scripts/compare_safetensors.py

#!/usr/bin/env python3
import safetensors.torch, torch # any tensor library would work

def safe_open(filename, framework):
    if filename.endswith('.json'):
        class IndexFile:
            def __init__(self, filename, framework):
                import json
                with open(filename) as fh:
                    index = json.load(fh)
                files = {
                    file: safetensors.safe_open(file, framework=framework)
                    for file in index['weight_map'].values()
                }
                self.weight_map = {k:files[v] for k,v in index['weight_map'].items()}
            def get_tensor(self, name):
                return self.weight_map[name].get_tensor(name)
            def get_slice(self, name):
                return self.weight_map[name].get_slice(name)
            def keys(self):
                return self.weight_map.keys()
        return IndexFile(filename, framework=framework)
    else:
        return safetensors.safe_open(filename, framework=framework)

def compare(*fns):
    global files, mismatching_keys, avgs, dists, errs

    files = [safe_open(files, framework='pt') for files in fns]

    assert set(files[0].keys()) == set(files[1].keys())

    print('dtypes ...')
    dtypes = {k: [safetensors.torch._TYPES[file.get_slice(k).get_dtype()] for file in files] for k in files[0].keys()}
    dtypes = {k: [min(dts, key=lambda dt: dt.itemsize),max(dts, key=lambda dt: dt.itemsize)] for k, dts in dtypes.items()}
    mismatching_dtypes = [k for k, dts in dtypes.items() if dts[0] is not dts[1]]
    cmp_dtypes = {k: torch.int8 if dts[0] is torch.bool else dts[0] for k, dts in dtypes.items()}
    print('midpoints ...')
    avgs = {k:((files[0].get_tensor(k) + files[1].get_tensor(k))/2).to(dtypes[k][0]) for k in files[0].keys()}
    print('dists ...')
    dists = {k:(files[0].get_tensor(k).to(cmp_dtypes[k]) - files[1].get_tensor(k).to(cmp_dtypes[k])).abs() for k in files[0].keys()}

    print('keys ...')
    mismatching_keys = [k for k, d in dists.items() if (d!=0).any()]

    print(f'{len(mismatching_keys)/len(files[0].keys())*100:.2f}% keys mismatch')
    print(f'{len(mismatching_dtypes)/len(files[0].keys())*100:.2f}% dtypes mismatch')

    #errs = {k:(dists[k] / avgs[k]).nan_to_num() for k in files[0].keys()}

    #print('greatest scalar error:', max([e.max().item() for e in errs.values()])*100, '%')
    #print('cumulative scalar error:', sum([e.sum().item() for e in errs.values()])*100, '%')
    #print('total error:', (sum([d.sum() for d in dists.values()]) / sum([a.sum() for a in avgs.values()])).item()*100, '%')
    #print('greatest scalar dist:'`

    embed_name = [x for x in ['model.embed_tokens'] if x + '.output' in dtypes][0]
    head_name = [x for x in ['lm_head'] if x + '.output' in dtypes][0]

    print('input embed dist:', dists[embed_name+'.output'].sum().item())
    #print('input embed error sum:', errs[embed_name+'.output'].sum())

    print('output head dist:', dists[head_name+'.output'].sum().item())
    #print('output head error sum:', dists[head_name+'.output'].sum())

    for idx in range(2):
        head_tokens = files[idx].get_tensor(head_name+'.output')[0].softmax(dim=-1).sort(dim=-1,descending=True)
        print('file',idx,'tokens: ', end='')
        for range_idx, range_ in enumerate([range(3), range(-3,0)]):
            if range_idx > 0:
                print('.. ', end='')
            for idx in range_:
                for token in range(head_tokens.values.shape[-2]):
                    if range_idx == 0:
                        print(f'{head_tokens.indices[token][idx]}({head_tokens.values[token][idx]*100:.3f}%) ',end='')
                    else: # the % format changes from f to e to show smaller values
                        print(f'{head_tokens.indices[token][idx]}({head_tokens.values[token][idx]*100:.3e}%) ',end='')
        print()



if __name__ == '__main__':
    import sys
    assert len(sys.argv[1:]) == 2
    compare(*sys.argv[1:])