[Rust Frontend] Bump llm-multimodal version (#47530)

Signed-off-by: Isotr0py <Isotr0py@outlook.com>

rust/src/chat/src/backend/hf.rs

@@ -303,7 +303,7 @@ mod tests {
             .unwrap()
             .join("preprocessor_config.json");
         write_json(&preprocessor_config_path, r#"{"size":[672,672]}"#);
-        files.preprocessor_config_path = Some(preprocessor_config_path);
+        files.preprocessor_config_path = Some(preprocessor_config_path.clone());
 
         let backend = HfChatBackend::from_resolved_model_files(
             files.clone(),
@@ -321,6 +321,9 @@ mod tests {
 
         assert!(backend.multimodal_model_info().is_none());
 
+        let invalid_preprocessor_config = r#"{"size":[672,672]"#;
+        write_json(&preprocessor_config_path, invalid_preprocessor_config);
+
         let error = HfChatBackend::from_resolved_model_files(
             files,
             "test-model".to_string(),

rust/src/chat/src/multimodal.rs

@@ -16,10 +16,11 @@ use std::sync::{Arc, LazyLock};
 
 use itertools::izip;
 use llm_multimodal::{
-    AsyncMultiModalTracker, FieldLayout, ImagePreProcessor, ImageProcessorRegistry, MediaConnector,
-    MediaConnectorConfig, MediaContentPart, Modality, ModelMetadata, ModelProcessorSpec,
-    ModelRegistry, PreProcessorConfig, PreprocessedImages, PromptReplacement, TokenResolver,
-    TrackedMedia,
+    AsyncMultiModalTracker, FieldLayout, MediaConnector, MediaConnectorConfig, MediaContentPart,
+    Modality, ModelMetadata, ModelProcessorSpec, ModelRegistry, PreProcessorConfig,
+    PreprocessedEncoderInputs as PreprocessedImages, PromptReplacement, Tokenizer as TokenResolver,
+    TrackedMedia, VisionPreProcessor as ImagePreProcessor,
+    VisionProcessorRegistry as ImageProcessorRegistry,
 };
 use tracing::warn;
 use vllm_engine_core_client::protocol::dtype::ModelDtype;
@@ -555,6 +556,10 @@ impl TokenResolver for TokenizerResolver {
     fn id_to_token(&self, id: u32) -> Option<String> {
         self.0.id_to_token(id)
     }
+
+    fn encode_text(&self, text: &str) -> Option<Vec<u32>> {
+        self.0.encode(text, false).ok()
+    }
 }
 
 #[cfg(test)]

rust/src/chat/src/multimodal/tensor.rs

@@ -1,7 +1,7 @@
 use std::collections::HashMap;
 
 use half::{bf16, f16};
-use llm_multimodal::{ModelSpecificValue, PreprocessedImages};
+use llm_multimodal::{ModelSpecificValue, PreprocessedEncoderInputs as PreprocessedImages};
 use vllm_engine_core_client::protocol::dtype::ModelDtype;
 use vllm_engine_core_client::protocol::multimodal::MmKwargValue as ProtocolKwargValue;
 use vllm_engine_core_client::protocol::tensor::{ShapeExt as _, WireTensor};
@@ -31,14 +31,14 @@ pub(super) fn collect_tensors(
     float_dtype: ModelDtype,
 ) -> Result<HashMap<String, KwargValue>> {
     let PreprocessedImages {
-        pixel_values,
+        encoder_input,
         model_specific,
         ..
     } = preprocessed;
 
     let pixel_values = {
-        let shape = pixel_values.shape().to_vec();
-        let data = pixel_values.into_iter().collect();
+        let shape = encoder_input.shape().to_vec();
+        let data = encoder_input.into_iter().collect();
         KwargValue::from_f32_tensor(data, shape, float_dtype)?
     };
 

rust/src/server/src/routes/tests.rs

@@ -559,7 +559,7 @@ fn qwen_multimodal_model_info() -> vllm_chat::multimodal::MultimodalModelInfo {
     ));
     fs::write(
         &config_path,
-        r#"{"model_type":"qwen2_vl","vision_token_id":151655}"#,
+        r#"{"model_type":"qwen2_vl","image_token_id":151655}"#,
     )
     .expect("write qwen test config");
     let info = vllm_chat::multimodal::MultimodalModelInfo::from_paths(