Record LLM prompt size metrics

README.md

@@ -75,7 +75,7 @@ Use **Run benchmark suite** after loading models to execute the current stack's
 
 Microphone startup is capped at 15 seconds. If browser permissions, fake media devices, or audio worklet startup hang, the active mic benchmark is cancelled and the event log records the failure instead of leaving a silent pending run. Loopback audio is also followed by an explicit ASR flush after the synthesized prompt and trailing silence have been fed, so sticky VAD/STT candidates fail or finish as rows instead of waiting for the global benchmark timeout.
 
-Exported result rows include runtime metadata under `stack.environment`: page URL, secure-context status, browser user agent, CPU concurrency hint, device memory hint when exposed, WebGPU availability, WebGPU adapter info/features/software-adapter classification when exposed by the browser, microphone or loopback sample rate, and sanitized microphone track settings such as channel count, sample rate, latency, and audio-processing flags when the browser reports them. The benchmark stack also records voice, TTS steps, VAD silence, partial-ASR state, full-stack model load time as `modelLoadMs`, LLM worker acknowledgement latency as `llmStartMs`, first-response latency, and scheduled full-audio completion as `audioEndMs` / `speechEndToAudioEndMs`. Keep that metadata with WebGPU results so latency numbers can be compared across machines.
+Exported result rows include runtime metadata under `stack.environment`: page URL, secure-context status, browser user agent, CPU concurrency hint, device memory hint when exposed, WebGPU availability, WebGPU adapter info/features/software-adapter classification when exposed by the browser, microphone or loopback sample rate, and sanitized microphone track settings such as channel count, sample rate, latency, and audio-processing flags when the browser reports them. The benchmark stack also records voice, TTS steps, VAD silence, partial-ASR state, full-stack model load time as `modelLoadMs`, LLM worker acknowledgement latency as `llmStartMs`, LLM prompt build time and input size as `llmPromptBuildMs` / `llmPromptTokens`, first-response latency, and scheduled full-audio completion as `audioEndMs` / `speechEndToAudioEndMs`. Keep that metadata with WebGPU results so latency numbers can be compared across machines.
 
 The latency panel counts **speech end** from the detected acoustic end of speech, not from the later worker event that closes the turn after trailing silence. The separate **VAD close delay** metric shows how much silence was observed before the turn was finalized.
 

app.js

@@ -1220,6 +1220,14 @@ function onLlmMessage(event) {
     setTile("llm", "Generating", "active");
     return;
   }
+  if (message.type === "prompt") {
+    if (message.turnId !== state.activeTurnId) return;
+    if (state.activeBenchmark) {
+      state.activeBenchmark.llmPromptBuildMs = message.promptBuildMs ?? null;
+      state.activeBenchmark.llmPromptTokens = message.inputTokens ?? null;
+    }
+    return;
+  }
   if (message.type === "token") {
     if (message.turnId !== state.activeTurnId) return;
     if (state.awaitingFirstToken) {
@@ -1535,6 +1543,8 @@ function startBenchmark(kind, prompt, { referenceText = "", timeoutMs = 120000 }
     sttCer: null,
     vadCloseDelayMs: null,
     llmStartMs: null,
+    llmPromptBuildMs: null,
+    llmPromptTokens: null,
     firstTokenMs: null,
     firstTtsQueuedMs: null,
     firstTtsSynthesisMs: null,

workers/llm-worker.js

@@ -62,12 +62,19 @@ function reportProgress(label) {
 async function generate(messages, turnId, { sentenceLimit = 1 } = {}) {
   stoppingCriteria = new InterruptableStoppingCriteria();
   self.postMessage({ type: "start", turnId });
+  const promptBuildStartedAt = performance.now();
   const promptMessages = normalizeMessages(messages);
   const inputs = tokenizer.apply_chat_template(promptMessages, {
     add_generation_prompt: true,
     enable_thinking: false,
     return_dict: true,
   });
+  self.postMessage({
+    type: "prompt",
+    turnId,
+    inputTokens: inputTokenCount(inputs),
+    promptBuildMs: performance.now() - promptBuildStartedAt,
+  });
 
   let firstTokenAt = 0;
   let tokenCount = 0;
@@ -138,3 +145,18 @@ function normalizeMessages(messages) {
   const content = messages.map((message) => `${message.role}: ${message.content}`).join("\n");
   return [{ role: "user", content }];
 }
+
+function inputTokenCount(inputs) {
+  const inputIds = inputs?.input_ids;
+  if (!inputIds) return null;
+  if (Array.isArray(inputIds)) {
+    return Array.isArray(inputIds[0]) ? inputIds[0].length : inputIds.length;
+  }
+  if (Array.isArray(inputIds.dims) && inputIds.dims.length > 0) {
+    return inputIds.dims[inputIds.dims.length - 1];
+  }
+  if (Number.isFinite(inputIds.size)) return inputIds.size;
+  if (Number.isFinite(inputIds.length)) return inputIds.length;
+  if (Number.isFinite(inputIds.data?.length)) return inputIds.data.length;
+  return null;
+}