| """生成 ZeroPoint 的参考数据 tests/golden/zeropoint_ref.json：固定的参数 + 输入 + 标准输出。 |
| |
| 所有实现（zerocg 并行形式、numpy 递推形式、以后内核里的 C 版本）都拿它对答案， |
| 谁和它对不上谁就有 bug。**只有在有意修改架构时才重新生成**，并在提交说明里写清楚为什么。 |
| |
| python3 tests/make_golden.py |
| """ |
| import json |
| import os |
| import sys |
|
|
| import numpy as np |
|
|
| HERE = os.path.dirname(os.path.abspath(__file__)) |
| sys.path.insert(0, os.path.join(HERE, "..")) |
| import zerocg as zc |
| from models import ZeroX |
|
|
| CONFIG = dict(vocab_size=13, n_embd=16, n_head=2, n_slot=4, n_layer=2) |
|
|
| if __name__ == "__main__": |
| zc.set_default_dtype(np.float64) |
| np.random.seed(2026) |
| model = ZeroX(**CONFIG).eval() |
| |
| for p in model.parameters(): |
| p.data = p.data + np.random.randn(*p.shape) * 0.3 |
| idx = np.random.randint(0, CONFIG["vocab_size"], size=(2, 70)) |
| logits = model(idx).data |
| out = os.path.join(HERE, "golden", "zeropoint_ref.json") |
| data = {"idx": idx.tolist(), "logits": logits.tolist()} |
| data.update({"param:" + k: v.tolist() for k, v in model.state_dict().items()}) |
| data.update({"config:" + k: int(v) for k, v in CONFIG.items()}) |
| with open(out, "w", encoding="utf-8") as f: |
| json.dump(data, f, separators=(",", ":"), ensure_ascii=False) |
| f.write("\n") |
| print(f"写入 {out}：输入 {idx.shape}，输出 {logits.shape}，{len(model.state_dict())} 个参数张量") |
|
|
