examples/benchmark.py

#!/usr/bin/env python3
"""
Sixpert K2 - Quick Benchmark Script
====================================
Runs basic performance benchmarks for Sixpert K2 (MoE).

Key advantage: Despite 8.9B total parameters, only ~1.2B are active
per token, making inference faster than dense models of similar size.

Usage:
    python benchmark.py --model SixpertK2.gguf
"""

import argparse
import time
import sys

try:
    from llama_cpp import Llama
except ImportError:
    print("Installing llama-cpp-python...")
    import subprocess
    subprocess.check_call([sys.executable, "-m", "pip", "install", "llama-cpp-python"])
    from llama_cpp import Llama


def benchmark_generation(model_path: str, tokens: int = 512):
    """Benchmark token generation speed."""
    print("\n=== Generation Benchmark ===")
    print(f"Generating {tokens} tokens...\n")

    llm = Llama(
        model_path=model_path,
        n_ctx=4096,
        n_gpu_layers=-1,
        verbose=False,
    )

    start = time.time()
    output = llm(
        "<|im_start|>user\nWrite a detailed essay about artificial intelligence and its impact on society.<|im_end|>\n<|im_start|>assistant\n",
        max_tokens=tokens,
        temperature=0.6,
        stream=False,
    )
    elapsed = time.time() - start

    tokens_per_sec = tokens / elapsed

    print(f"Generated: {tokens} tokens")
    print(f"Time: {elapsed:.2f}s")
    print(f"Speed: {tokens_per_sec:.1f} tokens/sec")
    print(f"Note: Only ~1.2B active params per token (MoE advantage)")


def benchmark_context(model_path: str, context_length: int = 16384):
    """Benchmark long-context processing speed."""
    print(f"\n=== Long-Context Benchmark ===")
    print(f"Processing {context_length} token context...\n")

    llm = Llama(
        model_path=model_path,
        n_ctx=context_length + 512,
        n_gpu_layers=-1,
        verbose=False,
    )

    # Create a long context prompt
    filler = "The evolution of artificial intelligence has been marked by several key milestones. " * (context_length // 15)
    prompt = f"<|im_start|>user\n{filler}\nBased on the above text, what are the main themes discussed?<|im_end|>\n<|im_start|>assistant\n"

    start = time.time()
    output = llm(prompt, max_tokens=200, stream=False)
    elapsed = time.time() - start

    prompt_tokens = output["usage"]["prompt_eval_count"]
    eval_time = output["usage"].get("prompt_eval_time", 1000) / 1000

    print(f"Context tokens: {prompt_tokens}")
    print(f"Processing time: {eval_time:.2f}s")
    print(f"Speed: {prompt_tokens / eval_time:.1f} tokens/sec")


def benchmark_reasoning(model_path: str):
    """Benchmark deep reasoning capability."""
    print(f"\n=== Deep Reasoning Benchmark ===")
    print(f"Testing multi-step reasoning...\n")

    llm = Llama(
        model_path=model_path,
        n_ctx=8192,
        n_gpu_layers=-1,
        verbose=False,
    )

    start = time.time()
    output = llm(
        "<|im_start|>user\nProve that there are infinitely many prime numbers. Provide a complete, rigorous mathematical proof.<|im_end|>\n<|im_start|>assistant\n",
        max_tokens=2048,
        temperature=0.3,
        stream=False,
    )
    elapsed = time.time() - start

    response_text = output["choices"][0]["text"]
    print(f"Response length: {len(response_text)} chars")
    print(f"Time: {elapsed:.2f}s")
    print(f"Tokens/sec: {2048 / elapsed:.1f}")
    print(f"\nFirst 200 chars: {response_text[:200]}...")


def main():
    parser = argparse.ArgumentParser(description="Sixpert K2 Benchmark")
    parser.add_argument("--model", type=str, default="SixpertK2.gguf", help="Path to GGUF model")
    parser.add_argument("--gen-tokens", type=int, default=512, help="Generation benchmark tokens")
    parser.add_argument("--ctx-length", type=int, default=16384, help="Context benchmark length")
    parser.add_argument("--all", action="store_true", help="Run all benchmarks")

    args = parser.parse_args()

    print("=" * 60)
    print("  Sixpert K2 Benchmark Suite")
    print("  Deep Reasoning Engine (MoE)")
    print("  Total: ~8.9B | Active: ~1.2B/token")
    print("=" * 60)

    benchmark_generation(args.model, args.gen_tokens)
    benchmark_context(args.model, args.ctx_length)
    benchmark_reasoning(args.model)

    print("\n" + "=" * 60)
    print("  Benchmark complete!")
    print("=" * 60)


if __name__ == "__main__":
    main()