IDs directly and skips tokenization.
+    ///
+    /// Returns both the preprocessed request and a hashmap of annotations.
+    pub async fn preprocess_embedding_request(
+        &self,
+        request: &NvCreateEmbeddingRequest,
+    ) -> Result<(PreprocessedEmbeddingRequest, HashMap<String, String>)> {
+        let mut annotations = HashMap::new();
+        let mut builder = PreprocessedEmbeddingRequest::builder();
+
+        let all_token_ids = match &request.inner.input {
+            async_openai::types::EmbeddingInput::String(s) => {
+                let encoding = tokio::task::block_in_place(|| self.tokenizer.encode(s))?;
+                vec![encoding.token_ids]
+            }
+            async_openai::types::EmbeddingInput::StringArray(arr) => {
+                let input_strs: Vec<String> = arr.to_vec();
+                let encodings = tokio::task::spawn_blocking({
+                    let tokenizer = self.tokenizer.clone();
+                    let strs = input_strs.clone();
+                    move || {
+                        tokenizer.encode_batch(&strs.iter().map(|s| s.as_str()).collect::<Vec<_>>())
+                    }
+                })
+                .await??;
+                let token_arrays: Vec<Vec<u32>> = encodings
+                    .into_iter()
+                    .map(|encoding| encoding.token_ids)
+                    .collect();
+                token_arrays
+            }
+            async_openai::types::EmbeddingInput::IntegerArray(token_ids) => vec![token_ids.clone()],
+            async_openai::types::EmbeddingInput::ArrayOfIntegerArray(token_arrays) => {
+                token_arrays.clone()
+            }
+        };
+
+        // Handle annotations
+        if request.has_annotation(ANNOTATION_TOKEN_IDS) {
+            annotations.insert(
+                ANNOTATION_TOKEN_IDS.to_string(),
+                serde_json::to_string(&all_token_ids)?,
+            );
+        }
+
+        builder.token_ids(all_token_ids);
+        builder.model(request.inner.model.clone());
+        builder.encoding_format(request.inner.encoding_format.as_ref().map(|f| match f {
+            EncodingFormat::Float => "float".to_string(),
+            EncodingFormat::Base64 => "base64".to_string(),
+        }));
+        builder.dimensions(request.inner.dimensions);
+
+        builder.annotations(request.annotations().unwrap_or_default());
+        builder.mdc_sum(Some(self.mdcsum.clone()));
+
+        Ok((builder.build()?, annotations))
+    }
+
     pub fn transform_postprocessor_stream<Resp: Send + Sync + 'static + std::fmt::Debug>(
         stream: ManyOut<Annotated<BackendOutput>>,
         generator: Box<dyn DeltaGeneratorExt<Resp>>,
@@ -367,6 +434,46 @@ impl OpenAIPreprocessor {
 
         ResponseStream::new(Box::pin(stream), context)
     }
+
+    /// Transform engine embedding output stream to OpenAI embedding response stream
+    pub fn transform_embedding_postprocessor_stream(
+        stream: ManyOut<Annotated<EmbeddingsEngineOutput>>,
+        original_request: NvCreateEmbeddingRequest,
+    ) -> ManyOut<Annotated<NvCreateEmbeddingResponse>> {
+        let context = stream.context();
+
+        let transformed_stream = stream.map(move |output| {
+            output.map_data(|engine_output| {
+                // Convert engine output to OpenAI response format
+                let embeddings: Vec<async_openai::types::Embedding> = engine_output
+                    .embeddings
+                    .into_iter()
+                    .enumerate()
+                    .map(|(index, embedding)| async_openai::types::Embedding {
+                        index: index as u32,
+                        object: "embedding".to_string(),
+                        embedding: embedding.into_iter().map(|f| f as f32).collect(),
+                    })
+                    .collect();
+
+                let response = NvCreateEmbeddingResponse {
+                    inner: async_openai::types::CreateEmbeddingResponse {
+                        object: "list".to_string(),
+                        model: original_request.inner.model.clone(),
+                        data: embeddings,
+                        usage: async_openai::types::EmbeddingUsage {
+                            prompt_tokens: engine_output.prompt_tokens,
+                            total_tokens: engine_output.total_tokens,
+                        },
+                    },
+                };
+
+                Ok(response)
+            })
+        });
+
+        ResponseStream::new(Box::pin(transformed_stream), context)
+    }
 }
 
 // for pals, we do not want to add the generation prompt to the formatted prompt
@@ -488,3 +595,51 @@ impl
         Ok(ResponseStream::new(Box::pin(stream), context))
     }
 }
+
+#[async_trait]
+impl
+    Operator<
+        SingleIn<NvCreateEmbeddingRequest>,
+        ManyOut<Annotated<NvCreateEmbeddingResponse>>,
+        SingleIn<PreprocessedEmbeddingRequest>,
+        ManyOut<Annotated<EmbeddingsEngineOutput>>,
+    > for OpenAIPreprocessor
+{
+    async fn generate(
+        &self,
+        request: SingleIn<NvCreateEmbeddingRequest>,
+        next: Arc<
+            dyn AsyncEngine<
+                SingleIn<PreprocessedEmbeddingRequest>,
+                ManyOut<Annotated<EmbeddingsEngineOutput>>,
+                Error,
+            >,
+        >,
+    ) -> Result<ManyOut<Annotated<NvCreateEmbeddingResponse>>, Error> {
+        // Unpack request
+        let (request, context) = request.into_parts();
+
+        // Preprocess the embedding request
+        let (preprocessed_request, annotations) =
+            self.preprocess_embedding_request(&request).await?;
+
+        // Forward to next stage
+        let preprocessed_request = context.map(|_| preprocessed_request);
+        let response_stream = next.generate(preprocessed_request).await?;
+
+        // Transform response stream back to OpenAI format
+        let stream = Self::transform_embedding_postprocessor_stream(response_stream, request);
+        let context = stream.context();
+
+        // Prepend annotations
+        let annotations_stream = stream::iter(
+            annotations
+                .into_iter()
+                .flat_map(|(k, v)| Annotated::from_annotation(k, &v))
+                .collect::<Vec<_>>(),
+        );
+
+        let combined_stream = annotations_stream.chain(stream);
+        Ok(ResponseStream::new(Box::pin(combined_stream), context))
+    }
+}

lib/llm/src/protocols/common/llm_backend.rs

@@ -133,3 +133,14 @@ impl LLMEngineOutput {
         }
     }
 }
+
+/// Raw output from embedding engines containing embedding vectors
+#[derive(Serialize, Deserialize, Debug, Clone, PartialEq)]
+pub struct EmbeddingsEngineOutput {
+    /// Generated embedding vectors (one per input text)
+    pub embeddings: Vec<Vec<f64>>,
+
+    /// Token usage information
+    pub prompt_tokens: u32,
+    pub total_tokens: u32,
+}

lib/llm/src/protocols/common/preprocessor.rs

@@ -68,3 +68,40 @@ impl PreprocessedRequest {
         PreprocessedRequestBuilder::default()
     }
 }
+
+/// [`PreprocessedEmbeddingRequest`] is the internal representation of an embedding request
+/// after preprocessing. Contains tokenized input ready for embedding engines.
+#[derive(Serialize, Deserialize, Debug, Clone, Builder)]
+pub struct PreprocessedEmbeddingRequest {
+    /// Tokenized input text as token IDs (one Vec per input text)
+    pub token_ids: Vec<Vec<TokenIdType>>,
+
+    /// Model to use for embedding
+    pub model: String,
+
+    /// Encoding format preference
+    pub encoding_format: Option<String>,
+
+    /// Number of dimensions for output embeddings (if supported)
+    pub dimensions: Option<u32>,
+
+    /// The computed checksum of the Model Deployment Card (MDC)
+    #[builder(default)]
+    pub mdc_sum: Option<String>,
+
+    /// User requested annotations for the request
+    #[builder(default)]
+    pub annotations: Vec<String>,
+}
+
+impl PreprocessedEmbeddingRequest {
+    pub fn has_annotation(&self, annotation: &str) -> bool {
+        self.annotations.contains(&annotation.to_string())
+    }
+}
+
+impl PreprocessedEmbeddingRequest {
+    pub fn builder() -> PreprocessedEmbeddingRequestBuilder {
+        PreprocessedEmbeddingRequestBuilder::default()
+    }
+}

lib/llm/src/tokenizers.rs

@@ -58,6 +58,7 @@ pub mod traits {
 
     pub trait Encoder: Send + Sync {
         fn encode(&self, input: &str) -> Result<Encoding>;
+        fn encode_batch(&self, inputs: &[&str]) -> Result<Vec<Encoding>>;
     }
 
     pub trait Decoder: Send + Sync {

lib/llm/src/tokenizers/hf.rs

@@ -54,6 +54,30 @@ impl Encoder for HuggingFaceTokenizer {
             spans,
         })
     }
+
+    fn encode_batch(&self, inputs: &[&str]) -> Result<Vec<Encoding>> {
+        let hf_encodings = self
+            .tokenizer
+            .encode_batch(inputs.to_vec(), false)
+            .map_err(|err| Error::msg(format!("Error encoding input: {}", err)))?;
+
+        let encodings = hf_encodings
+            .into_iter()
+            .map(|encoding| {
+                let token_ids = encoding.get_ids().to_vec();
+                let tokens = encoding.get_tokens().to_vec();
+                let spans = encoding.get_offsets().to_vec();
+
+                Encoding {
+                    token_ids,
+                    tokens,
+                    spans,
+                }
+            })
+            .collect();
+
+        Ok(encodings)
+    }
 }
 
 impl Decoder for HuggingFaceTokenizer {

lib/llm/src/tokenizers/sp.rs

@@ -73,6 +73,17 @@ impl Encoder for SentencePieceTokenizer {
             spans,
         })
     }
+
+    /// Encodes multiple string inputs into tokens using the SentencePiece model.
+    ///
+    /// # Arguments
+    /// * `inputs` - The texts to encode
+    ///
+    /// # Returns
+    /// * `Result<Vec<Encoding>>` - The encoded tokens for each input
+    fn encode_batch(&self, inputs: &[&str]) -> Result<Vec<Encoding>> {
+        inputs.iter().map(|input| self.encode(input)).collect()
+    }
 }
 
 impl Decoder for SentencePieceTokenizer {