fairbench-ai 0.3.0__py3-none-any.whl

fairbench/__init__.py

@@ -0,0 +1,4 @@
+"""FairBench - AI model benchmarking suite with SWE-bench integration."""
+
+__version__ = "0.3.0"
+__all__ = ["backend", "cli", "config", "data", "engine", "quality", "reporter"]

fairbench/backend.py

@@ -0,0 +1,211 @@
+"""Speed and memory benchmarking for AI models."""
+
+from __future__ import annotations
+
+import os
+import subprocess
+import time
+
+import psutil
+
+from fairbench.types import MemoryResult, SpeedResult
+
+
+def benchmark_speed(
+    model_path: str,
+    prompt: str = "The quick brown fox jumps over the lazy dog. " * 10,
+    n_ctx: int = 4096,
+    n_gpu_layers: int = -1,
+    n_threads: int = 0,
+    runs: int = 3,
+) -> SpeedResult:
+    """Measure TTFT, TPOT, and throughput using llama-cpp-python bindings.
+
+    Falls back to llama-bench CLI if bindings unavailable.
+    """
+    try:
+        from llama_cpp import Llama as _Llama  # type: ignore
+    except ImportError:
+        return _benchmark_with_llama_bench(model_path, prompt, runs)
+
+    return _benchmark_with_llama_cpp(
+        model_path, prompt, n_ctx, n_gpu_layers, n_threads, runs, _Llama
+    )
+
+
+def _benchmark_with_llama_cpp(
+    model_path: str,
+    prompt: str,
+    n_ctx: int,
+    n_gpu_layers: int,
+    n_threads: int,
+    runs: int,
+    Llama: type,  # type: ignore[type-arg]
+) -> SpeedResult:
+    """Benchmark using llama-cpp-python directly."""
+    ttfts: list[float] = []
+    tpots: list[float] = []
+    throughputs: list[float] = []
+
+    for i in range(runs):
+        try:
+            llm = Llama(  # type: ignore[no-untyped-call]
+                model_path=model_path,
+                n_ctx=n_ctx,
+                n_gpu_layers=n_gpu_layers,
+                n_threads=n_threads if n_threads > 0 else None,
+                verbose=False,
+            )
+
+            start = time.time()
+            response = llm.create_completion(
+                prompt=prompt,
+                max_tokens=512,
+                temperature=0.7,
+                stream=True,
+            )
+
+            ttft = time.time() - start
+            ttfts.append(ttft * 1000)  # ms
+
+            token_count = 0
+            last_time = time.time()
+            for chunk in response:
+                if time.time() - last_time > 0.01:
+                    elapsed = time.time() - last_time
+                    tpots.append((token_count / elapsed) if elapsed > 0 else 0)
+                    token_count = 0
+                    last_time = time.time()
+                token_count += 1
+
+            total_time = time.time() - start
+            throughput = token_count / total_time if total_time > 0 else 0
+            throughputs.append(throughput)
+
+            llm.close()
+        except (OSError, RuntimeError) as e:
+            print(f"    Run {i + 1} failed: {e}")
+
+    return SpeedResult(
+        ttft_avg=sum(ttfts) / len(ttfts) if ttfts else 0.0,
+        tpot_avg=sum(tpots) / len(tpots) if tpots else 0.0,
+        throughput_avg=sum(throughputs) / len(throughputs) if throughputs else 0.0,
+        runs=len(ttfts),
+    )
+
+
+def _benchmark_with_llama_bench(model_path: str, prompt: str, runs: int) -> SpeedResult:
+    """Fallback: use llama-bench CLI tool."""
+    try:
+        cmd = [
+            "llama-bench",
+            "-m",
+            model_path,
+            "-n",
+            "512",
+            "-ngl",
+            "0",
+            "-p",
+            str(len(prompt.split())),
+            "-t",
+            "0",
+            "-r",
+            str(runs),
+        ]
+        result = subprocess.run(cmd, check=False, capture_output=True, text=True, timeout=300)
+        # Parse output - llama-bench format varies by version
+        output = result.stdout
+        return SpeedResult(runs=runs, metadata={"raw_output": output})
+    except (FileNotFoundError, subprocess.TimeoutExpired) as e:
+        return SpeedResult(runs=runs, metadata={"error": str(e)})
+
+
+def benchmark_memory(
+    model_path: str,
+    prompt: str = "The quick brown fox jumps over the lazy dog. " * 10,
+    n_ctx: int = 4096,
+    n_gpu_layers: int = -1,
+    n_threads: int = 0,
+) -> MemoryResult:
+    """Measure VRAM and RAM peak usage during inference."""
+    try:
+        from llama_cpp import Llama as _Llama  # type: ignore
+    except ImportError:
+        return MemoryResult(metadata={"reason": "llama-cpp-python not installed"})
+
+    return _measure_memory_llama_cpp(
+        model_path,
+        prompt,
+        n_ctx,
+        n_gpu_layers,
+        n_threads,
+        _Llama,  # type: ignore[arg-type]
+    )
+
+
+def _measure_memory_llama_cpp(
+    model_path: str,
+    prompt: str,
+    n_ctx: int,
+    n_gpu_layers: int,
+    n_threads: int,
+    Llama: type,  # type: ignore[type-arg]
+) -> MemoryResult:
+    """Measure memory using llama-cpp-python with peak tracking."""
+    vram_samples: list[float] = []
+    ram_samples: list[float] = []
+
+    try:
+        # Get initial memory state
+        initial_ram = psutil.Process().memory_info().rss
+
+        llm = Llama(  # type: ignore[no-untyped-call]
+            model_path=model_path,
+            n_ctx=n_ctx,
+            n_gpu_layers=n_gpu_layers,
+            n_threads=n_threads if n_threads > 0 else None,
+            verbose=False,
+        )
+
+        # Measure VRAM via nvidia-smi if available
+        if os.path.exists("/usr/bin/nvidia-smi"):
+            result = subprocess.run(
+                [
+                    "nvidia-smi",
+                    "--query-gpu=memory.used",
+                    "--format=csv,noheader,nounits",
+                ],
+                check=False,
+                capture_output=True,
+                text=True,
+            )
+            if result.returncode == 0:
+                vram_samples = [
+                    float(line.strip())
+                    for line in result.stdout.strip().split("\n")
+                    if line.strip()
+                ]
+
+        # Run inference
+        llm.create_completion(
+            prompt=prompt,
+            max_tokens=256,
+            temperature=0.7,
+        )
+
+        # Peak RAM
+        current_ram = psutil.Process().memory_info().rss
+        ram_peak = (current_ram - initial_ram) / (1024 * 1024)  # MB
+        ram_samples.append(ram_peak)
+
+        llm.close()
+
+        return MemoryResult(
+            vram_peak_mb=sum(vram_samples) / len(vram_samples) if vram_samples else 0.0,
+            ram_peak_mb=ram_peak,
+            vram_avg_mb=sum(vram_samples) / len(vram_samples) if vram_samples else 0.0,
+            ram_avg_mb=sum(ram_samples) / len(ram_samples) if ram_samples else 0.0,
+            metadata={"model_size_gb": os.path.getsize(model_path) / (1024**3)},
+        )
+    except (OSError, RuntimeError) as e:
+        return MemoryResult(metadata={"error": str(e)})

fairbench/cli.py

@@ -0,0 +1,377 @@
+"""CLI for FairBench."""
+
+from __future__ import annotations
+
+import click
+
+from fairbench.config import load_config, save_config
+from fairbench.engine import BenchmarkEngine
+from fairbench.swebench import SWEBENCH_SOURCES
+from fairbench.types import BenchmarkCategory
+
+
+@click.group()
+@click.version_option(version="0.3.0")
+@click.option("--config", "-c", "config_path", help="Path to config file", default=None)
+@click.pass_context
+def cli(ctx, config_path):
+    """FairBench - AI model benchmarking suite with SWE-bench integration.
+
+    FairBench evaluates AI models across multiple dimensions:
+    - Quality: MMLU, HellaSwag, GSM8K, HumanEval, IFEval, TruthfulQA, ARC-Challenge, LiveBench
+    - Speed: TTFT, TPOT, throughput
+    - Memory: VRAM and RAM peak usage
+    - SWE-bench: GitHub issue resolution capability
+    """
+    ctx.ensure_object(dict)
+    ctx.obj["config"] = load_config(config_path)
+
+
+@cli.command()
+@click.argument("model_path")
+@click.option(
+    "--category",
+    "-k",
+    "category",
+    type=click.Choice(["quality", "speed", "memory", "swebench", "all"]),
+    default="all",
+    help="Benchmark category to run",
+)
+@click.option("--tasks", "-t", "tasks", multiple=True, help="Specific quality benchmarks to run")
+@click.option("--n-ctx", "-n", default=None, type=int, help="Context window size")
+@click.option("--n-gpu-layers", "-g", default=None, type=int, help="GPU layers to offload")
+@click.option("--n-threads", "-j", default=None, type=int, help="CPU threads")
+@click.option("--quant", "-q", default=None, help="Quantization type")
+@click.option("--lite/--no-lite", default=None, help="Use lite mode for quality benchmarks")
+@click.option(
+    "--swebench-split",
+    "-s",
+    "swebench_split",
+    type=click.Choice(list(SWEBENCH_SOURCES.keys())),
+    default=None,
+    help="SWE-bench split to use",
+)
+@click.option(
+    "--swebench-limit",
+    "-l",
+    "swebench_limit",
+    type=int,
+    default=None,
+    help="Limit SWE-bench instances",
+)
+@click.option("--publish/--no-publish", default=None, help="Override publish setting")
+@click.pass_context
+def run(
+    ctx,
+    model_path,
+    category,
+    tasks,
+    n_ctx,
+    n_gpu_layers,
+    n_threads,
+    quant,
+    lite,
+    swebench_split,
+    swebench_limit,
+    publish,
+):
+    """Run benchmarks on a model.
+
+    MODEL_PATH is the path to the GGUF model file.
+
+    Examples:
+    \b
+    fairbench run ~/models/qwen3-35b-q4_k_m.gguf
+    fairbench run ~/models/model.gguf --category quality --tasks mmlu hellaswag
+    fairbench run ~/models/model.gguf --category swebench --swebench-split lite
+    fairbench run ~/models/model.gguf --category all --lite
+    """
+    config = ctx.obj["config"]
+
+    # Override publish setting if specified
+    if publish is not None:
+        config.publish = publish
+
+    engine = BenchmarkEngine(config)
+
+    # Map category string to enum
+    category_map = {
+        "quality": BenchmarkCategory.QUALITY,
+        "speed": BenchmarkCategory.SPEED,
+        "memory": BenchmarkCategory.MEMORY,
+        "swebench": BenchmarkCategory.SWE_BENCH,
+        "all": None,  # Will run all
+    }
+
+    cat = category_map.get(category)
+
+    # Handle SWE-bench special cases
+    if category == "swebench" or (category == "all" and swebench_split):
+        lite_mode = "enabled" if lite else "disabled"
+        engine.run_benchmark(
+            model_path=model_path,
+            category=BenchmarkCategory.SWE_BENCH,
+            n_ctx=n_ctx,
+            n_gpu_layers=n_gpu_layers,
+            n_threads=n_threads,
+            quantization=quant,
+            lite_mode=lite_mode,
+        )
+        # Override split if specified
+        if swebench_split:
+            engine.swebench_client.config = config
+            engine.swebench_client.config.swebench_split = swebench_split
+        return
+
+    if category == "all":
+        # Run all benchmarks
+        engine.run_benchmark(
+            model_path=model_path,
+            category=BenchmarkCategory.QUALITY,
+            tasks=list(tasks) if tasks else None,
+            n_ctx=n_ctx,
+            n_gpu_layers=n_gpu_layers,
+            n_threads=n_threads,
+            quantization=quant,
+            lite_mode="enabled" if lite else "disabled",
+        )
+        engine.run_benchmark(
+            model_path=model_path,
+            category=BenchmarkCategory.SPEED,
+            n_ctx=n_ctx,
+            n_gpu_layers=n_gpu_layers,
+            n_threads=n_threads,
+            quantization=quant,
+        )
+        engine.run_benchmark(
+            model_path=model_path,
+            category=BenchmarkCategory.MEMORY,
+            n_ctx=n_ctx,
+            n_gpu_layers=n_gpu_layers,
+            n_threads=n_threads,
+            quantization=quant,
+        )
+        return
+
+    if cat is None:
+        click.echo("Invalid category")
+        return
+
+    engine.run_benchmark(
+        model_path=model_path,
+        category=cat,
+        tasks=list(tasks) if tasks else None,
+        n_ctx=n_ctx,
+        n_gpu_layers=n_gpu_layers,
+        n_threads=n_threads,
+        quantization=quant,
+        lite_mode="enabled" if lite else "disabled",
+    )
+
+
+@cli.command("list-benchmarks")
+@click.pass_context
+def list_benchmarks(ctx):
+    """List available quality benchmarks."""
+    engine = BenchmarkEngine(ctx.obj["config"])
+    benchmarks = engine.list_available_benchmarks()
+
+    click.echo("\nAvailable Quality Benchmarks:")
+    click.echo("-" * 50)
+    for name, info in benchmarks.items():
+        lite_marker = " [lite]" if info.get("lite") else ""
+        click.echo(f"  {name:20s} {info['name']:25s}{lite_marker}")
+        click.echo(f"    {info['description']}")
+    click.echo()
+
+
+@cli.command("list-swebench")
+@click.option(
+    "--split",
+    "-s",
+    "split",
+    type=click.Choice(list(SWEBENCH_SOURCES.keys())),
+    default="full",
+    help="SWE-bench split",
+)
+@click.option("--limit", "-l", type=int, default=10, help="Number of instances to show")
+@click.pass_context
+def list_swebench(ctx, split, limit):
+    """List available SWE-bench instances."""
+    engine = BenchmarkEngine(ctx.obj["config"])
+    instances = engine.get_swebench_instances(split=split, limit=limit)
+
+    click.echo(f"\nSWE-bench '{split}' instances (showing {len(instances)}):")
+    click.echo("-" * 70)
+    click.echo(f"  {'Instance ID':<35s} {'Repo':<25s} {'Language':<10s}")
+    click.echo("-" * 70)
+    for inst in instances:
+        click.echo(
+            f"  {inst['instance_id']:<35s} {inst['repo']:<25s} {inst['code_language']:<10s}"
+        )
+    click.echo()
+
+
+@cli.command("swebench-eval")
+@click.argument("model_path")
+@click.option(
+    "--split",
+    "-s",
+    "split",
+    type=click.Choice(list(SWEBENCH_SOURCES.keys())),
+    default="full",
+    help="SWE-bench split",
+)
+@click.option("--limit", "-l", type=int, default=5, help="Number of instances to evaluate")
+@click.option("--repo-filter", "-r", "repo_filter", help="Filter by repository")
+@click.pass_context
+def swebench_eval(ctx, model_path, split, limit, repo_filter):
+    """Run SWE-bench evaluation on a model.
+
+    MODEL_PATH is the path to the GGUF model file.
+    """
+    engine = BenchmarkEngine(ctx.obj["config"])
+
+    click.echo("\nSWE-bench Evaluation")
+    click.echo(f"Model: {model_path}")
+    click.echo(f"Split: {split}")
+    click.echo(f"Limit: {limit}")
+    if repo_filter:
+        click.echo(f"Repo filter: {repo_filter}")
+
+    # Fetch instances
+    instances = engine.swebench_client.fetch_instances(
+        split=split,
+        limit=limit,
+        repo_filter=repo_filter,
+    )
+
+    if not instances:
+        click.echo("No instances available")
+        return
+
+    click.echo(f"\nFound {len(instances)} instances")
+
+    # Run evaluation
+    results = engine.swebench_client.run_evaluation(
+        model_path=model_path,
+        instances=instances,
+        split=split,
+        limit=limit,
+    )
+
+    # Show results
+    stats = engine.swebench_client.compute_stats(results)
+    click.echo(f"\n{'=' * 50}")
+    click.echo("Results:")
+    click.echo(f"  Total: {stats.total_instances}")
+    click.echo(f"  Resolved: {stats.resolved}")
+    click.echo(f"  Failed: {stats.failed}")
+    click.echo(f"  Errors: {stats.errors}")
+    click.echo(f"  Score: {stats.avg_test_score:.1f}%")
+    click.echo(f"{'=' * 50}")
+
+
+@cli.command("config")
+@click.option("--show", "-s", "show", is_flag=True, help="Show current config")
+@click.option("--set", "-S", "set_kv", multiple=True, help="Set config key=value")
+@click.pass_context
+def config_cmd(ctx, show, set_kv):
+    """Manage FairBench configuration."""
+    config = ctx.obj["config"]
+
+    if show:
+        click.echo("\nCurrent Configuration:")
+        click.echo("-" * 40)
+        for key, value in config.__dict__.items():
+            click.echo(f"  {key}: {value}")
+        click.echo()
+        return
+
+    if set_kv:
+        for kv in set_kv:
+            if "=" not in kv:
+                click.echo(f"Invalid key=value: {kv}")
+                continue
+            key, value = kv.split("=", 1)
+            key = key.strip()
+            value = value.strip()
+
+            # Parse value types
+            if value.lower() in ("true", "false"):
+                value = value.lower() == "true"
+            else:
+                try:
+                    value = int(value)
+                except ValueError:
+                    try:
+                        value = float(value)
+                    except ValueError:
+                        pass  # Keep as string
+
+            if hasattr(config, key):
+                setattr(config, key, value)
+                click.echo(f"Set {key} = {value}")
+            else:
+                click.echo(f"Unknown config key: {key}")
+
+        # Save config
+        save_config(config)
+        click.echo("\nConfig saved")
+
+
+@cli.command("leaderboard")
+@click.option("--limit", "-l", type=int, default=20, help="Number of results to show")
+@click.pass_context
+def leaderboard(ctx, limit):
+    """Show the leaderboard."""
+    engine = BenchmarkEngine(ctx.obj["config"])
+    results = engine.reporter.get_leaderboard(limit=limit)
+
+    if not results:
+        click.echo("No results available")
+        return
+
+    click.echo("\nLeaderboard:")
+    click.echo("-" * 80)
+    click.echo(f"  {'Model':<30s} {'Quality':<10s} {'Speed':<10s} {'Memory':<10s} {'SWE':<10s}")
+    click.echo("-" * 80)
+
+    for result in results:
+        model = result.get("model_name", "unknown")[:30]
+        quality = result.get("quality_scores", {})
+        avg_quality = sum(quality.values()) / len(quality) if quality else 0
+        speed = result.get("speed", {}).get("throughput_tokens_per_sec", 0)
+        memory = result.get("memory", {}).get("vram_peak_mb", 0)
+        swebench = result.get("swebench", {})
+        sweb_score = (
+            (swebench.get("resolved", 0) / swebench.get("total", 1) * 100)
+            if swebench.get("total", 0) > 0
+            else 0
+        )
+
+        click.echo(
+            f"  {model:<30s} {avg_quality:<10.2f} {speed:<10.1f} {memory:<10.0f} {sweb_score:<10.1f}"
+        )
+    click.echo()
+
+
+@cli.command("results")
+@click.option("--category", "-c", help="Filter by category")
+@click.pass_context
+def results_cmd(ctx, category):
+    """List local benchmark results."""
+    data = BenchmarkEngine(ctx.obj["config"]).data
+    results = data.list_results()
+
+    if not results:
+        click.echo("No local results")
+        return
+
+    click.echo(f"\nLocal results ({len(results)}):")
+    for r in results:
+        click.echo(f"  {r}")
+
+
+if __name__ == "__main__":
+    cli()

fairbench/config.py

@@ -0,0 +1,106 @@
+"""Configuration management for FairBench."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import yaml
+
+from fairbench.types import BenchmarkConfig, LiteMode
+
+DEFAULT_CONFIG: dict[str, str | int | float | bool] = {
+    "api_url": "https://fairbench.vercel.app/api",
+    "default_n_ctx": 4096,
+    "default_n_gpu_layers": -1,
+    "default_quant": "q4_k_m",
+    "default_threads": 0,
+    "publish": True,
+    "results_dir": "/home/jarvispi/.fairbench/results",
+}  # type: ignore[dict-item]
+
+CONFIG_PATH = Path.home() / ".fairbench" / "config.yaml"
+
+
+def load_config(path: str | None = None) -> BenchmarkConfig:
+    """Load benchmark configuration from YAML file."""
+    config_path = Path(path) if path else CONFIG_PATH
+    if not config_path.exists():
+        return _default_config()
+
+    with open(config_path, "r") as f:
+        data = yaml.safe_load(f) or {}
+
+    return BenchmarkConfig(
+        api_url=data.get("api_url", DEFAULT_CONFIG["api_url"]),
+        default_n_ctx=data.get("default_n_ctx", DEFAULT_CONFIG["default_n_ctx"]),
+        default_n_gpu_layers=data.get(
+            "default_n_gpu_layers", DEFAULT_CONFIG["default_n_gpu_layers"]
+        ),
+        default_quant=data.get("default_quant", DEFAULT_CONFIG["default_quant"]),
+        default_threads=data.get("default_threads", DEFAULT_CONFIG["default_threads"]),
+        publish=data.get("publish", DEFAULT_CONFIG["publish"]),
+        results_dir=data.get("results_dir", DEFAULT_CONFIG["results_dir"]),
+        lite_mode=_parse_lite_mode(data.get("lite_mode")),
+        swebench_instances=data.get("swebench_instances", 0),
+        swebench_split=data.get("swebench_split", "full"),
+    )
+
+
+def save_config(config: BenchmarkConfig, path: str | None = None) -> None:
+    """Save benchmark configuration to YAML file."""
+    config_path = Path(path) if path else CONFIG_PATH
+    config_path.parent.mkdir(parents=True, exist_ok=True)
+
+    data = {
+        "api_url": config.api_url,
+        "default_n_ctx": config.default_n_ctx,
+        "default_n_gpu_layers": config.default_n_gpu_layers,
+        "default_quant": config.default_quant,
+        "default_threads": config.default_threads,
+        "publish": config.publish,
+        "results_dir": config.results_dir,
+    }
+
+    if config.lite_mode != LiteMode.AUTO:
+        data["lite_mode"] = config.lite_mode.value
+
+    if config.swebench_instances > 0:
+        data["swebench_instances"] = config.swebench_instances
+
+    if config.swebench_split != "full":
+        data["swebench_split"] = config.swebench_split
+
+    with open(config_path, "w") as f:
+        yaml.dump(data, f, default_flow_style=False)
+
+
+def _default_config() -> BenchmarkConfig:
+    """Return default configuration."""
+    return BenchmarkConfig(
+        api_url=str(DEFAULT_CONFIG["api_url"]),
+        default_n_ctx=int(DEFAULT_CONFIG["default_n_ctx"]),
+        default_n_gpu_layers=int(DEFAULT_CONFIG["default_n_gpu_layers"]),
+        default_quant=str(DEFAULT_CONFIG["default_quant"]),
+        default_threads=int(DEFAULT_CONFIG["default_threads"]),
+        publish=bool(DEFAULT_CONFIG["publish"]),
+        results_dir=str(DEFAULT_CONFIG["results_dir"]),
+    )
+
+
+def _parse_lite_mode(value: str | None) -> LiteMode:
+    """Parse lite mode from string value."""
+    if value is None:
+        return LiteMode.AUTO
+    mapping = {
+        "disabled": LiteMode.DISABLED,
+        "enabled": LiteMode.ENABLED,
+        "auto": LiteMode.AUTO,
+    }
+    return mapping.get(value.lower(), LiteMode.AUTO)
+
+
+def ensure_results_dir(results_dir: str | None = None) -> Path:
+    """Ensure results directory exists and return it."""
+    dir_path = Path(results_dir) if results_dir else CONFIG_PATH.parent / "results"
+    dir_path.mkdir(parents=True, exist_ok=True)
+    return dir_path