File size: 4,445 Bytes
5dc80b3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
#!/usr/bin/env python3
"""
Convert HRM checkpoints to safetensors and push to Hugging Face.

Usage:
    python push_to_hf.py \
        --baseline "checkpoints/Sudoku-extreme-1k-aug-1000 ACT-torch/HierarchicalReasoningModel_ACTV1 belligerent-squirrel/step_52080" \
        --tiered "checkpoints/Sudoku-extreme-1k-aug-1000 ACT-torch/HRM_Tiered realistic-dalmatian/step_52080" \
        --repo "Code2aum/HRM_based_evolution"
"""

import argparse
import os
import shutil
import torch
from safetensors.torch import save_file
from huggingface_hub import login, HfApi, upload_folder


def convert_checkpoint(ckpt_path, output_dir, model_name):
    """Load .pt checkpoint, save as safetensors + config."""
    os.makedirs(output_dir, exist_ok=True)

    # Load weights
    state_dict = torch.load(ckpt_path, map_location="cpu", weights_only=True)

    # Strip torch.compile prefix if present
    cleaned = {}
    for k, v in state_dict.items():
        key = k.removeprefix("_orig_mod.")
        cleaned[key] = v

    # Save as safetensors
    st_path = os.path.join(output_dir, "model.safetensors")
    save_file(cleaned, st_path)
    print(f"  ✓ Saved {st_path} ({os.path.getsize(st_path)/1e6:.1f} MB)")

    # Copy config
    ckpt_dir = os.path.dirname(ckpt_path)
    config_src = os.path.join(ckpt_dir, "all_config.yaml")
    if os.path.exists(config_src):
        shutil.copy2(config_src, os.path.join(output_dir, "config.yaml"))
        print(f"  ✓ Copied config.yaml")

    # Copy model source
    for src_file in ["hrm_act_v1.py", "hrm_tiered.py", "losses.py"]:
        src_path = os.path.join(ckpt_dir, src_file)
        if os.path.exists(src_path):
            shutil.copy2(src_path, os.path.join(output_dir, src_file))

    # Write a model card
    card = f"""---
tags:
  - hrm
  - hierarchical-reasoning
  - sudoku
  - pytorch
license: mit
---

# {model_name}

Hierarchical Reasoning Model trained on Sudoku-Extreme-1K (20,000 epochs).

## Architecture
- **Type**: {model_name}
- **Hidden Size**: 512
- **Heads**: 8
- **H/L Layers**: 4/4
- **H/L Cycles**: 2/2
- **Parameters**: ~27.3M

## Training
- **Dataset**: Sudoku-Extreme-1K (1000 puzzles, 1000 augmentations each)
- **Epochs**: 20,000
- **Batch Size**: 384
- **Learning Rate**: 7e-5
- **GPU**: NVIDIA RTX 4090

## Usage
```python
from safetensors.torch import load_file
state_dict = load_file("{model_name}/model.safetensors")
```
"""
    with open(os.path.join(output_dir, "README.md"), "w") as f:
        f.write(card)
    print(f"  ✓ Created README.md")


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--baseline", type=str, required=True)
    parser.add_argument("--tiered", type=str, required=True)
    parser.add_argument("--repo", type=str, default="Code2aum/HRM_based_evolution")
    parser.add_argument("--output-dir", type=str, default="hf_upload")
    parser.add_argument("--no-push", action="store_true", help="Convert only, skip push")
    args = parser.parse_args()

    print("=" * 60)
    print("  HRM → SafeTensors → Hugging Face")
    print("=" * 60)

    # Convert baseline
    base_dir = os.path.join(args.output_dir, "baseline_hrm_v1")
    print(f"\n  Converting Baseline...")
    convert_checkpoint(args.baseline, base_dir, "HRM_Baseline_V1")

    # Convert tiered
    tier_dir = os.path.join(args.output_dir, "tiered_hrm_sram_dram")
    print(f"\n  Converting Tiered...")
    convert_checkpoint(args.tiered, tier_dir, "HRM_Tiered_SRAM_DRAM")

    # Copy benchmark results if available
    bench_dir = "benchmark_results"
    if os.path.exists(bench_dir):
        dest = os.path.join(args.output_dir, "benchmark_results")
        if not os.path.exists(dest):
            shutil.copytree(bench_dir, dest)
            print(f"\n  ✓ Copied benchmark_results/")

    if args.no_push:
        print(f"\n  Files ready at: {args.output_dir}/")
        print("  Run without --no-push to upload to HF.")
        return

    # Push to HF
    print(f"\n  Logging in to Hugging Face...")
    login()

    print(f"\n  Uploading to {args.repo}...")
    api = HfApi()
    api.create_repo(repo_id=args.repo, repo_type="model", exist_ok=True)
    upload_folder(
        folder_path=args.output_dir,
        repo_id=args.repo,
        repo_type="model",
    )
    print(f"\n  ✓ Uploaded to https://huggingface.co/{args.repo}")

    print("\n" + "=" * 60)
    print("  Done!")
    print("=" * 60)


if __name__ == "__main__":
    main()