fix: GGUF local model GPU crashes and rough edges

android/app/src/main/cpp/gguf_runtime.cpp

@@ -431,9 +431,10 @@ static CreateResult try_create_session(const char* path, int n_gpu_layers, int n
         cparams.abort_callback = abort_callback;
         cparams.abort_callback_data = &operation->abort_flag;
         if (n_gpu_layers > 0) {
-            cparams.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_DISABLED;
-            cparams.type_k = GGML_TYPE_F16;
-            cparams.type_v = GGML_TYPE_F16;
+            // Standard attention hard-aborts on some Vulkan GPUs (uncatchable); flash attn avoids it and requires quantized KV.
+            cparams.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_ENABLED;
+            cparams.type_k = GGML_TYPE_Q4_0;
+            cparams.type_v = GGML_TYPE_Q4_0;
             cparams.offload_kqv = true;
         } else {
             cparams.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_ENABLED;
@@ -792,6 +793,8 @@ Java_dev_androidagent_localmodel_gguf_GgufNative_generate(
         sampler_guard.smpl = llama_sampler_chain_init(sparams);
         llama_sampler_chain_add(sampler_guard.smpl, llama_sampler_init_top_k(static_cast<int32_t>(top_k)));
         llama_sampler_chain_add(sampler_guard.smpl, llama_sampler_init_top_p(static_cast<float>(top_p), 1));
+        // Prevents small-model repetition loops; must come after top-k/top-p per llama.cpp's guidance.
+        llama_sampler_chain_add(sampler_guard.smpl, llama_sampler_init_penalties(64, 1.1f, 0.0f, 0.0f));
         llama_sampler_chain_add(sampler_guard.smpl, llama_sampler_init_temp(static_cast<float>(temperature)));
         llama_sampler_chain_add(sampler_guard.smpl, llama_sampler_init_dist(static_cast<uint32_t>(time(nullptr))));
 

android/app/src/main/java/dev/androidagent/AppShellActivity.kt

@@ -33,6 +33,7 @@ import androidx.core.view.WindowInsetsControllerCompat
 import androidx.lifecycle.lifecycleScope
 import dev.androidagent.chat.ChatAttachmentKind
 import dev.androidagent.chat.ChatAttachmentStore
+import dev.androidagent.localmodel.LocalModelImportStatus
 import dev.androidagent.localmodel.LocalModelStore
 import dev.androidagent.settings.DiagnosticsEventLog
 import dev.androidagent.settings.DiagnosticsEventLevel
@@ -102,6 +103,7 @@ class AppShellActivity : ComponentActivity() {
         }
         localModelImportJob?.cancel()
         DiagnosticsEventLog.append(DiagnosticsEventLevel.Info, "Importing local model")
+        LocalModelImportStatus.publish("Importing local model…")
         localModelImportJob = lifecycleScope.launch {
             try {
                 var lastProgressMarker = -1L
@@ -116,14 +118,21 @@ class AppShellActivity : ComponentActivity() {
                             "${copied / (1024 * 1024)} MB"
                         }
                         DiagnosticsEventLog.append(DiagnosticsEventLevel.Info, "Importing local model: $label")
+                        LocalModelImportStatus.publish("Importing local model… $label")
                     }
                 }
                 pendingLocalModelPathField?.setText(path)
                 DiagnosticsEventLog.append(DiagnosticsEventLevel.Success, "Imported local model")
+                LocalModelImportStatus.publish("Imported ✓ ${LocalModelStore.displayName(path)}")
+                mainHandler.postDelayed({ LocalModelImportStatus.publish(null) }, 2500)
             } catch (cancelled: CancellationException) {
+                LocalModelImportStatus.publish(null)
                 throw cancelled
             } catch (error: Exception) {
-                DiagnosticsEventLog.append(DiagnosticsEventLevel.Error, error.message ?: "Import failed")
+                val message = error.message ?: "Import failed"
+                DiagnosticsEventLog.append(DiagnosticsEventLevel.Error, message)
+                LocalModelImportStatus.publish("Import failed: $message")
+                mainHandler.postDelayed({ LocalModelImportStatus.publish(null) }, 4000)
             }
         }
     }

android/app/src/main/java/dev/androidagent/MainActivity.kt

@@ -28,6 +28,7 @@ import androidx.core.view.WindowInsetsCompat
 import androidx.core.view.WindowInsetsControllerCompat
 import androidx.lifecycle.lifecycleScope
 import dev.androidagent.avatar.AvatarLibrary
+import dev.androidagent.localmodel.LocalModelImportStatus
 import dev.androidagent.localmodel.LocalModelStore
 import dev.androidagent.settings.SettingsHost
 import dev.androidagent.settings.SettingsUi
@@ -67,21 +68,31 @@ class MainActivity : ComponentActivity() {
         }
         localModelImportJob?.cancel()
         setupBanner.text = "Importing local model..."
+        LocalModelImportStatus.publish("Importing local model…")
         localModelImportJob = lifecycleScope.launch {
             try {
                 var lastProgressMarker = -1L
                 val path = LocalModelStore.importModel(this@MainActivity, uri) { copied, total ->
                     val marker = if (total != null && total > 0L) copied * 100L / total else copied / PROGRESS_STEP_BYTES
                     if (marker == lastProgressMarker) return@importModel
                     lastProgressMarker = marker
-                    mainHandler.post { setupBanner.text = modelImportProgress(copied, total) }
+                    mainHandler.post {
+                        setupBanner.text = modelImportProgress(copied, total)
+                        LocalModelImportStatus.publish(modelImportProgress(copied, total))
+                    }
                 }
                 pendingLocalModelPathField?.setText(path)
                 setupBanner.text = "Imported local model."
+                LocalModelImportStatus.publish("Imported ✓ ${LocalModelStore.displayName(path)}")
+                mainHandler.postDelayed({ LocalModelImportStatus.publish(null) }, 2500)
             } catch (cancelled: CancellationException) {
+                LocalModelImportStatus.publish(null)
                 throw cancelled
             } catch (error: Exception) {
-                setupBanner.text = error.message ?: "Could not import local model."
+                val message = error.message ?: "Could not import local model."
+                setupBanner.text = message
+                LocalModelImportStatus.publish("Import failed: $message")
+                mainHandler.postDelayed({ LocalModelImportStatus.publish(null) }, 4000)
             }
         }
     }

android/app/src/main/java/dev/androidagent/localmodel/LocalAgentController.kt

@@ -115,7 +115,8 @@ class LocalAgentController(
         val systemPrompt = LocalPromptBuilder.systemPrompt(
             basePrompt = config.systemPrompt,
             toolsAllowed = toolsAllowed,
-            toolDescriptionsJson = tools.toolDescriptions(runtimeProfile, toolAccess).toString()
+            toolDescriptionsJson = tools.toolDescriptions(runtimeProfile, toolAccess).toString(),
+            isTinyModel = LocalModelSize.isTiny(LocalModelStore.displayName(config.localModelPath))
         )
         val transcript = selectNewestHistory(
             history = history,

android/app/src/main/java/dev/androidagent/localmodel/LocalModelImportStatus.kt

@@ -0,0 +1,19 @@
+package dev.androidagent.localmodel
+
+// Only one screen is visible at a time, so a single overwrite-on-build slot is enough.
+object LocalModelImportStatus {
+    @Volatile
+    private var listener: ((String?) -> Unit)? = null
+
+    fun observe(onUpdate: (String?) -> Unit) {
+        listener = onUpdate
+    }
+
+    fun clearObserver(onUpdate: (String?) -> Unit) {
+        if (listener === onUpdate) listener = null
+    }
+
+    fun publish(status: String?) {
+        listener?.invoke(status)
+    }
+}

android/app/src/main/java/dev/androidagent/localmodel/LocalModelSize.kt

@@ -0,0 +1,14 @@
+package dev.androidagent.localmodel
+
+// Parses param count from the display name; file size isn't reliable across quant schemes.
+internal object LocalModelSize {
+    private val PARAM_COUNT = Regex("""(\d+(?:\.\d+)?)\s*[Bb](?:\b|-)""")
+
+    // Below this, the tool-call JSON protocol degenerates into non-answers.
+    private const val TINY_THRESHOLD_BILLIONS = 2.0
+
+    fun isTiny(displayName: String): Boolean {
+        val billions = PARAM_COUNT.find(displayName)?.groupValues?.get(1)?.toDoubleOrNull() ?: return false
+        return billions < TINY_THRESHOLD_BILLIONS
+    }
+}

android/app/src/main/java/dev/androidagent/localmodel/LocalModelStore.kt

@@ -10,6 +10,7 @@ import dev.androidagent.storage.StoredBlob
 import kotlinx.coroutines.Dispatchers
 import kotlinx.coroutines.isActive
 import kotlinx.coroutines.withContext
+import org.json.JSONObject
 import java.io.File
 import java.io.IOException
 import java.io.InputStream
@@ -25,7 +26,8 @@ object LocalModelStore {
     )
     private val GGUF_LIMITS = modelLimits(
         maxItemBytes = 12L * 1024L * 1024L * 1024L,
-        maxAggregateBytes = 24L * 1024L * 1024L * 1024L
+        maxAggregateBytes = 24L * 1024L * 1024L * 1024L,
+        maxBlobCount = 8
     )
 
     suspend fun importModel(
@@ -102,10 +104,53 @@ object LocalModelStore {
         formatForDisplayName(path) != null && File(path.trim()).isFile
 
     fun displayName(path: String): String {
-        val name = File(path.trim()).name
-        return name.removeSuffix(".litertlm").removeSuffix(".gguf").ifBlank { "Local model" }
+        val trimmed = path.trim()
+        if (trimmed.isBlank()) return "No model selected"
+        val file = File(trimmed)
+        val fallback = file.name.removeSuffix(".litertlm").removeSuffix(".gguf").ifBlank { "Local model" }
+        val metadataFile = File(file.parentFile, "${file.nameWithoutExtension}.json")
+        val storedDisplayName = runCatching {
+            if (metadataFile.isFile) {
+                JSONObject(metadataFile.readText()).optString("displayName").takeIf { it.isNotBlank() }
+            } else {
+                null
+            }
+        }.getOrNull()
+        return storedDisplayName ?: fallback
     }
 
+    // Newest first.
+    internal fun listImportedModels(context: Context): List<ImportedModel> {
+        return ModelFormat.values().flatMap { format ->
+            val directory = File(context.filesDir, format.directoryName)
+            directory.listFiles().orEmpty()
+                .filter { it.name.endsWith(".json") }
+                .mapNotNull { metadataFile ->
+                    runCatching {
+                        val json = JSONObject(metadataFile.readText())
+                        val id = json.getString("id")
+                        val payloadFile = File(directory, "$id${format.extension}")
+                        if (!payloadFile.isFile) return@runCatching null
+                        ImportedModel(
+                            path = payloadFile.absolutePath,
+                            displayName = json.optString("displayName").takeIf { it.isNotBlank() } ?: id,
+                            format = format,
+                            sizeBytes = json.optLong("sizeBytes", payloadFile.length()),
+                            createdAt = json.optLong("createdAt", payloadFile.lastModified())
+                        )
+                    }.getOrNull()
+                }
+        }.sortedByDescending { it.createdAt }
+    }
+
+    internal data class ImportedModel(
+        val path: String,
+        val displayName: String,
+        val format: ModelFormat,
+        val sizeBytes: Long,
+        val createdAt: Long
+    )
+
     internal enum class ModelFormat(
         val extension: String,
         val directoryName: String,
@@ -121,10 +166,10 @@ object LocalModelStore {
         val mimeType: String?
     )
 
-    private fun modelLimits(maxItemBytes: Long, maxAggregateBytes: Long) = BlobStoreLimits(
+    private fun modelLimits(maxItemBytes: Long, maxAggregateBytes: Long, maxBlobCount: Int = 3) = BlobStoreLimits(
         minItemBytes = MIN_MODEL_BYTES,
         maxItemBytes = maxItemBytes,
-        maxBlobCount = 3,
+        maxBlobCount = maxBlobCount,
         maxAggregateBytes = maxAggregateBytes,
         freeSpaceReserveBytes = 512L * 1024L * 1024L,
         retentionMillis = Long.MAX_VALUE

android/app/src/main/java/dev/androidagent/localmodel/LocalPromptBuilder.kt

@@ -4,8 +4,13 @@ internal object LocalPromptBuilder {
     fun systemPrompt(
         basePrompt: String,
         toolsAllowed: Boolean,
-        toolDescriptionsJson: String
+        toolDescriptionsJson: String,
+        isTinyModel: Boolean = false
     ): String {
+        // Sub-2B models can't reliably use the tool-call protocol below; skip to a plain prompt.
+        if (isTinyModel) {
+            return "Answer the user's question directly and completely, in plain conversational language."
+        }
         val toolPolicy = if (toolsAllowed) {
             """
             Tool mode:

android/app/src/main/java/dev/androidagent/settings/screens/LocalModelSettingsScreen.kt

@@ -9,6 +9,8 @@ import android.widget.LinearLayout
 import dev.androidagent.AgentConfigStore
 import dev.androidagent.LocalModelBackend
 import dev.androidagent.R
+import dev.androidagent.localmodel.LocalModelImportStatus
+import dev.androidagent.localmodel.LocalModelStore
 import dev.androidagent.settings.ColorUtils
 import dev.androidagent.settings.SettingsComponents
 import dev.androidagent.settings.SettingsComponents.BadgeTone
@@ -75,8 +77,46 @@ object LocalModelSettingsScreen {
             exposeToAccessibility(R.id.openclaw_local_model_path_field, "Local model path")
         }
 
+        val modelPickerPaths = mutableListOf<String>()
+        fun modelPickerChoices(): List<String> {
+            val models = LocalModelStore.listImportedModels(activity)
+            modelPickerPaths.clear()
+            modelPickerPaths.addAll(models.map { it.path })
+            return if (models.isEmpty()) {
+                listOf("No models imported yet")
+            } else {
+                models.map { "${it.displayName} (${it.sizeBytes / (1024 * 1024)} MB)" }
+            }
+        }
+        val initialModelChoices = modelPickerChoices()
+        val initialModelSelection = modelPickerPaths.indexOf(config.localModelPath).coerceAtLeast(0)
+        val modelPickerSpinner = SettingsUi.styledSpinner(activity, initialModelChoices, initialModelSelection, tokens)
+        root.addView(modelPickerSpinner, SettingsComponents.verticalMargin(activity, bottom = DesignTokens.Spacing.md))
+        SettingsUi.onSpinnerSelectionChanged(modelPickerSpinner) { index ->
+            modelPickerPaths.getOrNull(index)?.let { path -> pathInput.setText(path) }
+        }
+
         // Import model card
-        root.addView(buildImportCard(activity, tokens, callbacks, pathInput, config.localModelPath))
+        val importProgressLabel = SettingsComponents.body(activity, tokens, "").apply {
+            visibility = View.GONE
+        }
+        root.addView(buildImportCard(activity, tokens, callbacks, pathInput, config.localModelPath, importProgressLabel))
+        LocalModelImportStatus.observe { status ->
+            if (status == null) {
+                importProgressLabel.visibility = View.GONE
+                val choices = modelPickerChoices()
+                (modelPickerSpinner.adapter as? android.widget.ArrayAdapter<String>)?.let { adapter ->
+                    adapter.clear()
+                    adapter.addAll(choices)
+                    adapter.notifyDataSetChanged()
+                }
+                val selection = modelPickerPaths.indexOf(pathInput.text.toString()).coerceAtLeast(0)
+                modelPickerSpinner.setSelection(selection)
+            } else {
+                importProgressLabel.visibility = View.VISIBLE
+                importProgressLabel.text = status
+            }
+        }
 
         // Backend card
         val backends = LocalModelBackend.values().toList()
@@ -158,7 +198,8 @@ object LocalModelSettingsScreen {
         tokens: ThemeTokens,
         callbacks: Callbacks,
         pathInput: EditText,
-        currentPath: String
+        currentPath: String,
+        progressLabel: android.widget.TextView
     ): LinearLayout {
         val card = SettingsComponents.card(activity, tokens, padding = DesignTokens.Spacing.md)
         val row = LinearLayout(activity).apply {
@@ -191,6 +232,7 @@ object LocalModelSettingsScreen {
         }.exposeToAccessibility(R.id.openclaw_local_model_import_button, "Import local model"))
 
         card.addView(row)
+        card.addView(progressLabel, SettingsComponents.verticalMargin(activity, top = DesignTokens.Spacing.sm))
         // Hidden text field driver
         card.addView(pathInput)
         return card