chore(bench): chat prompt-eval benchmark (bench-chat.mjs)

Isolates prompt-eval cost: tiny output, growing input, measures time-to-first-token + a cold-vs-warm
cache test. Used to root-cause the Windows chat slowness — proved prompt-eval is ~1.3s/1k tokens and
the KV cache gives 15x only on a stable prefix (which led to finding capture was hammering the single
engine on Windows). Companion to scripts/bench-capture.mjs.

Author: alichherawalla

scripts/bench-chat.mjs

@@ -0,0 +1,130 @@
+#!/usr/bin/env node
+// Does a bigger conversation history actually slow each turn? Isolate PROMPT-eval cost: hold the
+// OUTPUT tiny (max_tokens=16) and grow the INPUT (a filler "history") across sizes, measuring
+// time-to-first-token (TTFT ≈ prompt eval + 1 token) and the reported prompt_tokens. If TTFT
+// scales with prompt tokens, the "2nd response onwards is slower" report is prompt-eval growth
+// from bloated history (which the Auto max-output default enlarges), not a fixed per-turn regression.
+//
+//   node scripts/bench-chat.mjs                 # default sweep
+//   node scripts/bench-chat.mjs --port 8439 --repeats 3
+
+import { performance } from 'node:perf_hooks'
+
+const arg = (f, d) => {
+  const i = process.argv.indexOf(f)
+  return i >= 0 && process.argv[i + 1] ? process.argv[i + 1] : d
+}
+const PORT = Number(arg('--port', '8439'))
+const REPEATS = Number(arg('--repeats', '3'))
+const ENDPOINT = `http://127.0.0.1:${PORT}/v1/chat/completions`
+// Prompt-token targets standing in for a growing chat history. ~500 ≈ turn 1; 8k–32k ≈ a few
+// Auto-length answers deep. (Qwen3.5-2B trained ctx is 256K, so all fit.)
+const TARGETS = [500, 2000, 4000, 8000, 16000, 32000]
+
+// ~4 chars/token of natural-ish filler; oversize then the model reports the real prompt_tokens.
+const SENTENCE =
+  'The off-grid system balances solar capacity against battery storage while the controller logs each cycle. '
+const fillerForTokens = (tok) => SENTENCE.repeat(Math.ceil((tok * 4) / SENTENCE.length))
+
+let nonce = 0
+async function ttft(promptText, { unique = true } = {}) {
+  // A unique prefix per call defeats the server's prompt cache so we measure COLD prompt-eval
+  // (the cache-miss cost a turn pays when its prefix changed). unique:false reuses the prefix to
+  // measure the WARM (cache-hit) cost.
+  const tag = unique ? `#${++nonce} ` : ''
+  const body = JSON.stringify({
+    messages: [
+      { role: 'system', content: 'You are concise.' },
+      { role: 'user', content: `${tag}${promptText}\n\nReply with exactly: OK` }
+    ],
+    max_tokens: 16,
+    temperature: 0,
+    stream: true,
+    stream_options: { include_usage: true },
+    chat_template_kwargs: { enable_thinking: false }
+  })
+  const t0 = performance.now()
+  const res = await fetch(ENDPOINT, {
+    method: 'POST',
+    headers: { 'content-type': 'application/json' },
+    body
+  })
+  if (!res.ok) throw new Error(`HTTP ${res.status}`)
+  const reader = res.body.getReader()
+  const dec = new TextDecoder()
+  let firstAt = 0
+  let buf = ''
+  let promptTokens = 0
+  let done = false
+  while (!done) {
+    const { value, done: d } = await reader.read()
+    if (d) break
+    buf += dec.decode(value, { stream: true })
+    let nl
+    while ((nl = buf.indexOf('\n')) >= 0) {
+      const line = buf.slice(0, nl).trim()
+      buf = buf.slice(nl + 1)
+      if (!line.startsWith('data:')) continue
+      const payload = line.slice(5).trim()
+      if (payload === '[DONE]') {
+        done = true
+        break
+      }
+      try {
+        const j = JSON.parse(payload)
+        const delta = j.choices?.[0]?.delta?.content
+        if (delta && firstAt === 0) firstAt = performance.now()
+        if (j.usage?.prompt_tokens) promptTokens = j.usage.prompt_tokens
+      } catch {
+        /* keepalive */
+      }
+    }
+  }
+  const total = performance.now() - t0
+  return { ttft: firstAt ? firstAt - t0 : total, total, promptTokens }
+}
+
+const median = (xs) => [...xs].sort((a, b) => a - b)[Math.floor(xs.length / 2)]
+
+async function main() {
+  console.log(`\nChat prompt-eval benchmark · ${ENDPOINT} · ${REPEATS} repeats\n`)
+  console.log('warming up…')
+  await ttft('warmup').catch(() => {})
+  console.log('done\n')
+  console.log('target   promptTok   TTFT(med ms)   ms/1k promptTok   total(med ms)')
+  const rows = []
+  for (const target of TARGETS) {
+    const filler = fillerForTokens(target)
+    const runs = []
+    for (let r = 0; r < REPEATS; r++) runs.push(await ttft(filler))
+    const pt = runs[runs.length - 1].promptTokens
+    const tt = median(runs.map((r) => r.ttft))
+    const tot = median(runs.map((r) => r.total))
+    rows.push({ target, pt, tt, tot })
+    console.log(
+      `${String(target).padStart(6)}   ${String(pt).padStart(8)}   ${tt.toFixed(0).padStart(11)}   ${((tt / pt) * 1000).toFixed(1).padStart(15)}   ${tot.toFixed(0).padStart(12)}`
+    )
+  }
+  const first = rows[0]
+  const last = rows[rows.length - 1]
+  console.log(
+    `\nCOLD TTFT grew ${(last.tt / first.tt).toFixed(1)}× from ${first.pt} to ${last.pt} prompt tokens ` +
+      `(${(last.pt / first.pt).toFixed(1)}× the tokens). ~linear ⇒ a cache-MISS turn scales with history.`
+  )
+  // Decisive: does the SAME big prompt, sent twice, hit the cache the 2nd time? If warm << cold, the
+  // engine reuses the KV prefix across turns → a growing history is NOT re-eval'd each turn (so the
+  // slowdown must come from something that BUSTS the prefix, e.g. changing injected context).
+  const bigFiller = fillerForTokens(16000)
+  const cold = await ttft(bigFiller, { unique: true })
+  const warm1 = await ttft(bigFiller, { unique: false })
+  const warm2 = await ttft(bigFiller, { unique: false })
+  console.log(
+    `\nCache test @~${cold.promptTokens} promptTok:  cold=${cold.ttft.toFixed(0)}ms  ` +
+      `warm=${Math.min(warm1.ttft, warm2.ttft).toFixed(0)}ms  ` +
+      `→ ${warm1.ttft < cold.ttft * 0.5 ? 'CACHE REUSED across identical prefixes (history growth is cheap when the prefix is stable)' : 'NO cache benefit (every turn re-evals — history growth is expensive)'}`
+  )
+}
+main().catch((e) => {
+  console.error('bench failed:', e.message)
+  process.exit(1)
+})