feat: add tool calling smoke tests and single-shot demo mode

michalharakal · claude · michalharakal · commit 34eda36bc8a0 · 2026-04-11T12:20:14.000+02:00
Add tool calling test phase to smoke-test.sh that runs --demo with a
prompt for models with toolCalling config. Add Qwen3-8B-Q4 to smoke
test config.

- Add ToolCallingDemo.runSingleShot() for non-interactive tool calling
- Wire --demo with positional prompt to single-shot mode in kllama and skainet CLIs
- Add tool calling section to smoke-test.sh with [Tool Call] detection
- Add skainet runner to smoke-test.sh runner_args
- Increase kllama-cli memory to -Xmx42g -XX:MaxDirectMemorySize=64g
- Add Qwen3-8B-Q4 and toolCalling config to smoke-models.json

Co-Authored-By: Claude Opus 4.6 (1M context) &lt;noreply@anthropic.com&gt;
diff --git a/llm-apps/kllama-cli/build.gradle.kts b/llm-apps/kllama-cli/build.gradle.kts
@@ -34,5 +34,5 @@ tasks.withType<Test>().configureEach {
 }
 
 tasks.withType<JavaExec>().configureEach {
-    jvmArgs("--enable-preview", "--add-modules", "jdk.incubator.vector", "-Xmx12g")
+    jvmArgs("--enable-preview", "--add-modules", "jdk.incubator.vector", "-Xmx42g", "-XX:MaxDirectMemorySize=64g")
 }
diff --git a/llm-apps/skainet-cli/build.gradle.kts b/llm-apps/skainet-cli/build.gradle.kts
@@ -50,5 +50,5 @@ tasks.withType<Test>().configureEach {
 }
 
 tasks.withType<JavaExec>().configureEach {
-    jvmArgs("--enable-preview", "--add-modules", "jdk.incubator.vector", "-Xmx12g", "-XX:MaxDirectMemorySize=64g")
+    jvmArgs("--enable-preview", "--add-modules", "jdk.incubator.vector", "-Xmx48g", "-XX:MaxDirectMemorySize=64g")
 }
diff --git a/llm-apps/skainet-cli/src/main/kotlin/sk/ainet/apps/skainet/cli/Main.kt b/llm-apps/skainet-cli/src/main/kotlin/sk/ainet/apps/skainet/cli/Main.kt
@@ -209,7 +209,11 @@ fun main(args: Array<String>) {
             when {
                 cliArgs.demoMode -> {
                     val demo = ToolCallingDemo(runtime, tokenizer, cliArgs.templateName, metadata)
-                    demo.run(maxTokens = cliArgs.steps, temperature = cliArgs.temperature)
+                    if (cliArgs.prompt != null) {
+                        demo.runSingleShot(cliArgs.prompt, maxTokens = cliArgs.steps, temperature = cliArgs.temperature)
+                    } else {
+                        demo.run(maxTokens = cliArgs.steps, temperature = cliArgs.temperature)
+                    }
                 }
                 cliArgs.agentMode -> {
                     val agentCli = AgentCli(runtime, tokenizer, cliArgs.templateName, metadata)
diff --git a/llm-runtime/kllama/src/jvmMain/kotlin/sk/ainet/apps/kllama/cli/Main.kt b/llm-runtime/kllama/src/jvmMain/kotlin/sk/ainet/apps/kllama/cli/Main.kt
@@ -496,7 +496,11 @@ fun main(args: Array<String>) {
             when {
                 cliArgs.demoMode -> {
                     val demo = ToolCallingDemo(runtime, tokenizer, cliArgs.templateName, metadata)
-                    demo.run(maxTokens = cliArgs.steps, temperature = cliArgs.temperature)
+                    if (cliArgs.prompt != null) {
+                        demo.runSingleShot(cliArgs.prompt, maxTokens = cliArgs.steps, temperature = cliArgs.temperature)
+                    } else {
+                        demo.run(maxTokens = cliArgs.steps, temperature = cliArgs.temperature)
+                    }
                 }
                 cliArgs.agentMode -> {
                     val agentCli = AgentCli(runtime, tokenizer, cliArgs.templateName, metadata)
diff --git a/llm-runtime/kllama/src/jvmMain/kotlin/sk/ainet/apps/kllama/cli/ToolCallingDemo.kt b/llm-runtime/kllama/src/jvmMain/kotlin/sk/ainet/apps/kllama/cli/ToolCallingDemo.kt
@@ -46,6 +46,70 @@ public class ToolCallingDemo<T : DType>(
     /**
      * Run the tool-calling demo with `list_files` and `calculator` tools.
      */
+    /**
+     * Run a single non-interactive tool calling round. Used by smoke tests.
+     *
+     * @param prompt The user prompt.
+     * @param maxTokens Maximum tokens per round.
+     * @param temperature Sampling temperature.
+     */
+    public fun runSingleShot(
+        prompt: String,
+        maxTokens: Int = 256,
+        temperature: Float = 0.7f
+    ) {
+        val registry = ToolRegistry()
+        registry.register(ListFilesTool())
+        registry.register(CalculatorTool())
+
+        println("Tool Calling Smoke Test")
+        println("Available tools: ${registry.definitions().joinToString { it.name }}")
+        println("Prompt: \"$prompt\"")
+        println("---")
+
+        val agentLoop = session.createAgentLoop(registry, maxTokens, temperature)
+
+        val systemPrompt = """You are a helpful assistant with access to tools.
+When the user asks about files or directories, use the list_files tool to look up the actual contents.
+When the user asks to calculate something, use the calculator tool.
+Always use a tool when one is relevant — do not guess file listings."""
+
+        val messages = mutableListOf(
+            ChatMessage(role = ChatRole.SYSTEM, content = systemPrompt),
+            ChatMessage(role = ChatRole.USER, content = prompt)
+        )
+
+        val listener = object : AgentListener {
+            override fun onToken(token: String) {
+                print(token)
+                System.out.flush()
+            }
+            override fun onAssistantMessage(text: String) { println() }
+            override fun onToolCalls(calls: List<ToolCall>) {
+                for (call in calls) println("[Tool Call] ${call.name}(${call.arguments})")
+            }
+            override fun onToolResult(call: ToolCall, result: String) {
+                println("[Tool Result] ${call.name} -> $result")
+                print("Assistant: ")
+                System.out.flush()
+            }
+            override fun onComplete(finalResponse: String) {}
+        }
+
+        print("Assistant: ")
+        System.out.flush()
+
+        agentLoop.runWithEncoder(
+            messages = messages,
+            encode = { text -> tokenizer.encode(text) },
+            listener = listener
+        )
+        println()
+    }
+
+    /**
+     * Run the interactive tool-calling demo.
+     */
     public fun run(
         maxTokens: Int = 512,
         temperature: Float = 0.7f
diff --git a/tests/smoke/smoke-models.json b/tests/smoke/smoke-models.json
@@ -6,31 +6,30 @@
   },
   "models": [
     {
-      "name": "Llama-3.2-1B-Q8",
+      "name": "TinyLlama-1.1B-Q8",
       "runner": "kllama",
-      "model": "TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF/tinyllama-1.1b-chat-v1.0.Q8_0.gguf",
+      "model": "tinyllama-1.1b-chat-v1.0.Q8_0.gguf",
       "format": "gguf"
     },
     {
-      "name": "Gemma-2B-SafeTensors",
-      "runner": "kgemma",
-      "model": "unsloth/gemma-3-270m-it-GGUF/gemma-3-270m-it-Q8_0.gguf",
+      "name": "Qwen3-1.7B-Q8",
+      "runner": "kqwen",
+      "model": "Qwen3-1.7B-Q8_0.gguf",
       "format": "gguf",
-      "steps": 16
+      "toolCalling": {
+        "prompt": "What is 2 + 2?",
+        "steps": 256
+      }
     },
     {
-      "name": "Qwen3-1.7B",
-      "runner": "qwen",
-      "model": "Qwen3-1.7B-GGUF/Qwen3-1.7B-Q8_0.gguf",
-      "format": "safetensors",
-      "prompt": "Hello world",
-    },
-    {
-      "name": "BERT-MiniLM",
-      "runner": "kbert",
-      "model": "~/.cache/huggingface/models/all-MiniLM-L6-v2",
-      "format": "safetensors",
-      "prompt": "Hello world"
+      "name": "Qwen3-8B-Q4",
+      "runner": "kllama",
+      "model": "Qwen3-8B-Q4_K_M.gguf",
+      "format": "gguf",
+      "toolCalling": {
+        "prompt": "What is 2 + 2?",
+        "steps": 256
+      }
     }
   ]
 }
diff --git a/tests/smoke/smoke-test.sh b/tests/smoke/smoke-test.sh
@@ -69,6 +69,7 @@ runner_args() {
   local runner="$1" model="$2" prompt="$3" steps="$4" temp="$5" doc="${6:-}" output="${7:-}"
 
   case "$runner" in
+    skainet)  echo "-m ${model} -s ${steps} -k ${temp} \"${prompt}\"" ;;
     kllama)   echo "-m ${model} -s ${steps} -k ${temp} \"${prompt}\"" ;;
     kgemma)   echo "${model} \"${prompt}\" ${steps} ${temp}" ;;
     kqwen)    echo "${model} \"${prompt}\" ${steps} ${temp}" ;;
@@ -254,9 +255,93 @@ print(f'M_OUTPUT={repr(m.get(\"output\", \"\"))}')
     separator
   done
 
+  # ── Tool Calling Tests ───────────────────────────────────────────
+  TC_COUNT=$(python3 -c "
+import json
+cfg = json.load(open('${CONFIG_FILE}'))
+print(sum(1 for m in cfg['models'] if m.get('toolCalling')))
+")
+
+  declare -a tc_results=()
+  tc_pass=0
+  tc_fail=0
+
+  if [[ "$TC_COUNT" -gt 0 ]]; then
+    echo ""
+    echo -e "${BOLD}Tool Calling Tests${RESET} ($TC_COUNT models)"
+    separator
+
+    kllama_task=$(runner_task "kllama")
+
+    for i in $(seq 0 $((MODEL_COUNT - 1))); do
+      eval "$(python3 -c "
+import json
+cfg = json.load(open('${CONFIG_FILE}'))
+m = cfg['models'][$i]
+tc = m.get('toolCalling')
+if tc is None:
+    print('TC_ENABLED=false')
+else:
+    print('TC_ENABLED=true')
+    print(f'TC_PROMPT={repr(tc.get(\"prompt\", \"What is 2 + 2?\"))}')
+    print(f'TC_STEPS={tc.get(\"steps\", 256)}')
+    print(f'M_NAME={repr(m[\"name\"])}')
+    print(f'M_MODEL={repr(m[\"model\"])}')
+")"
+
+      [[ "$TC_ENABLED" != "true" ]] && continue
+
+      M_MODEL=$(expand_path "$M_MODEL")
+
+      echo -e "\n${BOLD}Model:${RESET}  $M_NAME (tool calling)"
+      echo -e "${BOLD}Prompt:${RESET} \"$TC_PROMPT\""
+
+      if [[ ! -e "$M_MODEL" ]]; then
+        echo -e "  ${RED}FAIL${RESET} (model path not found)"
+        tc_fail=$((tc_fail + 1))
+        tc_results+=("FAIL|$M_NAME|not found|-")
+        separator
+        continue
+      fi
+
+      start_ts=$(python3 -c 'import time; print(time.time())')
+      output_file=$(mktemp)
+      exit_code=0
+
+      $GRADLE "$kllama_task" --quiet \
+        --args="-m ${M_MODEL} --demo -s ${TC_STEPS} -k 0.7 \"${TC_PROMPT}\"" \
+        > "$output_file" 2>&1 || exit_code=$?
+
+      end_ts=$(python3 -c 'import time; print(time.time())')
+      wall_sec=$(python3 -c "print(f'{$end_ts - $start_ts:.1f}')")
+
+      if [[ $exit_code -ne 0 ]]; then
+        echo -e "  ${RED}FAIL${RESET} (exit $exit_code, wall ${wall_sec}s)"
+        tail -5 "$output_file" | sed 's/^/  │ /'
+        tc_fail=$((tc_fail + 1))
+        tc_results+=("FAIL|$M_NAME|exit $exit_code|${wall_sec}s")
+      elif grep -q '\[Tool Call\]' "$output_file"; then
+        tool_name=$(grep -oE '\[Tool Call\] [a-z_]+' "$output_file" | head -1 | sed 's/\[Tool Call\] //')
+        echo -e "  ${GREEN}OK${RESET}   tool called: ${CYAN}${tool_name}${RESET}  wall: ${wall_sec}s"
+        grep '\[Tool Call\]' "$output_file" | head -2 | sed 's/^/  │ /'
+        grep '\[Tool Result\]' "$output_file" | head -2 | sed 's/^/  │ /'
+        tc_pass=$((tc_pass + 1))
+        tc_results+=("OK|$M_NAME|$tool_name|${wall_sec}s")
+      else
+        echo -e "  ${YELLOW}WARN${RESET} (no tool call detected, wall ${wall_sec}s)"
+        tail -5 "$output_file" | sed 's/^/  │ /'
+        tc_fail=$((tc_fail + 1))
+        tc_results+=("WARN|$M_NAME|no tool call|${wall_sec}s")
+      fi
+
+      rm -f "$output_file"
+      separator
+    done
+  fi
+
   # ── Summary ──────────────────────────────────────────────────────
   echo ""
-  echo -e "${BOLD}Summary${RESET}"
+  echo -e "${BOLD}Summary — Generation${RESET}"
   separator
   printf "  %-6s %-30s %-8s %8s %10s %8s\n" "Status" "Model" "Runner" "Size" "tok/s" "Wall"
   separator
@@ -272,6 +357,29 @@ print(f'M_OUTPUT={repr(m.get(\"output\", \"\"))}')
   done
   separator
   echo -e "  ${GREEN}Pass: $pass${RESET}  ${RED}Fail: $fail${RESET}  Total: ${MODEL_COUNT}"
+
+  if [[ "$TC_COUNT" -gt 0 ]]; then
+    echo ""
+    echo -e "${BOLD}Summary — Tool Calling${RESET}"
+    separator
+    printf "  %-6s %-30s %-15s %8s\n" "Status" "Model" "Tool" "Wall"
+    separator
+    for r in "${tc_results[@]}"; do
+      IFS='|' read -r status name tool wall <<< "$r"
+      if [[ "$status" == "OK" ]]; then
+        color="$GREEN"
+      elif [[ "$status" == "WARN" ]]; then
+        color="$YELLOW"
+      else
+        color="$RED"
+      fi
+      printf "  ${color}%-6s${RESET} %-30s %-15s %8s\n" \
+        "$status" "${name:0:30}" "$tool" "$wall"
+    done
+    separator
+    echo -e "  ${GREEN}Pass: $tc_pass${RESET}  ${RED}Fail: $tc_fail${RESET}  Total: ${TC_COUNT}"
+  fi
+
   echo ""
   exit 0
 fi

Original file line number	Diff line number	Diff line change
`@@ -34,5 +34,5 @@ tasks.withType<Test>().configureEach {`
`34`	`34`	`}`
`35`	`35`
`36`	`36`	`tasks.withType<JavaExec>().configureEach {`
`37`		`- jvmArgs("--enable-preview", "--add-modules", "jdk.incubator.vector", "-Xmx12g")`
	`37`	`+ jvmArgs("--enable-preview", "--add-modules", "jdk.incubator.vector", "-Xmx42g", "-XX:MaxDirectMemorySize=64g")`
`38`	`38`	`}`
Original file line number	Diff line number	Diff line change
`@@ -50,5 +50,5 @@ tasks.withType<Test>().configureEach {`
`50`	`50`	`}`
`51`	`51`
`52`	`52`	`tasks.withType<JavaExec>().configureEach {`
`53`		`- jvmArgs("--enable-preview", "--add-modules", "jdk.incubator.vector", "-Xmx12g", "-XX:MaxDirectMemorySize=64g")`
	`53`	`+ jvmArgs("--enable-preview", "--add-modules", "jdk.incubator.vector", "-Xmx48g", "-XX:MaxDirectMemorySize=64g")`
`54`	`54`	`}`