|
| 1 | +/* |
| 2 | + * Copyright (c) Meta Platforms, Inc. and affiliates. |
| 3 | + * All rights reserved. |
| 4 | + * |
| 5 | + * This source code is licensed under the BSD-style license found in the |
| 6 | + * LICENSE file in the root directory of this source tree. |
| 7 | + */ |
| 8 | + |
| 9 | +package com.example.executorchllamademo |
| 10 | + |
| 11 | +import android.content.Context |
| 12 | +import android.util.Log |
| 13 | +import androidx.compose.ui.semantics.SemanticsProperties |
| 14 | +import androidx.compose.ui.test.assertIsEnabled |
| 15 | +import androidx.compose.ui.test.hasContentDescription |
| 16 | +import androidx.compose.ui.test.junit4.createAndroidComposeRule |
| 17 | +import androidx.compose.ui.test.onAllNodesWithText |
| 18 | +import androidx.compose.ui.test.onNodeWithContentDescription |
| 19 | +import androidx.compose.ui.test.onNodeWithTag |
| 20 | +import androidx.compose.ui.test.onNodeWithText |
| 21 | +import androidx.compose.ui.test.performClick |
| 22 | +import androidx.compose.ui.test.performTextInput |
| 23 | +import androidx.test.core.app.ApplicationProvider |
| 24 | +import androidx.test.ext.junit.runners.AndroidJUnit4 |
| 25 | +import androidx.test.filters.LargeTest |
| 26 | +import org.junit.Assert.assertTrue |
| 27 | +import org.junit.Before |
| 28 | +import org.junit.Ignore |
| 29 | +import org.junit.Rule |
| 30 | +import org.junit.Test |
| 31 | +import org.junit.runner.RunWith |
| 32 | + |
| 33 | +/** |
| 34 | + * Preset model sanity test that validates the preset model download and chat workflow. |
| 35 | + * |
| 36 | + * This test validates: |
| 37 | + * 1. Navigate from Welcome screen to Preset model screen |
| 38 | + * 2. Select Stories 110M and download it |
| 39 | + * 3. After download completes, tap to load and enter chat view |
| 40 | + * 4. Type "Once upon a time" and generate a response |
| 41 | + */ |
| 42 | +@RunWith(AndroidJUnit4::class) |
| 43 | +@LargeTest |
| 44 | +class PresetSanityTest { |
| 45 | + |
| 46 | + companion object { |
| 47 | + private const val TAG = "PresetSanityTest" |
| 48 | + private const val RESPONSE_TAG = "LLAMA_RESPONSE" |
| 49 | + } |
| 50 | + |
| 51 | + @get:Rule |
| 52 | + val composeTestRule = createAndroidComposeRule<WelcomeActivity>() |
| 53 | + |
| 54 | + @Before |
| 55 | + fun setUp() { |
| 56 | + // Clear SharedPreferences before test to ensure a clean state |
| 57 | + val context = ApplicationProvider.getApplicationContext<Context>() |
| 58 | + val prefs = context.getSharedPreferences( |
| 59 | + context.getString(R.string.demo_pref_file_key), |
| 60 | + Context.MODE_PRIVATE |
| 61 | + ) |
| 62 | + prefs.edit().clear().commit() |
| 63 | + } |
| 64 | + |
| 65 | + /** |
| 66 | + * Types text into the chat input field using testTag. |
| 67 | + */ |
| 68 | + private fun typeInChatInput(text: String) { |
| 69 | + composeTestRule.onNodeWithTag("chat_input_field").performClick() |
| 70 | + composeTestRule.waitForIdle() |
| 71 | + composeTestRule.onNodeWithTag("chat_input_field").performTextInput(text) |
| 72 | + composeTestRule.waitForIdle() |
| 73 | + } |
| 74 | + |
| 75 | + /** |
| 76 | + * Waits for generation to complete by checking for tokens-per-second metrics. |
| 77 | + */ |
| 78 | + private fun waitForGenerationComplete(timeoutMs: Long = 120000): Boolean { |
| 79 | + return try { |
| 80 | + composeTestRule.waitUntil(timeoutMillis = timeoutMs) { |
| 81 | + val tpsNodes = composeTestRule.onAllNodesWithText("t/s", substring = true) |
| 82 | + .fetchSemanticsNodes() |
| 83 | + val tokpsNodes = composeTestRule.onAllNodesWithText("tok/s", substring = true) |
| 84 | + .fetchSemanticsNodes() |
| 85 | + tpsNodes.isNotEmpty() || tokpsNodes.isNotEmpty() |
| 86 | + } |
| 87 | + Log.i(TAG, "Generation complete - found generation metrics") |
| 88 | + true |
| 89 | + } catch (e: Exception) { |
| 90 | + Log.e(TAG, "Generation timed out after ${timeoutMs}ms") |
| 91 | + false |
| 92 | + } |
| 93 | + } |
| 94 | + |
| 95 | + /** |
| 96 | + * Waits for the model to be loaded by checking for success or error messages. |
| 97 | + */ |
| 98 | + private fun waitForModelLoaded(timeoutMs: Long = 60000): Boolean { |
| 99 | + return try { |
| 100 | + var wasSuccess = false |
| 101 | + composeTestRule.waitUntil(timeoutMillis = timeoutMs) { |
| 102 | + val successNodes = composeTestRule.onAllNodesWithText("Successfully loaded", substring = true) |
| 103 | + .fetchSemanticsNodes() |
| 104 | + val errorNodes = composeTestRule.onAllNodesWithText("Model load failure", substring = true) |
| 105 | + .fetchSemanticsNodes() |
| 106 | + wasSuccess = successNodes.isNotEmpty() |
| 107 | + successNodes.isNotEmpty() || errorNodes.isNotEmpty() |
| 108 | + } |
| 109 | + if (wasSuccess) { |
| 110 | + Log.i(TAG, "Model loaded successfully") |
| 111 | + } else { |
| 112 | + Log.e(TAG, "Model load failed") |
| 113 | + } |
| 114 | + wasSuccess |
| 115 | + } catch (e: Exception) { |
| 116 | + Log.e(TAG, "Model loading timed out after ${timeoutMs}ms: ${e.message}") |
| 117 | + false |
| 118 | + } |
| 119 | + } |
| 120 | + |
| 121 | + /** |
| 122 | + * Verifies that the model generated a non-empty response. |
| 123 | + */ |
| 124 | + private fun assertModelResponseNotEmpty(timeoutMs: Long = 10000) { |
| 125 | + try { |
| 126 | + composeTestRule.waitUntil(timeoutMillis = timeoutMs) { |
| 127 | + val tpsNodes = composeTestRule.onAllNodesWithText("t/s", substring = true) |
| 128 | + .fetchSemanticsNodes() |
| 129 | + val tokpsNodes = composeTestRule.onAllNodesWithText("tok/s", substring = true) |
| 130 | + .fetchSemanticsNodes() |
| 131 | + tpsNodes.isNotEmpty() || tokpsNodes.isNotEmpty() |
| 132 | + } |
| 133 | + Log.i(TAG, "Model response verified - found generation metrics") |
| 134 | + } catch (e: Exception) { |
| 135 | + throw AssertionError("Model response appears to be empty - no generation metrics found after ${timeoutMs}ms") |
| 136 | + } |
| 137 | + } |
| 138 | + |
| 139 | + /** |
| 140 | + * Logs the model response text for CI output. |
| 141 | + */ |
| 142 | + private fun logModelResponse() { |
| 143 | + try { |
| 144 | + Log.i(RESPONSE_TAG, "BEGIN_RESPONSE") |
| 145 | + val responseNodes = composeTestRule.onAllNodesWithText("t/s", substring = true) |
| 146 | + .fetchSemanticsNodes() |
| 147 | + for (node in responseNodes) { |
| 148 | + val text = node.config.getOrElse(SemanticsProperties.Text) { emptyList() } |
| 149 | + .joinToString(" ") { it.text } |
| 150 | + if (text.isNotBlank()) { |
| 151 | + Log.i(RESPONSE_TAG, text) |
| 152 | + } |
| 153 | + } |
| 154 | + Log.i(RESPONSE_TAG, "END_RESPONSE") |
| 155 | + } catch (e: Exception) { |
| 156 | + Log.d(TAG, "Could not log model response: ${e.message}") |
| 157 | + } |
| 158 | + } |
| 159 | + |
| 160 | + /** |
| 161 | + * Tests the complete preset model download and chat workflow: |
| 162 | + * 1. From Welcome screen, tap "Preset model" card |
| 163 | + * 2. Find Stories 110M and tap Download |
| 164 | + * 3. Wait for download to complete |
| 165 | + * 4. Tap the card to load model and enter chat |
| 166 | + * 5. Type "Once upon a time" and send |
| 167 | + * 6. Verify response is generated |
| 168 | + */ |
| 169 | + @Ignore("Temporarily disabled") |
| 170 | + @Test |
| 171 | + fun testPresetModelDownloadAndChat() { |
| 172 | + composeTestRule.waitForIdle() |
| 173 | + |
| 174 | + // Step 1: From Welcome screen, tap "Preset model" card |
| 175 | + Log.i(TAG, "Step 1: Navigating to Preset model screen") |
| 176 | + composeTestRule.onNodeWithText("Preset model").performClick() |
| 177 | + composeTestRule.waitUntil(timeoutMillis = 5000) { |
| 178 | + composeTestRule.onAllNodesWithText("Download Preset Model").fetchSemanticsNodes().isNotEmpty() |
| 179 | + } |
| 180 | + |
| 181 | + // Step 2: Find Stories 110M and tap Download |
| 182 | + Log.i(TAG, "Step 2: Finding Stories 110M and starting download") |
| 183 | + composeTestRule.onNodeWithText("Stories 110M").assertExists() |
| 184 | + |
| 185 | + // Check if already downloaded (Ready to use) or needs download |
| 186 | + val readyNodes = composeTestRule.onAllNodesWithText("Ready to use", substring = true) |
| 187 | + .fetchSemanticsNodes() |
| 188 | + |
| 189 | + if (readyNodes.isEmpty()) { |
| 190 | + // Need to download - click Download button |
| 191 | + composeTestRule.onNodeWithText("Download").performClick() |
| 192 | + |
| 193 | + // Step 3: Wait for download to complete (may take a while for large files) |
| 194 | + Log.i(TAG, "Step 3: Waiting for download to complete") |
| 195 | + composeTestRule.waitUntil(timeoutMillis = 300000) { // 5 minutes timeout for download |
| 196 | + composeTestRule.onAllNodesWithText("Ready to use", substring = true) |
| 197 | + .fetchSemanticsNodes().isNotEmpty() |
| 198 | + } |
| 199 | + Log.i(TAG, "Download completed") |
| 200 | + } else { |
| 201 | + Log.i(TAG, "Model already downloaded, skipping download step") |
| 202 | + } |
| 203 | + |
| 204 | + // Step 4: Tap the card to load model and enter chat |
| 205 | + Log.i(TAG, "Step 4: Tapping card to load model") |
| 206 | + composeTestRule.onNodeWithText("Stories 110M").performClick() |
| 207 | + |
| 208 | + // Wait for Activity transition - MainActivity needs time to launch and set content |
| 209 | + // The SelectPresetModelActivity calls finish() after starting MainActivity |
| 210 | + Thread.sleep(2000) |
| 211 | + |
| 212 | + // Wait for model to load and chat screen to appear |
| 213 | + Log.i(TAG, "Waiting for model to load") |
| 214 | + val modelLoaded = waitForModelLoaded(90000) |
| 215 | + assertTrue("Model should be loaded successfully", modelLoaded) |
| 216 | + Log.i(TAG, "Model loaded successfully") |
| 217 | + |
| 218 | + // Step 5: Type "Once upon a time" and send |
| 219 | + Log.i(TAG, "Step 5: Typing prompt and sending") |
| 220 | + typeInChatInput("Once upon a time") |
| 221 | + |
| 222 | + // Wait for send button to be enabled |
| 223 | + composeTestRule.waitUntil(timeoutMillis = 5000) { |
| 224 | + try { |
| 225 | + composeTestRule.onNodeWithContentDescription("Send").assertIsEnabled() |
| 226 | + true |
| 227 | + } catch (e: AssertionError) { |
| 228 | + false |
| 229 | + } |
| 230 | + } |
| 231 | + |
| 232 | + composeTestRule.onNodeWithContentDescription("Send").performClick() |
| 233 | + composeTestRule.waitForIdle() |
| 234 | + |
| 235 | + // Step 6: Wait for generation to complete and verify response |
| 236 | + Log.i(TAG, "Step 6: Waiting for generation to complete") |
| 237 | + val generationComplete = waitForGenerationComplete(120000) |
| 238 | + assertTrue("Generation should complete", generationComplete) |
| 239 | + |
| 240 | + assertModelResponseNotEmpty() |
| 241 | + logModelResponse() |
| 242 | + |
| 243 | + Log.i(TAG, "Preset model sanity test completed successfully") |
| 244 | + } |
| 245 | +} |
0 commit comments