add tool-calling with qwen wip

Signed-off-by: Varun Khare <varun.khare@nimbledgehq.ai>

Author: Varun Khare

android/app/build.gradle.kts

@@ -100,8 +100,8 @@ dependencies {
     debugImplementation(libs.androidx.ui.tooling)
     debugImplementation(libs.androidx.ui.test.manifest)
 
-    implementation("dev.deliteai:nimblenet_ktx:0.0.1-dev-1751902318")
-    implementation("dev.deliteai:nimblenet_core:0.0.1-dev-1751904491")
+    implementation("dev.deliteai:nimblenet_ktx:0.0.1-dev-1751904492")
+    implementation("dev.deliteai:nimblenet_core:0.0.1-dev-1751904492")
 
     implementation("com.halilibo.compose-richtext:richtext-ui-material3:1.0.0+")
     implementation("com.halilibo.compose-richtext:richtext-commonmark:1.0.0+")

android/app/src/main/assets/config.json

@@ -0,0 +1,40 @@
+{
+  "architectures": [
+    "Qwen3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "eos_token_id": 151645,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "intermediate_size": 6144,
+  "max_position_embeddings": 40960,
+  "max_window_layers": 28,
+  "model_type": "qwen3",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 8,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": null,
+  "rope_theta": 1000000,
+  "sliding_window": null,
+  "tie_word_embeddings": true,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.52.0.dev0",
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 151936,
+  "transformers.js_config": {
+    "kv_cache_dtype": {
+      "q4f16": "float16",
+      "fp16": "float16"
+    },
+    "use_external_data_format": {
+      "model.onnx": true,
+      "model_fp16.onnx": true
+    }
+  }
+}

android/app/src/main/assets/tokenizer.json

@@ -9,25 +9,21 @@ package dev.deliteai.assistant.domain.features.llm
 import dev.deliteai.NimbleNet
 import dev.deliteai.datamodels.NimbleNetTensor
 import dev.deliteai.impl.common.DATATYPE
+import org.json.JSONObject
 
 object LLMManager {
     suspend fun feedInput(input: String, isVoiceInitiated: Boolean) {
         val res = NimbleNet.runMethod(
-            "prompt_llm",
+            "prompt_for_tool_calling",
             inputs = hashMapOf(
-                "query" to NimbleNetTensor(input, DATATYPE.STRING, null),
-                "is_voice_initiated" to NimbleNetTensor(
-                    if (isVoiceInitiated) 1 else 0,
-                    DATATYPE.INT32,
-                    null
-                )
+                "prompt" to NimbleNetTensor(input, DATATYPE.STRING, null),
             ),
         )
-        check(res.status) { "NimbleNet.runMethod('prompt_llm') failed with status: ${res.status}" }
+        check(res.status) { "NimbleNet.runMethod('prompt_for_tool_calling') failed with status: ${res.status}" }
     }
 
     suspend fun getNextMap(): Map<String, NimbleNetTensor> {
-        val res2 = NimbleNet.runMethod("get_next_str", hashMapOf())
+        val res2 = NimbleNet.runMethod("get_token_stream", hashMapOf())
         check(res2.status) { "NimbleNet.runMethod('get_next_str') failed with error: ${res2.error?.message}" }
         return res2.payload
             ?: throw IllegalStateException("NimbleNet.runMethod('get_next_str') returned null payload")

android/app/src/main/kotlin/dev/deliteai/assistant/domain/features/llm/LLMManager.kt

@@ -52,12 +52,13 @@ class ChatRepository {
     suspend fun getModelName() = LLMService.getLLMName()
 
     fun getLLMText(textInput: String) = flow {
-        LLMService.feedInput(textInput, false)
+        repositoryScope.launch(Dispatchers.Default) {  LLMService.feedInput(textInput, false) }
         do {
             val outputMap = LLMService.getNextMap()
-            val currentOutputString = (outputMap["str"] as NimbleNetTensor).data.toString()
+            val currentOutputString = (outputMap["token_stream"] as NimbleNetTensor).data.toString()
+            Log.d(TAG, "token stream $currentOutputString")
             emit(GenerateResponseJobStatus.NextItem(currentOutputString))
-        } while (!outputMap.containsKey("finished"))
+        } while (!outputMap.containsKey("<|im_end|>"))
 
         emit(GenerateResponseJobStatus.Finished())
         Log.d(TAG, "startFeedbackLoop: LLM finished output")

android/app/src/main/kotlin/dev/deliteai/assistant/domain/repositories/ChatRepository.kt

@@ -42,24 +42,10 @@ fun ScrollableTextSuggestions(
 ) {
     val allSuggestions = remember {
         listOf(
-            "Design workout routine",
-            "Recommend wine pairings",
-            "Write a short poem",
-            "Draft party menu",
-            "Create smoothie blends",
-            "Generate gift ideas",
-            "Craft cocktail ideas",
-            "Mix mocktail recipes",
-            "Suggest hiking essentials",
-            "Plan a game night",
-            "Prep for camping",
-            "Plan a movie marathon",
-            "Invent signature cocktail",
-            "Craft lunchbox ideas",
-            "Who are you?",
-            "Plan a solo trip",
-            "Curate weekend playlist",
-            "Plan a beach day"
+            "How is the weather here?",
+            "Where am I located currently?",
+            "Multiply 23 and 35",
+            "What is time right now?"
         )
     }
 

android/app/src/main/kotlin/dev/deliteai/assistant/presentation/components/ScrollableTextSuggestions.kt

@@ -77,8 +77,9 @@ suspend fun initializeNimbleNetAndWaitForIsReady(application: Application, ct: S
         host = BuildConfig.NIMBLENET_CONFIG_HOST,
         deviceId = getInternalDeviceId(application),
         clientSecret = BuildConfig.NIMBLENET_CONFIG_CLIENT_SECRET,
-        debug = false,
-        compatibilityTag = ct,
+        debug = true,
+        online = true,
+        compatibilityTag = "QWEN_ONNX",
         libraryVariant = NIMBLENET_VARIANTS.STATIC,
         showDownloadProgress = true
     )
@@ -93,17 +94,21 @@ suspend fun initializeNimbleNetAndWaitForIsReady(application: Application, ct: S
 
 //used only during the app init
 private suspend fun passLexiconToTheWorkflowScript(context: Context) {
-    val resString = context.assets.open("lexicon.json").bufferedReader().use { it.readText() }
-    val lexiconArray = JSONObject(resString)
+    val generationConfig = JSONObject(context.assets.open("config.json").bufferedReader().use { it.readText() })
+    val tokenizerConfig = JSONObject(context.assets.open("tokenizer.json").bufferedReader().use { it.readText() })
+
     val res = NimbleNet.runMethod(
-        "init",
+        "init_generation_mixin",
         inputs = hashMapOf(
-            "lexicon" to NimbleNetTensor(
-                shape = null, data = lexiconArray, datatype = DATATYPE.JSON
+            "tokenizer_config" to NimbleNetTensor(
+                shape = null, data = tokenizerConfig, datatype = DATATYPE.JSON
+            ),
+            "generation_config" to NimbleNetTensor(
+                shape = null, data = generationConfig, datatype = DATATYPE.JSON
             )
         )
     )
-    check(res.status) { "NimbleNet.runMethod('init') failed with status: ${res.status}" }
+    check(res.status) { "NimbleNet.runMethod('init_generation_mixin') failed with status: ${res.status}" }
 }
 
 fun MutableList<String>.mergeChunks(): MutableList<String> {

android/app/src/main/kotlin/dev/deliteai/assistant/utils/Utils.kt

@@ -25,6 +25,7 @@ dependencyResolutionManagement {
         maven {
             url = uri("https://deliteai.s3.ap-south-1.amazonaws.com/releases-dev/android/")
         }
+        mavenLocal()
     }
 }