Fix Ollama multi-turn, tool calling, and cross-backend tool argument encoding

Sources/AnyLanguageModel/LanguageModelSession.swift

@@ -123,19 +123,19 @@ public final class LanguageModelSession: @unchecked Sendable {
             state.withLock { $0.endResponding() }
         }
     }
-    
+
     nonisolated func growStreamingTranscript(text: String) {
         withMutation(keyPath: \.transcript) {
             state.withLock { $0.transcript.appendStreamingResponse(text) }
         }
     }
-    
+
     nonisolated func appendTranscriptEntry(_ entry: Transcript.Entry) {
         withMutation(keyPath: \.transcript) {
             state.withLock { $0.transcript.append(entry) }
         }
     }
-    
+
     nonisolated private func wrapRespond<T>(_ operation: () async throws -> T) async throws -> T {
         beginResponding()
         do {
@@ -176,7 +176,9 @@ public final class LanguageModelSession: @unchecked Sendable {
                         }
 
                         session.withMutation(keyPath: \.transcript) {
-                            session.state.withLock { $0.transcript.finalizeStreamedTranscript(textContent, assetIDs: []) }
+                            session.state.withLock {
+                                $0.transcript.finalizeStreamedTranscript(textContent, assetIDs: [])
+                            }
                         }
                     }
                 } catch {

Sources/AnyLanguageModel/Models/GeminiLanguageModel.swift

@@ -723,7 +723,9 @@ private func toGeneratedContent(_ value: [String: JSONValue]?) throws -> Generat
 }
 
 private func fromGeneratedContent(_ content: GeneratedContent) throws -> [String: JSONValue] {
-    let data = try JSONEncoder().encode(content)
+    // `GeneratedContent`'s `Codable` conformance is a lossless persistence format for
+    // round-tripping a `Transcript`, not a wire format, so go through `jsonString`.
+    let data = Data(content.jsonString.utf8)
     let jsonValue = try JSONDecoder().decode(JSONValue.self, from: data)
 
     guard case .object(let dict) = jsonValue else {

Sources/AnyLanguageModel/Models/LlamaLanguageModel.swift

@@ -672,9 +672,9 @@ import Foundation
             // Force CPU-only execution to avoid Metal GPU issues
             params.n_gpu_layers = 0
 
-            // Try to reduce memory usage
-            params.use_mmap = true
-            params.use_mlock = false
+            // Try to reduce memory usage by memory-mapping the weights without locking them
+            // in RAM. Replaces the separate `use_mmap` / `use_mlock` flags removed upstream.
+            params.load_mode = LLAMA_LOAD_MODE_MMAP
             return params
         }