#!/usr/bin/env python3
"""
llama.cpp router benchmark — LAPTOP (lw) edition → writes a results ledger for tuning.

This is the CPU analog of the fl (RX 7600 XT) bench.py. lw has NO discrete GPU, so all
the ROCm / VRAM / GTT / freeze-guard / force-unload machinery from the fl version is GONE
— on a CPU box there is no display-driving GPU to starve. The only resource ceiling is
system RAM, so instead of VRAM this records per-model **RAM footprint** (MemAvailable
delta, from /proc/meminfo) alongside decode/prefill/TTFT.

Per model it records decode/prefill t/s, TTFT, RAM used (Δ) / RAM free, and the preset's
knobs (ctx, threads, spec-type) parsed from config.ini. For spec/MTP presets the draft
acceptance % is appended to the spec column, read from the response's `timings` object
(the child server also prints a `draft acceptance` log line; the table is easier to compare).

Output: bench-results.md (latest run, overwritten) + bench-history.md (every run, appended)
— both next to this script. Run ON THE LAPTOP against its own server. Stdlib only.

  ./bench.py                            # all models the server lists
  ./bench.py -m id1,id2                 # only these
  ./bench.py -x some-id                 # skip some (e.g. not downloaded yet)
  ./bench.py -n 512 --ctx 8000          # longer gen + a long-context decode column

Note: the laptop runs `--models-max 1` (LRU eviction), so mid-sweep the RAM Δ is NET —
each load evicts the previous model, and a smaller model after a bigger one reads negative.
Only the first model after a server restart shows a true footprint.
Compare a change: edit config.ini → restart server → re-run → diff runs in bench-history.md.
"""
import argparse, json, os, re, subprocess, sys, time, urllib.request, urllib.error
from datetime import datetime

HERE = os.path.dirname(os.path.abspath(__file__))


def mem_available():
    """(available_bytes, total_bytes) from /proc/meminfo; (None, None) if unavailable."""
    try:
        info = {}
        for ln in open("/proc/meminfo", encoding="utf-8"):
            k, _, rest = ln.partition(":")
            m = re.search(r"(\d+)\s*kB", rest)
            if m:
                info[k.strip()] = int(m.group(1)) * 1024
        return info.get("MemAvailable"), info.get("MemTotal")
    except OSError:
        return None, None


def with_retries(fn, retries, wait, label=""):
    """Retry on 5xx/connection errors (model still mounting). A deterministic 'failed to
    load' from the router is NOT retried — that's a broken preset, not a slow mount."""
    last = None
    for i in range(retries + 1):
        try:
            return fn()
        except urllib.error.HTTPError as e:
            if e.code < 500:              # 4xx = real client error
                raise
            body = ""
            try:
                body = e.read().decode("utf-8", "ignore")
            except Exception:
                pass
            if "failed to load" in body:  # deterministic → don't burn retries
                raise RuntimeError("model failed to load — check the llama-server log") from None
            last = e
        except (urllib.error.URLError, ConnectionError, TimeoutError) as e:
            last = e
        if i < retries:
            print(f"    … {label} not ready ({last}); waiting {wait}s (try {i + 1}/{retries})", file=sys.stderr)
            time.sleep(wait)
    raise last


def _open(url, key, payload=None, timeout=600):
    data = json.dumps(payload).encode() if payload is not None else None
    hdrs = {"Authorization": f"Bearer {key}"}
    if data:
        hdrs["Content-Type"] = "application/json"
    return urllib.request.urlopen(urllib.request.Request(url, data=data, headers=hdrs), timeout=timeout)


def list_models(base, key):
    d = json.load(_open(base.rstrip("/") + "/models", key))
    return [m["id"] for m in d.get("data", [])]


def parse_config(path):
    """{preset: {ctx, threads, spec}} from the router config.ini."""
    sections, sec = {}, None
    try:
        lines = open(path, encoding="utf-8").read().splitlines()
    except OSError:
        return {}
    for ln in lines:
        s = ln.strip()
        if s.startswith("[") and s.endswith("]"):
            sec = s[1:-1]; sections[sec] = {}
        elif sec and "=" in s and not s.startswith("#"):
            k, v = s.split("=", 1)
            sections[sec][k.strip()] = v.split("#", 1)[0].strip()
    return {sec: {"ctx": kv.get("ctx-size", "-"),
                  "threads": kv.get("threads", "-"),
                  "spec": kv.get("spec-type", "-") or "-"}
            for sec, kv in sections.items()}


def run(base, key, model, prompt, n, temp, timeout):
    payload = {"model": model, "messages": [{"role": "user", "content": prompt}],
               "max_tokens": n, "temperature": temp, "stream": True,
               "stream_options": {"include_usage": True},
               # llama.cpp extension: makes the server embed its own timings in the stream →
               # the table shows the SAME predicted_per_second as the llama.cpp web UI
               # (without it we fall back to wall-clock estimates, which read lower).
               "timings_per_token": True}
    t0 = time.perf_counter(); ttft = None; ntok = 0; usage = None; timings = None
    for raw in _open(base.rstrip("/") + "/chat/completions", key, payload, timeout):
        line = raw.decode("utf-8", "ignore").strip()
        if not line.startswith("data:"):
            continue
        body = line[5:].strip()
        if body == "[DONE]":
            break
        try:
            c = json.loads(body)
        except ValueError:
            continue
        ch = c.get("choices") or [{}]
        d = ch[0].get("delta", {}) if ch else {}
        # reasoning models stream `reasoning_content` first and may never emit `content`
        if d.get("content") or d.get("reasoning_content"):
            if ttft is None:
                ttft = time.perf_counter() - t0
            ntok += 1
        if c.get("usage"):
            usage = c["usage"]
        if c.get("timings"):
            timings = c["timings"]
    total = time.perf_counter() - t0
    comp = (usage or {}).get("completion_tokens") or ntok
    ptok = (usage or {}).get("prompt_tokens")
    gen_s = total - (ttft or total)
    tg = comp / gen_s if gen_s > 0 else 0.0
    pp = (ptok / ttft) if (ptok and ttft) else None
    acc = None
    if timings:  # llama.cpp's own numbers are authoritative when the router forwards them
        tg = timings.get("predicted_per_second", tg)
        pp = timings.get("prompt_per_second", pp)
        ptok = timings.get("prompt_n", ptok)
        if timings.get("prompt_ms"):
            ttft = timings["prompt_ms"] / 1000
        if timings.get("draft_n"):  # spec decode active — mirror the child's acceptance stat
            acc = timings.get("draft_n_accepted", 0) / timings["draft_n"]
    return {"tg": tg, "pp": pp, "ttft": ttft or 0.0, "ptok": ptok or 0, "acc": acc}


def gb(b):
    return f"{b / 1e9:.1f} GB" if b else "-"


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("-u", "--url", default="http://127.0.0.1:11343/v1",
                    help="run this ON the laptop against its own server (localhost)")
    ap.add_argument("-k", "--key", default="no-key-required")
    ap.add_argument("-m", "--models", default="all", help="comma-separated ids, or 'all'")
    ap.add_argument("-x", "--skip", default="", help="comma-separated ids to skip")
    ap.add_argument("-n", "--tokens", type=int, default=256)
    ap.add_argument("-t", "--temp", type=float, default=0.3)
    ap.add_argument("--ctx", type=int, default=0, help="also record decode at ~this many prompt tokens")
    ap.add_argument("--timeout", type=int, default=600, help="per-request stall timeout (s)")
    ap.add_argument("--config", default=os.path.join(HERE, "config.ini"))
    ap.add_argument("--out", default=os.path.join(HERE, "bench-results.md"))
    ap.add_argument("--history", default=os.path.join(HERE, "bench-history.md"))
    ap.add_argument("--retries", type=int, default=4, help="retries while a model is still mounting")
    ap.add_argument("--retry-wait", type=int, default=15, help="seconds between retries")
    ap.add_argument("--settle", type=float, default=1.0,
                    help="seconds to wait after warmup before reading RAM (let allocation settle)")
    a = ap.parse_args()

    models = list_models(a.url, a.key) if a.models == "all" else [m.strip() for m in a.models.split(",")]
    skip = {s.strip() for s in a.skip.split(",") if s.strip()}
    models = [m for m in models if m not in skip]
    cfg = parse_config(a.config)

    avail0, total_ram = mem_available()

    task = "Write a Python function that merges two sorted lists, with a short docstring and one example."
    filler = ("The quick brown fox jumps over the lazy dog. " * max(1, a.ctx // 9)) if a.ctx else None

    cols = ["model", "cfg ctx", "threads", "spec", "decode t/s", "prefill t/s", "TTFT s", "RAM Δ", "RAM free"]
    if a.ctx:
        cols.append(f"decode@{a.ctx // 1000}k")

    latest = open(a.out, "w", encoding="utf-8")
    hist = open(a.history, "a", encoding="utf-8")
    meta = (f"_Run {datetime.now():%Y-%m-%d %H:%M} · server `{a.url}` · RAM total {gb(total_ram)} · "
            f"baseline avail {gb(avail0)} · gen {a.tokens} tok (CPU inference, no GPU)_")

    def emit(line, both=True):
        print(line)
        latest.write(line + "\n"); latest.flush()
        if both:
            hist.write(line + "\n"); hist.flush()

    latest.write("# llama.cpp benchmark results — laptop (lw), latest run\n\n")
    hist.write(f"\n## run {datetime.now():%Y-%m-%d %H:%M}\n\n")
    emit(meta + "\n")
    emit("| " + " | ".join(cols) + " |")
    emit("|" + "|".join(["---"] * len(cols)) + "|")

    for m in models:
        c = cfg.get(m, {})
        base_cells = [m, c.get("ctx", "-"), c.get("threads", "-"), c.get("spec", "-")]
        pre_avail, _ = mem_available()   # per-model baseline (a model may already be resident)
        try:
            # warmup — retries wait out the mount; 'failed to load' aborts immediately
            with_retries(lambda: run(a.url, a.key, m, "hi", 8, a.temp, a.timeout), a.retries, a.retry_wait, m)
            if a.settle:
                time.sleep(a.settle)
            post_avail, _ = mem_available()
            free = post_avail
            delta = (pre_avail - post_avail) if (pre_avail is not None and post_avail is not None) else None
            # with --models-max 1 the Δ is net-of-eviction: negative = replaced a bigger model
            if delta is None:
                delta_cell = "-"
            elif delta > 0.2e9:
                delta_cell = gb(delta)
            elif delta < -0.2e9:
                delta_cell = f"-{gb(-delta)} (net swap)"
            else:
                delta_cell = "~0 (resident/net)"
            r = with_retries(lambda: run(a.url, a.key, m, task, a.tokens, a.temp, a.timeout),
                             a.retries, a.retry_wait, m)
            cells = base_cells.copy()
            if r.get("acc") is not None:
                cells[3] = f"{cells[3]} ({r['acc'] * 100:.0f}% acc)"
            cells += [f"{r['tg']:.1f}", (f"{r['pp']:.0f}" if r['pp'] else "-"),
                      f"{r['ttft']:.2f}", delta_cell, gb(free)]
            if a.ctx:
                rc = with_retries(lambda: run(a.url, a.key, m, filler + "\n\n" + task, a.tokens, a.temp, a.timeout),
                                  a.retries, a.retry_wait, m)
                cells.append(f"{rc['tg']:.1f}")
        except Exception as e:
            err = f"HTTP {e.code} (check server log)" if isinstance(e, urllib.error.HTTPError) else f"ERROR {e}"
            cells = base_cells + [err] + [""] * (len(cols) - len(base_cells) - 1)
        emit("| " + " | ".join(str(x) for x in cells) + " |")

    latest.write("\n_Tuning hints (CPU box): decode reads the model's active bytes every token — smaller "
                 "quants and small-active MoE are directly faster. **RAM free** is the real ceiling. "
                 "Sanity-check: decode t/s × model GB = effective GB/s; dense models measure ~8-10 GB/s "
                 "here (the practical wall), a model far below that is compute-bound (MoE/hybrid) → more "
                 "`threads` may help it; `threads-batch = 4` may help prefill either way. With "
                 "`--models-max 1`, mid-sweep RAM Δ is net-of-eviction. For spec/MTP models the draft "
                 "acceptance % is shown in the spec column. ⚠ Absolute numbers swing ±20-30% with desktop "
                 "load/thermals — only same-run rows are directly comparable._\n")
    latest.close(); hist.close()
    print(f"\nwrote {a.out} (+ appended {a.history})")


if __name__ == "__main__":
    main()