Download python/board_check.py from HY-2012/M5-TTS.AXERA: direct link, hf CLI and curl.
- Browser
- Download file 2.84 kB
-
https://huggingface.co/HY-2012/M5-TTS.AXERA/resolve/main/python/board_check.py
- Command line
-
hf download hf://HY-2012/M5-TTS.AXERA/python/board_check.py
-
curl -L -o board_check.py https://huggingface.co/HY-2012/M5-TTS.AXERA/resolve/main/python/board_check.py
2.84 kB
| #!/usr/bin/env python3 | |
| """AX650 per-graph accuracy check: axmodel vs the FP32 ONNX reference I/O. | |
| python3 board_check.py [--models models] [--eval eval_io] | |
| """ | |
| import argparse | |
| import os | |
| import sys | |
| import numpy as np | |
| sys.path.insert(0, os.path.dirname(os.path.abspath(__file__))) | |
| from m5_ax.tts import AxEngine # noqa: E402 | |
| GRAPHS = ["lm_prefill", "lm_decode", "depth_prefill", "depth_step", | |
| "codec_encoder", "codec_decoder"] | |
| def metrics(a, b): | |
| a = np.asarray(a, np.float64).ravel() | |
| b = np.asarray(b, np.float64).ravel() | |
| cos = float(a @ b / (np.linalg.norm(a) * np.linalg.norm(b) + 1e-12)) | |
| snr = float(10 * np.log10(np.sum(a * a) / (np.sum((a - b) ** 2) + 1e-12))) | |
| return cos, snr | |
| def main(): | |
| here = os.path.dirname(os.path.abspath(__file__)) | |
| ap = argparse.ArgumentParser() | |
| ap.add_argument("--models", default=os.path.join(here, "models")) | |
| ap.add_argument("--eval", default=os.path.join(here, "eval_io")) | |
| a = ap.parse_args() | |
| rows = [] | |
| for name in GRAPHS: | |
| path = os.path.join(a.models, name + ".axmodel") | |
| folder = os.path.join(a.eval, name) | |
| if not os.path.exists(path) or not os.path.isdir(folder): | |
| print(f"== {name}: missing, skip") | |
| continue | |
| engine = AxEngine(path) | |
| ins = {f[:-4][3:]: np.load(os.path.join(folder, f)) for f in sorted(os.listdir(folder)) | |
| if f.startswith("in_")} | |
| outs = {f[:-4][4:]: np.load(os.path.join(folder, f)) for f in sorted(os.listdir(folder)) | |
| if f.startswith("out_")} | |
| feeds = [] | |
| for i in range(next(iter(ins.values())).shape[0]): | |
| feeds.append({k: v[i] for k, v in ins.items()}) | |
| got_all = [engine.run(feed) for feed in feeds] | |
| names = engine.output_names | |
| print(f"== {name}") | |
| for i in range(len(feeds)): | |
| for out_name, value in zip(names, got_all[i]): | |
| if out_name not in outs: | |
| continue | |
| ref = outs[out_name][i] | |
| got = np.asarray(value).reshape(ref.shape) | |
| if np.issubdtype(ref.dtype, np.integer): | |
| same = float(np.mean(np.round(got) == ref)) | |
| rows.append((name, out_name, i, same, 99.0)) | |
| print(f" s{i} {out_name:12s} int-match {same*100:.3f}%") | |
| else: | |
| cos, snr = metrics(ref, got) | |
| rows.append((name, out_name, i, cos, snr)) | |
| print(f" s{i} {out_name:12s} cosine {cos:.6f} SNR {snr:6.2f} dB") | |
| bad = [r for r in rows if (r[3] < 0.99 and r[3] <= 1.0) or (r[3] <= 1.0 and r[4] < 20 and r[4] != 99.0)] | |
| print(f"\n{len(rows)} tensors checked, {len(bad)} below 0.99/20 dB") | |
| for r in bad: | |
| print(" LOW", r) | |
| if __name__ == "__main__": | |
| main() | |