← back to Exo

bench/exo_bench.py

937 lines

# type: ignore
#!/usr/bin/env python3
"""Tool-calling eval for exo's OpenAI-compatible API.

Tests whether models correctly:
- Trigger tool calls when appropriate
- Return valid JSON arguments matching function schemas
- Handle multi-turn tool use (call -> result -> final answer)
- Avoid calling tools when unnecessary

Start exo with a model first, then run:
    uv run python tool_call_eval.py --model <model-id>
    uv run python tool_call_eval.py --model <model-id> --host 10.0.0.5 --port 52415
    uv run python tool_call_eval.py --model <model-id> --repeat 3
    uv run python tool_call_eval.py --model <model-id> --scenarios weather_simple calculator_multi_turn
"""

from __future__ import annotations

import argparse
import contextlib
import itertools
import json
import sys
import threading
import time
from collections.abc import Callable
from concurrent.futures import ThreadPoolExecutor, as_completed
from pathlib import Path
from statistics import mean
from typing import Any

from exo_tools.client import ExoClient, ExoHttpError
from exo_tools.harness import (
    add_common_instance_args,
    capture_cluster_snapshot,
    find_existing_instance,
    instance_id_from_instance,
    node_ids_from_instance,
    nodes_used_in_instance,
    resolve_model_short_id,
    run_planning_phase,
    settle_and_fetch_placements,
    wait_for_instance_gone,
    wait_for_instance_ready,
)
from loguru import logger
from transformers import AutoTokenizer

# Monkey-patch for transformers 5.x compatibility
# Kimi's tokenization_kimi.py imports bytes_to_unicode from the old location
# which was moved in transformers 5.0.0rc2
try:
    import transformers.models.gpt2.tokenization_gpt2 as gpt2_tokenization
    from transformers.convert_slow_tokenizer import bytes_to_unicode

    if not hasattr(gpt2_tokenization, "bytes_to_unicode"):
        gpt2_tokenization.bytes_to_unicode = bytes_to_unicode  # type: ignore[attr-defined]
except ImportError:
    pass  # transformers < 5.0 or bytes_to_unicode not available


def load_tokenizer_for_bench(model_id: str) -> Any:
    """
    Load tokenizer for benchmarking, with special handling for Kimi models.

    Kimi uses a custom TikTokenTokenizer that transformers 5.x can't load via AutoTokenizer.
    This function replicates the logic from utils_mlx.py for bench compatibility.
    """
    model_id_lower = model_id.lower()

    if "kimi-k2" in model_id_lower:
        import importlib.util
        import types

        from huggingface_hub import snapshot_download

        # Download/get the model path
        model_path = Path(
            snapshot_download(
                model_id,
                allow_patterns=["*.json", "*.py", "*.tiktoken", "*.model", "*.jinja"],
            )
        )

        sys.path.insert(0, str(model_path))

        # Load tool_declaration_ts first (tokenization_kimi imports it with relative import)
        tool_decl_path = model_path / "tool_declaration_ts.py"
        if tool_decl_path.exists():
            spec = importlib.util.spec_from_file_location(
                "tool_declaration_ts", tool_decl_path
            )
            if spec and spec.loader:
                tool_decl_module = importlib.util.module_from_spec(spec)
                sys.modules["tool_declaration_ts"] = tool_decl_module
                spec.loader.exec_module(tool_decl_module)

        # Load tokenization_kimi with patched source (convert relative to absolute import)
        tok_path = model_path / "tokenization_kimi.py"
        source = tok_path.read_text()
        source = source.replace("from .tool_declaration_ts", "from tool_declaration_ts")
        spec = importlib.util.spec_from_file_location("tokenization_kimi", tok_path)
        if spec:
            tok_module = types.ModuleType("tokenization_kimi")
            tok_module.__file__ = str(tok_path)
            sys.modules["tokenization_kimi"] = tok_module
            exec(compile(source, tok_path, "exec"), tok_module.__dict__)  # noqa: S102
            TikTokenTokenizer = tok_module.TikTokenTokenizer  # noqa: N806
        else:
            from tokenization_kimi import TikTokenTokenizer  # type: ignore[import-not-found]  # noqa: I001

        hf_tokenizer: Any = TikTokenTokenizer.from_pretrained(model_path)

        # Patch encode to use internal tiktoken model directly
        # transformers 5.x has a bug in the encode->pad path for slow tokenizers
        def _patched_encode(text: str, **kwargs: object) -> list[int]:
            # Pass allowed_special="all" to handle special tokens like <|im_user|>
            return list(hf_tokenizer.model.encode(text, allowed_special="all"))

        hf_tokenizer.encode = _patched_encode

        return hf_tokenizer

    # TODO: Change back to using only transformers
    try:
        return AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
    except (AttributeError, ValueError):
        from huggingface_hub import snapshot_download
        from transformers import PretrainedConfig

        model_path = Path(
            snapshot_download(
                model_id,
                allow_patterns=[
                    "*.json",
                    "*.py",
                    "tokenizer.model",
                    "*.tiktoken",
                    "tiktoken.model",
                    "*.txt",
                    "*.jsonl",
                    "*.jinja",
                ],
            )
        )
        stub_kwargs: dict[str, Any] = {}
        config_file = model_path / "config.json"
        if config_file.exists():
            with open(config_file) as f:
                raw = json.load(f)
            for key in (
                "model_type",
                "max_position_embeddings",
                "vocab_size",
                "bos_token_id",
                "eos_token_id",
                "pad_token_id",
            ):
                if key in raw:
                    stub_kwargs[key] = raw[key]
        return AutoTokenizer.from_pretrained(
            str(model_path),
            config=PretrainedConfig(**stub_kwargs),
            trust_remote_code=True,
        )


def format_peak_memory(b: float) -> str:
    for unit in ["B", "KB", "MB", "GB", "TB"]:
        if b < 1024.0:
            return f"{b:.2f}{unit}"
        b /= 1024.0
    raise ValueError("You're using petabytes of memory. Something went wrong...")


_SAMPLER_METRICS = ("gpuUsage", "temp", "sysPower", "pcpuUsage", "ecpuUsage")


class SystemMetricsSampler:
    def __init__(self, client: ExoClient, node_ids: list[str], interval_s: float = 1.0):
        self._client = client
        self._node_ids = node_ids
        self._interval_s = interval_s
        self._samples: dict[str, list[tuple[float, dict[str, float]]]] = {
            nid: [] for nid in node_ids
        }
        self._stop = threading.Event()
        self._thread: threading.Thread | None = None

    def start(self) -> None:
        self._stop.clear()
        self._thread = threading.Thread(target=self._poll_loop, daemon=True)
        self._thread.start()

    def stop(self) -> None:
        self._stop.set()
        if self._thread:
            self._thread.join(timeout=5)

    def _poll_loop(self) -> None:
        while not self._stop.is_set():
            t = time.monotonic()
            for nid in self._node_ids:
                try:
                    data = self._client.get_node_system(nid)
                    if data:
                        self._samples[nid].append(
                            (t, {k: data.get(k, 0.0) for k in _SAMPLER_METRICS})
                        )
                except Exception:
                    pass
            self._stop.wait(self._interval_s)

    def energy_between(self, t0: float, t1: float) -> float:
        total_joules = 0.0
        for _nid, samples in self._samples.items():
            window = [(t, s["sysPower"]) for t, s in samples if t0 <= t <= t1]
            if len(window) >= 2:
                for i in range(1, len(window)):
                    dt = window[i][0] - window[i - 1][0]
                    avg_power = (window[i][1] + window[i - 1][1]) / 2
                    total_joules += avg_power * dt
            elif len(window) == 1:
                total_joules += window[0][1] * (t1 - t0)
        return total_joules

    def summarize(self) -> dict[str, dict[str, dict[str, float]]]:
        result: dict[str, dict[str, dict[str, float]]] = {}
        for nid, samples in self._samples.items():
            if not samples:
                continue
            metrics: dict[str, dict[str, float]] = {}
            for key in _SAMPLER_METRICS:
                values = [s[key] for t, s in samples]
                metrics[key] = {
                    "min": round(min(values), 2),
                    "max": round(max(values), 2),
                    "mean": round(mean(values), 2),
                    "samples": len(values),
                }
            result[nid] = metrics
        return result

    def print_summary(self, placement_label: str) -> None:
        summary = self.summarize()
        if not summary:
            return
        logger.info(f"--- System Metrics ({placement_label}) ---")
        for nid, metrics in summary.items():
            gpu = metrics.get("gpuUsage", {})
            temp = metrics.get("temp", {})
            power = metrics.get("sysPower", {})
            logger.info(
                f"  {nid}:  "
                f"GPU {gpu.get('mean', 0) * 100:.0f}% avg ({gpu.get('min', 0) * 100:.0f}–{gpu.get('max', 0) * 100:.0f}%)  |  "
                f"{temp.get('mean', 0):.1f}°C avg  |  "
                f"{power.get('mean', 0):.1f}W avg"
            )


def parse_int_list(values: list[str]) -> list[int]:
    items: list[int] = []
    for v in values:
        for part in v.split(","):
            part = part.strip()
            if part:
                items.append(int(part))
    return items


def run_one_completion(
    client: ExoClient,
    model_id: str,
    pp_hint: int,
    tg: int,
    prompt_sizer: PromptSizer,
    *,
    use_prefix_cache: bool = False,
    stream: bool = False,
) -> tuple[dict[str, Any], int]:
    content, pp_tokens = prompt_sizer.build(pp_hint)
    payload: dict[str, Any] = {
        "model": model_id,
        "messages": [{"role": "user", "content": content}],
        "max_tokens": tg,
        "logprobs": False,
        "use_prefix_cache": use_prefix_cache,
    }

    if not stream:
        payload["stream"] = False
        t0 = time.perf_counter()
        out = client.post_bench_chat_completions(payload)
        elapsed = time.perf_counter() - t0

        stats = out.get("generation_stats")
        power_usage = out.get("power_usage")
        choices = out.get("choices") or [{}]
        message = choices[0].get("message", {}) if choices else {}
        content = message.get("content") or ""
        preview = content[:200] if content else ""
    else:
        tokens = 0
        first_token_time = None
        t0 = time.perf_counter()
        text_parts: list[str] = []
        stats = None

        for raw_line in client.stream_bench_chat_completions(payload):
            line = raw_line.strip()
            if line.startswith(": generation_stats "):
                with contextlib.suppress(json.JSONDecodeError):
                    stats = json.loads(line[len(": generation_stats ") :])
                continue
            if not line.startswith("data: "):
                continue
            data = line[6:]
            if data == "[DONE]":
                break
            try:
                chunk = json.loads(data)
                delta = chunk.get("choices", [{}])[0].get("delta", {})
                if delta.get("content"):
                    if first_token_time is None:
                        first_token_time = time.perf_counter()
                    tokens += 1
                    text_parts.append(delta["content"])
            except json.JSONDecodeError:
                pass

        elapsed = time.perf_counter() - t0
        preview = "".join(text_parts)[:200]
        power_usage = None

        if not stats:
            ttft = (first_token_time - t0) if first_token_time else elapsed
            gen_time = elapsed - ttft if tokens > 1 else elapsed
            gen_tps = (tokens - 1) / gen_time if tokens > 1 and gen_time > 0 else 0.0
            prompt_tps = pp_tokens / ttft if ttft > 0 else 0.0
            stats = {
                "prompt_tokens": pp_tokens,
                "generation_tokens": tokens,
                "prompt_tps": round(prompt_tps, 2),
                "generation_tps": round(gen_tps, 2),
                "peak_memory_usage": {"inBytes": 0},
            }

    return {
        "elapsed_s": elapsed,
        "output_text_preview": preview,
        "stats": stats,
        "power_usage": power_usage,
    }, pp_tokens


class PromptSizer:
    def __init__(self, tokenizer: Any, atom: str = "a "):
        self.tokenizer = tokenizer
        self.atom = atom
        self.count_fn = PromptSizer._make_counter(tokenizer)
        self.base_tokens = self.count_fn("")

    @staticmethod
    def _make_counter(tokenizer: Any) -> Callable[[str], int]:
        def count_fn(user_content: str) -> int:
            messages = [{"role": "user", "content": user_content}]
            try:
                ids = tokenizer.apply_chat_template(
                    messages, tokenize=True, add_generation_prompt=True
                )
            except ValueError:
                # Models without a Jinja chat template (e.g. DeepSeek V4 which
                # ships its own Python encoder). Use the exo-side V4 encoder.
                from exo.worker.engines.mlx.deepseek_v4_encoding import (
                    encode_messages as encode_v4,
                )

                prompt = encode_v4(messages, thinking_mode="thinking")
                ids = tokenizer.encode(prompt, add_special_tokens=False)
            # Fix for transformers 5.x
            if hasattr(ids, "input_ids"):
                ids = ids.input_ids
            return int(len(ids))

        return count_fn

    def build(self, target_prompt_tokens: int) -> tuple[str, int]:
        target = int(target_prompt_tokens)
        if target < self.base_tokens:
            raise RuntimeError(
                f"Target ({target}) is smaller than template overhead ({self.base_tokens})."
            )

        # Estimate tokens per atom using a sample
        sample_count = 100
        sample_content = self.atom * sample_count
        sample_tokens = self.count_fn(sample_content) - self.base_tokens
        tokens_per_atom = sample_tokens / sample_count

        # Estimate starting point
        needed_tokens = target - self.base_tokens
        estimated_atoms = int(needed_tokens / tokens_per_atom)

        # Binary search to find exact atom count
        low, high = 0, estimated_atoms * 2 + 100
        while low < high:
            mid = (low + high) // 2
            tok = self.count_fn(self.atom * mid)
            if tok < target:
                low = mid + 1
            else:
                high = mid

        content = self.atom * low
        tok = self.count_fn(content)
        logger.info(f"{tok=}")

        if tok != target:
            raise RuntimeError(
                f"Overshot: got {tok} tokens (target {target}). "
                f"Pick a different atom (try ' a' or '\\n' or '0 ')."
            )

        return content, tok


def main() -> int:
    ap = argparse.ArgumentParser(
        prog="exo-bench",
        description="Benchmark exo model throughput across placement previews.",
    )
    add_common_instance_args(ap)
    ap.add_argument(
        "--pp",
        nargs="+",
        required=True,
        help="Prompt-size hints (ints). Accepts commas.",
    )
    ap.add_argument(
        "--tg",
        nargs="+",
        required=True,
        help="Generation lengths (ints). Accepts commas.",
    )
    ap.add_argument(
        "--repeat", type=int, default=1, help="Repetitions per (pp,tg) pair."
    )
    ap.add_argument(
        "--concurrency",
        nargs="+",
        default=["1"],
        help="Concurrency levels (ints). Accepts commas. E.g. --concurrency 1,2,4,8. Default 1.",
    )
    ap.add_argument(
        "--warmup",
        type=int,
        default=0,
        help="Warmup runs per placement (uses first pp/tg).",
    )
    ap.add_argument(
        "--json-out",
        default="bench/results.json",
        help="Write raw per-run results JSON to this path.",
    )
    ap.add_argument("--stdout", action="store_true", help="Write results to stdout")
    ap.add_argument(
        "--dry-run", action="store_true", help="List selected placements and exit."
    )
    ap.add_argument(
        "--all-combinations",
        action="store_true",
        help="Force all pp×tg combinations (cartesian product) even when lists have equal length.",
    )
    ap.add_argument(
        "--stream",
        action="store_true",
        help="Use /bench/chat/completions with streaming SSE response (bench=True still applies: no EOS detection, no KV cache).",
    )
    ap.add_argument(
        "--no-system-metrics",
        action="store_true",
        help="Disable GPU utilization, temperature, and power collection during inference.",
    )
    ap.add_argument(
        "--metrics-interval",
        type=float,
        default=1.0,
        help="System metrics polling interval in seconds (default: 1.0).",
    )
    ap.add_argument(
        "--use-prefix-cache",
        action="store_true",
        help="Enable KV prefix cache during bench (default: disabled for cold-cache measurements).",
    )
    args = ap.parse_args()

    pp_list = parse_int_list(args.pp)
    tg_list = parse_int_list(args.tg)
    if not pp_list or not tg_list:
        logger.error("pp and tg lists must be non-empty")
        return 2
    if args.repeat <= 0:
        logger.error("--repeat must be >= 1")
        return 2
    concurrency_list = parse_int_list(args.concurrency)
    if not concurrency_list or any(c <= 0 for c in concurrency_list):
        logger.error("--concurrency values must be >= 1")
        return 2

    if args.use_prefix_cache:
        logger.warning(
            "--use-prefix-cache: prompt TPS will be approximate. See METHODOLOGY.md for details."
        )
        if pp_list != sorted(pp_list):
            logger.warning(
                "--pp values are not in ascending order: prompt TPS will be less accurate. Use ascending --pp for best results."
            )

    # Log pairing mode
    use_combinations = args.all_combinations or len(pp_list) != len(tg_list)
    if use_combinations:
        logger.info(
            f"pp/tg mode: combinations (product) - {len(pp_list) * len(tg_list)} pairs"
        )
    else:
        logger.info(f"pp/tg mode: tandem (zip) - {len(pp_list)} pairs")

    client = ExoClient(args.host, args.port, timeout_s=args.timeout)
    short_id, full_model_id = resolve_model_short_id(
        client, args.model, force_download=args.force_download
    )

    tokenizer = load_tokenizer_for_bench(full_model_id)
    if tokenizer is None:
        raise RuntimeError("[exo-bench] tokenizer load failed")

    try:
        prompt_sizer = PromptSizer(tokenizer)
        logger.debug(f"[exo-bench] loaded tokenizer: {full_model_id} for prompt sizer")
    except Exception:
        logger.error("[exo-bench] tokenizer usable but prompt sizing failed")
        raise

    # Optionally reuse a running instance for this model
    reused_instance_id: str | None = None
    if args.reuse_instance:
        existing = find_existing_instance(client, full_model_id)
        if existing:
            reused_instance_id = existing
            logger.info(f"Reusing existing instance {reused_instance_id}")
        else:
            logger.warning(
                "--reuse-instance: no existing instance found, creating a new one"
            )

    if reused_instance_id is not None:
        # Use the existing instance directly — skip placement iteration
        selected = []
        download_duration_s = None
    else:
        selected = settle_and_fetch_placements(
            client, full_model_id, args, settle_timeout=args.settle_timeout
        )

        if not selected:
            logger.error("No valid placements matched your filters.")
            return 1

        selected.sort(
            key=lambda p: (
                str(p.get("instance_meta", "")),
                str(p.get("sharding", "")),
                nodes_used_in_instance(p["instance"]),
            ),
            reverse=True,
        )

        logger.debug(f"exo-bench model: short_id={short_id} full_id={full_model_id}")
        logger.info(f"placements: {len(selected)}")
        for p in selected:
            logger.info(
                f"  - {p['sharding']} / {p['instance_meta']} / nodes={nodes_used_in_instance(p['instance'])}"
            )

        if args.dry_run:
            return 0

        settle_deadline = (
            time.monotonic() + args.settle_timeout if args.settle_timeout > 0 else None
        )

        logger.info("Planning phase: checking downloads...")
        download_duration_s = run_planning_phase(
            client,
            full_model_id,
            selected[0],
            args.danger_delete_downloads,
            args.timeout,
            settle_deadline,
        )
        if download_duration_s is not None:
            logger.info(f"Download: {download_duration_s:.1f}s (freshly downloaded)")
        else:
            logger.info("Download: model already cached")

    cluster_snapshot = capture_cluster_snapshot(client)
    all_rows: list[dict[str, Any]] = []
    all_system_metrics: dict[str, dict[str, dict[str, float]]] = {}

    # If reusing an existing instance, run a single benchmark pass against it
    if reused_instance_id is not None:
        selected = [None]

    for preview in selected:
        created_instance = False
        if preview is not None:
            instance = preview["instance"]
            instance_id = instance_id_from_instance(instance)

            sharding = str(preview["sharding"])
            instance_meta = str(preview["instance_meta"])
            n_nodes = nodes_used_in_instance(instance)

            logger.info("=" * 80)
            logger.info(
                f"PLACEMENT: {sharding} / {instance_meta} / nodes={n_nodes} / instance_id={instance_id}"
            )

            # Delete any existing instances to free resources before placing
            try:
                state = client.request_json("GET", "/state")
                for old_id in list(state.get("instances", {}).keys()):
                    logger.info(f"Deleting stale instance {old_id}")
                    with contextlib.suppress(ExoHttpError):
                        client.request_json("DELETE", f"/instance/{old_id}")
                if state.get("instances"):
                    time.sleep(2)
            except Exception as e:
                logger.warning(f"Failed to clean up stale instances: {e}")

            client.request_json("POST", "/instance", body={"instance": instance})
            try:
                wait_for_instance_ready(client, instance_id)
            except (RuntimeError, TimeoutError) as e:
                logger.error(f"Failed to initialize placement: {e}")
                with contextlib.suppress(ExoHttpError):
                    client.request_json("DELETE", f"/instance/{instance_id}")
                continue

            time.sleep(1)
            created_instance = True
        else:
            instance_id = reused_instance_id
            sharding = "reused"
            instance_meta = "reused"
            n_nodes = 0
            logger.info("=" * 80)
            logger.info(f"Using existing instance {instance_id}")

        sampler: SystemMetricsSampler | None = None
        if not args.no_system_metrics and preview is not None:
            nids = node_ids_from_instance(instance)
            sampler = SystemMetricsSampler(
                ExoClient(args.host, args.port, timeout_s=30),
                nids,
                interval_s=args.metrics_interval,
            )
            sampler.start()

        def _do_one(c: ExoClient, pp: int, tg: int) -> tuple[dict[str, Any], int]:
            return run_one_completion(
                c,
                full_model_id,
                pp,
                tg,
                prompt_sizer,
                use_prefix_cache=args.use_prefix_cache,
                stream=args.stream,
            )

        try:
            for i in range(args.warmup):
                _do_one(client, pp_list[0], tg_list[0])
                logger.debug(f"  warmup {i + 1}/{args.warmup} done")

            # If pp and tg lists have same length, run in tandem (zip)
            # Otherwise (or if --all-combinations), run all combinations (cartesian product)
            if use_combinations:
                pp_tg_pairs = list(itertools.product(pp_list, tg_list))
            else:
                pp_tg_pairs = list(zip(pp_list, tg_list, strict=True))

            for pp, tg in pp_tg_pairs:
                for concurrency in concurrency_list:
                    logger.info(f"--- pp={pp} tg={tg} concurrency={concurrency} ---")
                    runs: list[dict[str, Any]] = []
                    inference_windows: list[tuple[float, float]] = []
                    for r in range(args.repeat):
                        time.sleep(3)

                        if concurrency <= 1:
                            # Sequential: single request
                            try:
                                inf_t0 = time.monotonic()
                                row, actual_pp_tokens = _do_one(client, pp, tg)
                                inference_windows.append((inf_t0, time.monotonic()))
                            except Exception as e:
                                logger.error(e)
                                continue
                            row.update(
                                {
                                    "model_short_id": short_id,
                                    "model_id": full_model_id,
                                    "placement_sharding": sharding,
                                    "placement_instance_meta": instance_meta,
                                    "placement_nodes": n_nodes,
                                    "instance_id": instance_id,
                                    "pp_tokens": actual_pp_tokens,
                                    "tg": tg,
                                    "repeat_index": r,
                                    "concurrency": 1,
                                    **(
                                        {"download_duration_s": download_duration_s}
                                        if download_duration_s is not None
                                        else {}
                                    ),
                                }
                            )
                            runs.append(row)
                            all_rows.append(row)
                        else:
                            # Concurrent: fire N requests in parallel
                            # Pre-build prompt once, barrier ensures simultaneous dispatch
                            content, actual_pp = prompt_sizer.build(pp)
                            pre_built_payload: dict[str, Any] = {
                                "model": full_model_id,
                                "messages": [{"role": "user", "content": content}],
                                "stream": False,
                                "max_tokens": tg,
                                "logprobs": False,
                                "use_prefix_cache": args.use_prefix_cache,
                            }
                            barrier = threading.Barrier(concurrency)
                            batch_start = threading.Event()
                            batch_t0: float = 0.0
                            batch_results: list[tuple[dict[str, Any], int]] = []
                            batch_errors = 0

                            def _run_concurrent(
                                idx: int,
                                _barrier: threading.Barrier = barrier,
                                _batch_start: threading.Event = batch_start,
                                _payload: dict[str, Any] = pre_built_payload,
                                _actual_pp: int = actual_pp,
                            ) -> tuple[dict[str, Any], int]:
                                nonlocal batch_t0
                                c = ExoClient(
                                    args.host, args.port, timeout_s=args.timeout
                                )
                                if _barrier.wait() == 0:
                                    batch_t0 = time.perf_counter()
                                    _batch_start.set()
                                else:
                                    _batch_start.wait()
                                t0 = batch_t0
                                out = c.post_bench_chat_completions(_payload)
                                elapsed = time.perf_counter() - t0
                                stats = out.get("generation_stats")
                                power_usage = out.get("power_usage")
                                choices = out.get("choices") or [{}]
                                message = (
                                    choices[0].get("message", {}) if choices else {}
                                )
                                text = message.get("content") or ""
                                return {
                                    "elapsed_s": elapsed,
                                    "output_text_preview": text[:200],
                                    "stats": stats,
                                    "power_usage": power_usage,
                                }, _actual_pp

                            inf_t0 = time.monotonic()
                            with ThreadPoolExecutor(max_workers=concurrency) as pool:
                                futures = {
                                    pool.submit(_run_concurrent, i): i
                                    for i in range(concurrency)
                                }
                                for fut in as_completed(futures):
                                    try:
                                        batch_results.append(fut.result())
                                    except Exception as e:
                                        logger.error(f"Concurrent request failed: {e}")
                                        batch_errors += 1
                            batch_wall_s = (
                                max(x["elapsed_s"] for x, _ in batch_results)
                                if batch_results
                                else time.perf_counter() - batch_t0
                            )
                            inference_windows.append((inf_t0, time.monotonic()))

                            for idx, (row, actual_pp_tokens) in enumerate(
                                batch_results
                            ):
                                row.update(
                                    {
                                        "model_short_id": short_id,
                                        "model_id": full_model_id,
                                        "placement_sharding": sharding,
                                        "placement_instance_meta": instance_meta,
                                        "placement_nodes": n_nodes,
                                        "instance_id": instance_id,
                                        "pp_tokens": actual_pp_tokens,
                                        "tg": tg,
                                        "repeat_index": r,
                                        "concurrency": concurrency,
                                        "concurrent_index": idx,
                                        **(
                                            {"download_duration_s": download_duration_s}
                                            if download_duration_s is not None
                                            else {}
                                        ),
                                    }
                                )
                                runs.append(row)
                                all_rows.append(row)

                            if batch_results:
                                valid_gen_tps = [
                                    x["stats"]["generation_tps"]
                                    for x, _ in batch_results
                                    if x["stats"]["generation_tps"] > 0
                                ]
                                per_req_tps = (
                                    max(valid_gen_tps) if valid_gen_tps else 0.0
                                )
                                agg_gen_tps = per_req_tps * concurrency
                                logger.info(
                                    f"[concurrent {concurrency}x]  "
                                    f"agg_gen_tps={agg_gen_tps:.2f}  "
                                    f"per_req_tps={per_req_tps:.2f}  "
                                    f"wall_s={batch_wall_s:.2f}  "
                                    f"errors={batch_errors}"
                                )

                    if runs:
                        prompt_tps = mean(x["stats"]["prompt_tps"] for x in runs)
                        valid_gen = [
                            x["stats"]["generation_tps"]
                            for x in runs
                            if x["stats"]["generation_tps"] > 0
                        ]
                        per_req_tps = max(valid_gen) if valid_gen else 0.0
                        gen_tps = per_req_tps * concurrency
                        ptok = mean(x["stats"]["prompt_tokens"] for x in runs)
                        gtok = mean(x["stats"]["generation_tokens"] for x in runs)

                        def _peak_bytes(s: dict[str, Any]) -> float:
                            pm = s["peak_memory_usage"]
                            return pm.get("inBytes") or pm.get("in_bytes", 0)

                        peak = mean(_peak_bytes(x["stats"]) for x in runs)
                        summary = (
                            f"prompt_tps={prompt_tps:.2f} gen_tps={gen_tps:.2f}    "
                            f"prompt_tokens={ptok} gen_tokens={gtok}    "
                            f"peak_memory={format_peak_memory(peak)}"
                        )
                        if sampler and inference_windows:
                            joules = sum(
                                sampler.energy_between(t0, t1)
                                for t0, t1 in inference_windows
                            )
                            inf_seconds = sum(t1 - t0 for t0, t1 in inference_windows)
                            avg_watts = joules / inf_seconds if inf_seconds > 0 else 0
                            summary += f"    energy={joules:.1f}J ({avg_watts:.1f}W avg over {inf_seconds:.1f}s inference)"

                        # mean() not sum() across concurrent runs: each
                        # request's PowerSampler observes the same shared
                        # cluster state, so they all report the same figure.
                        prefill_energies = [
                            (x.get("power_usage") or {}).get("prefill_energy_joules")
                            for x in runs
                        ]
                        gen_energies = [
                            (x.get("power_usage") or {}).get("generation_energy_joules")
                            for x in runs
                        ]
                        prefill_vals = [e for e in prefill_energies if e is not None]
                        gen_vals = [e for e in gen_energies if e is not None]
                        if prefill_vals and gen_vals:
                            avg_pref = mean(prefill_vals)
                            avg_gen = mean(gen_vals)
                            summary += (
                                f"    prefill_energy={avg_pref:.1f}J  "
                                f"gen_energy={avg_gen:.1f}J"
                            )
                        logger.info(f"{summary}\n")
                    time.sleep(2)
        finally:
            if sampler:
                sampler.stop()
                placement_label = f"{sharding}/{instance_meta}/{n_nodes} nodes"
                sampler.print_summary(placement_label)
                placement_metrics = sampler.summarize()
                if placement_metrics:
                    all_system_metrics.update(placement_metrics)

            if created_instance and instance_id is not None:
                try:
                    client.request_json("DELETE", f"/instance/{instance_id}")
                except ExoHttpError as e:
                    if e.status != 404:
                        raise
                wait_for_instance_gone(client, instance_id)
                logger.debug(f"Deleted instance {instance_id}")

                time.sleep(5)

    output: dict[str, Any] = {"runs": all_rows}
    if cluster_snapshot:
        output["cluster"] = cluster_snapshot
    if all_system_metrics:
        output["system_metrics"] = all_system_metrics

    if args.stdout:
        json.dump(output, sys.stdout, indent=2, ensure_ascii=False)
    elif args.json_out:
        with open(args.json_out, "w", encoding="utf-8") as f:
            json.dump(output, f, indent=2, ensure_ascii=False)
        logger.debug(f"\nWrote results JSON: {args.json_out}")

    return 0


if __name__ == "__main__":
    raise SystemExit(main())