Inline images for multimodal models. (#1666)

Author: Narsil

Cargo.lock

@@ -20,13 +20,14 @@ async def idefics(idefics_handle):
 def get_chicken():
     with open("integration-tests/images/chicken_on_money.png", "rb") as image_file:
         encoded_string = base64.b64encode(image_file.read())
-    return f"data:image/png;base64,{encoded_string}"
+    return f"data:image/png;base64,{encoded_string.decode('utf-8')}"
 
 
 @pytest.mark.asyncio
 async def test_idefics(idefics, response_snapshot):
+    chicken = get_chicken()
     response = await idefics.generate(
-        "User:![](https://huggingface.co/spaces/HuggingFaceM4/idefics_playground/resolve/main/example_images/chicken_on_money.png?download=true)Can you tell me a very short story based on the image?",
+        f"User:![]({chicken})Can you tell me a very short story based on the image?",
         max_new_tokens=10,
         decoder_input_details=True,
     )
@@ -37,9 +38,10 @@ async def test_idefics(idefics, response_snapshot):
 
 @pytest.mark.asyncio
 async def test_idefics_load(idefics, generate_load, response_snapshot):
+    chicken = get_chicken()
     responses = await generate_load(
         idefics,
-        "User:![](https://huggingface.co/spaces/HuggingFaceM4/idefics_playground/resolve/main/example_images/chicken_on_money.png?download=true)Can you tell me a very short story based on the image?",
+        f"User:![]({chicken})Can you tell me a very short story based on the image?",
         max_new_tokens=10,
         n=4,
     )

integration-tests/models/test_idefics.py

@@ -46,6 +46,8 @@ ngrok = { version = "0.13.1", features = ["axum"], optional = true }
 init-tracing-opentelemetry = { version = "0.14.1", features = ["opentelemetry-otlp"] }
 minijinja = { git = "https://github.com/mitsuhiko/minijinja.git", branch = "main", commit = "5cd4efb" }
 futures-util = "0.3.30"
+regex = "1.10.3"
+once_cell = "1.19.0"
 
 [build-dependencies]
 vergen = { version = "8.2.5", features = ["build", "git", "gitcl"] }

router/Cargo.toml

@@ -13,6 +13,7 @@ use tokenizers::TruncationDirection;
 use tokio::sync::mpsc;
 use tokio::sync::oneshot;
 use tracing::{instrument, Span};
+use {once_cell::sync::Lazy, regex::Regex};
 
 /// Validation
 #[derive(Debug, Clone)]
@@ -409,10 +410,14 @@ async fn round_robin_task(
 /// Start tokenization workers
 fn tokenizer_worker(tokenizer: Tokenizer, mut receiver: mpsc::UnboundedReceiver<TokenizerRequest>) {
     // Loop over requests
+    let is_multimodal = {
+        let vocab = tokenizer.get_vocab(true);
+        vocab.contains_key("<image>")
+    };
     while let Some(((inputs, truncate), response_tx, parent_span)) = receiver.blocking_recv() {
         parent_span.in_scope(|| {
             response_tx
-                .send(prepare_input(inputs, truncate, &tokenizer))
+                .send(prepare_input(inputs, truncate, &tokenizer, is_multimodal))
                 .unwrap_or(())
         })
     }
@@ -423,15 +428,22 @@ fn prepare_input(
     mut inputs: String,
     truncate: Option<usize>,
     tokenizer: &Tokenizer,
+    is_multimodal: bool,
 ) -> Result<(tokenizers::Encoding, String), ValidationError> {
+    let simplified_query = if is_multimodal {
+        static RE: Lazy<Regex> = Lazy::new(|| Regex::new(r"!\[\]\([^\)]*\)").unwrap());
+        RE.replace_all(&inputs, "<image>").into()
+    } else {
+        inputs.clone()
+    };
     // Get the number of tokens in the input
     let mut encoding = tokenizer
-        .encode(inputs.clone(), true)
+        .encode(simplified_query, true)
         .map_err(|err| ValidationError::Tokenizer(err.to_string()))?;
 
     // Optionally truncate
     if let Some(truncate) = truncate {
-        if truncate < encoding.len() {
+        if truncate < encoding.len() && !is_multimodal {
             encoding.truncate(truncate, 0, TruncationDirection::Left);
             inputs = tokenizer
                 .decode(encoding.get_ids(), false)