No public description

PiperOrigin-RevId: 715054197

mediapipe/tasks/cc/genai/inference/c/llm_inference_engine.h

@@ -107,6 +107,9 @@ typedef struct {
   // initialization may finish before weights have finished uploading which
   // might push some of the weight upload time into input processing.
   bool wait_for_weight_uploads;
+
+  // Whether the submodel should be used if available.
+  bool use_submodel;
 } LlmModelSettings;
 
 // LlmSessionConfig configures how to execute the model.

mediapipe/tasks/ios/genai/inference/sources/LlmInference.swift

@@ -76,7 +76,8 @@ import MediaPipeTasksGenAIC
             max_top_k: options.maxTopk,
             llm_activation_data_type: kLlmActivationDataTypeDefault,
             num_draft_tokens: 0,
-            wait_for_weight_uploads: options.waitForWeightUploads)
+            wait_for_weight_uploads: options.waitForWeightUploads,
+            use_submodel: options.useSubmodel)
           return try LlmTaskRunner(modelSettings: modelSetting)
         }
       }
@@ -240,6 +241,9 @@ extension LlmInference {
     /// time into input processing.
     @objc public var waitForWeightUploads: Bool = false
 
+    // Whether to use the submodel if available.
+    @objc public var useSubmodel: Bool = false
+
     /// Creates a new instance of `Options` with the given `modelPath` and default values of
     /// `maxTokens`, `maxTopk`, `supportedLoraRanks`.
     /// This function is only intended to be used from Objective C.

mediapipe/tasks/java/com/google/mediapipe/tasks/core/jni/llm.cc

@@ -77,6 +77,7 @@ LlmModelSettings ParseModelSettings(void* bytes, int size) {
   output.llm_activation_data_type = kLlmActivationDataTypeDefault;
   output.num_draft_tokens = 0;
   output.wait_for_weight_uploads = false;
+  output.use_submodel = false;
   return output;
 }