Download tools/compare_hf.py from Angshul/Falcon-H1-0.5B-Rust: direct link, hf CLI and curl.
- Browser
- Download file 1.72 kB
-
https://huggingface.co/Angshul/Falcon-H1-0.5B-Rust/resolve/main/tools/compare_hf.py
- Command line
-
hf download hf://Angshul/Falcon-H1-0.5B-Rust/tools/compare_hf.py
-
curl -L -o compare_hf.py https://huggingface.co/Angshul/Falcon-H1-0.5B-Rust/resolve/main/tools/compare_hf.py
1.72 kB
| """Compare Rust token-ID /v1/logits with the official 0.5B Hugging Face model. | |
| Run a Rust --checkpoint server first. This parity check is deliberately a report: | |
| there is no unsupported numerical-parity claim until the numbers are inspected. | |
| """ | |
| import argparse | |
| import json | |
| import urllib.request | |
| import numpy as np | |
| import torch | |
| from transformers import AutoModelForCausalLM | |
| def main(): | |
| p = argparse.ArgumentParser() | |
| p.add_argument('--repo', default='tiiuae/Falcon-H1-0.5B-Instruct') | |
| p.add_argument('--api', default='http://127.0.0.1:8080/v1/logits') | |
| p.add_argument('--ids', nargs='+', type=int, default=[1, 452, 1223]) | |
| p.add_argument('--device', default='cuda' if torch.cuda.is_available() else 'cpu') | |
| args = p.parse_args() | |
| model = AutoModelForCausalLM.from_pretrained(args.repo, torch_dtype=torch.float32).eval().to(args.device) | |
| with torch.no_grad(): | |
| y = model(input_ids=torch.tensor([args.ids], dtype=torch.long, device=args.device)).logits[0, -1, :].float().cpu().numpy() | |
| req = urllib.request.Request(args.api, data=json.dumps({'input_ids': args.ids}).encode(), headers={'Content-Type':'application/json'}) | |
| with urllib.request.urlopen(req, timeout=600) as resp: | |
| z = np.asarray(json.load(resp)['logits'], dtype='float32') | |
| if y.shape != z.shape: | |
| raise ValueError(f'vocab shape mismatch: HF {y.shape}, Rust {z.shape}') | |
| print('reference_top1', int(np.argmax(y))) | |
| print('rust_top1', int(np.argmax(z))) | |
| print('max_abs_error', float(np.max(np.abs(y-z)))) | |
| print('mean_abs_error', float(np.mean(np.abs(y-z)))) | |
| print('cosine_similarity', float(np.dot(y,z) / (np.linalg.norm(y)*np.linalg.norm(z)))) | |
| if __name__ == '__main__': | |
| main() | |