Instructions to use ixim/Image21-MLX-8bit with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use ixim/Image21-MLX-8bit with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] huggingface-cli download --local-dir Image21-MLX-8bit ixim/Image21-MLX-8bit
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Atomic Chat
File size: 5,206 Bytes
4f03424 | 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 | """Stage only verified weights, reproducible tools and paired evaluation evidence."""
import argparse
import json
import shutil
from pathlib import Path
from scripts.common import REVISION,RUNTIME_REVISION,sha256,header,write_json
from scripts.report import REQUIRED_CASES,pair_records,records
def manifest(root):
root=Path(root)
return [dict(path=p.relative_to(root).as_posix(),size=p.stat().st_size,sha256=sha256(p))
for p in sorted(root.rglob('*')) if p.is_file() and p.name!='MANIFEST.json'
and not any(x.startswith('.') or x=='__pycache__' for x in p.relative_to(root).parts)]
def validate_manifest(root):
root=Path(root).resolve()
rows=json.loads((root/'MANIFEST.json').read_text())
names=set()
for row in rows:
name=row['path']; original=root/name; path=original.resolve()
if original.is_symlink() or not path.is_relative_to(root) or name in names:
raise ValueError(f'Invalid manifest path: {name}')
names.add(name)
if path.stat().st_size!=row['size'] or sha256(path)!=row['sha256']:
raise ValueError(f'Manifest mismatch: {name}')
actual={p.relative_to(root).as_posix() for p in root.rglob('*') if p.is_file()
and not any(x.startswith('.') or x=='__pycache__' for x in p.relative_to(root).parts)}
if actual!=names|{'MANIFEST.json'}: raise ValueError('Unexpected/missing release files')
return rows
def validate_evaluation(root):
root=Path(root)
pairs=list(pair_records(records(root/'bf16'),records(root/'8bit')))
if {a['case_id'] for a,b in pairs}!=REQUIRED_CASES: raise ValueError('Incomplete seven-case suite')
for a,b in pairs:
if a['steps']!=40 or a['width']!=1024 or a['height']!=1024 or not a['warmup']:
raise ValueError('Release requires the full 1024px/40-step warmed protocol')
if b['mode']!='RGBA' or b['rgb_std']<1: raise ValueError('Broken output')
review=json.loads((root/'visual-review.json').read_text())
if review.get('approved_for_research_release') is not True:
raise ValueError('Visual review incomplete or failed')
if set(review.get('cases',{}))!=REQUIRED_CASES: raise ValueError('Missing visual case reviews')
for name in ('bf16','8bit'):
if review['results_sha256'][name]!=sha256(root/name/'results.jsonl'):
raise ValueError('Visual review does not match current results')
def main():
ap=argparse.ArgumentParser()
ap.add_argument('--model',type=Path,default=Path('models/Image21-MLX-8bit'))
ap.add_argument('--evaluation',type=Path,default=Path('artifacts/eval'))
ap.add_argument('--platform',choices=('huggingface','modelscope'),required=True)
ap.add_argument('--output',type=Path)
args=ap.parse_args(); out=args.output or Path('release')/args.platform
if out.exists(): raise FileExistsError(out)
validate_evaluation(args.evaluation)
conversion=json.loads((args.model/'conversion.json').read_text())
if conversion['status']!='converted_and_roundtrip_verified' or conversion['source_revision']!=REVISION or conversion['runtime_revision']!=RUNTIME_REVISION:
raise ValueError('Unverified conversion')
for comp in conversion['components'].values():
if not comp['exact_roundtrip']: raise ValueError('Roundtrip not verified')
for p in args.model.glob('*/*.safetensors'):
if 'modification_notice' not in header(p).get('__metadata__',{}): raise ValueError('Missing modification notice')
shutil.copytree(args.model,out,copy_function=shutil.copy2)
shutil.copytree(args.evaluation,out/'evaluation',ignore=shutil.ignore_patterns('._*','*.tmp'))
# Publish useful machine measurements without the workstation's absolute paths.
for precision in ('bf16','8bit'):
path=out/'evaluation'/precision/'environment.json'
env=json.loads(path.read_text())
env['model']=f'Image21-MLX-{precision}'
write_json(path,env)
audit=json.loads(Path('artifacts/source-audit.json').read_text())
audit.pop('source')
audit['source_model']='Qwen/Qwen-Image-2.1'
audit['note']='Public source-file hashes and precision inventory; local source directory omitted.'
write_json(out/'evaluation'/'source-files.json',audit)
shutil.copytree('scripts',out/'scripts',ignore=shutil.ignore_patterns('__pycache__'))
shutil.copytree('benchmarks',out/'benchmarks')
for name in ('requirements.txt','requirements.lock.txt'):
shutil.copy2(name,out/name)
shutil.copy2(f'cards/{args.platform}.md',out/'README.md')
shutil.copy2('artifacts/precision-probe.json',out/'evaluation'/'precision-probe.json')
for name in ('cache-parity-bf16.json','cache-parity-8bit.json','offload-parity.json'):
shutil.copy2(Path('artifacts')/name,out/'evaluation'/name)
for row in json.loads((args.evaluation/'runtime-manifest.json').read_text()):
if sha256(out/row['path'])!=row['sha256']:
raise ValueError(f'Runtime changed after evaluation: {row["path"]}')
write_json(out/'MANIFEST.json',manifest(out))
rows=validate_manifest(out)
print(f'Staged and verified {len(rows)} files in {out}')
if __name__=='__main__': main()
|