Do not use Sessions for LLM Inference app (#456)

Sessions lead to crashes in append-only chat apps. We use should the non-Session API as this creates a new implicit session for every message, thereby reducing the chance that we run out of tokens in the KV cache.

examples/llm_inference/ios/InferenceExample/ConversationViewModel.swift

@@ -119,7 +119,7 @@ class ConversationViewModel: ObservableObject {
     do {
       let model = try OnDeviceModel(model: Model.gemma)
       self.model = model
-      chat = try Chat(model: model)
+      chat = try Chat(inference: model.inference)
     } catch let error as InferenceError {
       self.error = error
     } catch {
@@ -154,7 +154,7 @@ class ConversationViewModel: ObservableObject {
       return
     }
     do {
-      chat = try Chat(model: model)
+      chat = try Chat(inference: model.inference)
       messages.removeAll()
     } catch {
       self.error = InferenceError.mediaPipeTasksError(error: error)

examples/llm_inference/ios/InferenceExample/OnDeviceModel.swift

@@ -29,15 +29,11 @@ struct OnDeviceModel {
 /// Represents a chat session using an instance of `OnDeviceModel`.  It manages a MediaPipe 
 /// `LlmInference.Session` under the hood and passes all response generation queries to the session.
 final class Chat {
-  /// The on device model using which this chat session was created.
-  private let model: OnDeviceModel
-
-  /// MediaPipe session managed by the current instance.
-  private var session: LlmInference.Session
-
-  init(model: OnDeviceModel) throws {
-    self.model = model
-    session = try LlmInference.Session(llmInference: model.inference)
+  /// The on device inference engine for this chat session
+  private let inference: LlmInference
+
+  init(inference: LlmInference) throws {
+    self.inference = inference
   }
 
   /// Sends a streaming response generation query to the underlying MediaPipe 
@@ -48,8 +44,7 @@ final class Chat {
   /// - Throws: A MediaPipe `GenAiInferenceError` if the query cannot be added to the current 
   /// session.
   func sendMessage(_ text: String) async throws -> AsyncThrowingStream<String, any Error> {
-    try session.addQueryChunk(inputText: text)
-    let resultStream = session.generateResponseAsync()
+    let resultStream = inference.generateResponseAsync(inputText: text)
     return resultStream
   }
 }