import time
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

# 1. 路径定义
MODEL_36_LAYERS = "./Qwen3-4B-Thinking-2507"  # 原始 36 层 Base
MODEL_32_LAYERS = "./QiMing-Polaris-Qwen3-4B-Final-Accelerated"  # 终极 32 层硬剪枝模型

prompt = "What is the 'Burden-based Training' method?"


def test_speed(model_path, name):
    print(f"\n正在加载模型 [{name}]...")
    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
    model = AutoModelForCausalLM.from_pretrained(
        model_path, dtype=torch.bfloat16, device_map="cuda", trust_remote_code=True
    )

    inputs = tokenizer(prompt, return_tensors="pt").to("cuda")

    # 预热
    with torch.no_grad():
        _ = model.generate(**inputs, max_new_tokens=10)

    # 强制生成 200 个 Token (不提前结束，精确测量 TPS)
    torch.cuda.synchronize()
    start_time = time.time()

    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            min_new_tokens=200,
            max_new_tokens=200,
            do_sample=False,  # 贪婪采样，排除随机性
        )

    torch.cuda.synchronize()
    end_time = time.time()

    latency = end_time - start_time
    tps = 200 / latency
    print(
        f"⚡ [{name}] 生成 200 Tokens 耗时: {latency:.4f} 秒 | 吞吐率: {tps:.2f} Tokens/s"
    )

    # 释放显存
    del model
    torch.cuda.empty_cache()
    return tps


# 运行对比
tps_36 = test_speed(MODEL_36_LAYERS, "原始 36 层模型")
tps_32 = test_speed(MODEL_32_LAYERS, "剪枝 32 层终极模型")

speedup = ((tps_32 - tps_36) / tps_36) * 100
print("\n" + "=" * 50)
print(f"🏆 物理剪枝硬加速结果: 每秒生成速度提升了 +{speedup:.2f}% !")
print("=" * 50)