Add editable LLM penalties and preserve sampling through backup restore

deploy/configure_llm_sampling.py

@@ -0,0 +1,73 @@
+#!/usr/bin/env python3
+"""Explicit offline application of reviewed model-card defaults; never load weights."""
+import argparse
+import copy
+import json
+import os
+from pathlib import Path
+import subprocess
+import time
+import uuid
+
+QWEN_SOURCES={
+ 'jpetrina/Qwen3.8-27B-IQ4_XS-pure-GGUF',
+ 'vmarcelo/Qwen3.8-27B-MIX_GGUF',
+ 'JonathanColetti/Qwen3.8-27B-Uncensored-GGUF',
+}
+NEUTRAL=dict(repeat_penalty=1.0,presence_penalty=0.0,frequency_penalty=0.0)
+
+def recommendations(entry):
+ repo=entry.get('repo')
+ if repo in QWEN_SOURCES:
+  return dict(temperature=1.0,top_p=.95,top_k=20,**NEUTRAL)
+ if repo=='ggml-org/gemma-4-31B-it-GGUF':
+  return dict(temperature=1.0,top_p=.95,top_k=64,**NEUTRAL)
+ return None
+
+def configure(rows,entries,add_large_ultra=False):
+ result=copy.deepcopy(rows);report=[]
+ for row in result:
+  if row.get('kind')!='chat':continue
+  recommended=recommendations(entries.get(row['model_id'],{}))
+  if recommended is None:
+   report.append(dict(name=row['name'],state='unknown-source'));continue
+  changed=any(row['parameters'].get(k)!=v for k,v in recommended.items())
+  if changed:
+   row['parameters'].update(recommended);row['revision']+=1;row['updated_at']=time.time()
+  report.append(dict(name=row['name'],state='updated' if changed else 'unchanged',sampling=recommended))
+ if add_large_ultra:
+  medium=next((r for r in result if r.get('name')=='Qwen3.8 27B - Medium' and r.get('kind')=='chat'),None)
+  if not medium or entries.get(medium['model_id'],{}).get('repo')!='jpetrina/Qwen3.8-27B-IQ4_XS-pure-GGUF':
+   raise ValueError('Das geprüfte Medium-Profil fehlt; keine Profile erstellt.')
+  for suffix,context,split,micro in [('Large',192000,[86,14],256),('Ultra',262144,[80,20],128)]:
+   name='Qwen3.8 27B - '+suffix
+   if any(r['name']==name for r in result):continue
+   row=copy.deepcopy(medium);row.update(id=uuid.uuid4().hex,revision=1,name=name,updated_at=time.time())
+   row['parameters'].update(context=context,slots=1,batch=2048,ubatch=micro,tensor_split=split,split_mode='layer',gpu_offload='full',gpu_reserve_mode='none',gpu_reserve_mib={})
+   if suffix=='Ultra':row['parameters']['vision_device']='cpu'
+   result.append(row);report.append(dict(name=name,state='created',sampling=recommendations(entries[medium['model_id']])))
+ return result,report
+
+def main():
+ parser=argparse.ArgumentParser(description=__doc__)
+ parser.add_argument('--installation',type=Path,required=True)
+ parser.add_argument('--apply',action='store_true')
+ parser.add_argument('--add-large-ultra',action='store_true')
+ args=parser.parse_args();base=args.installation.resolve();state=base/'state';path=state/'profiles.json'
+ original=path.read_bytes();rows=json.loads(original)
+ entries={row['model_id']:json.loads((state/'models'/row['model_id']/'entry.json').read_text()) for row in rows}
+ configured,report=configure(rows,entries,args.add_large_ultra)
+ if args.apply and configured!=rows:
+  manifest=json.loads((base/'installation.json').read_text())
+  container=json.loads(subprocess.check_output(['docker','inspect',manifest['name']],text=True))[0]
+  if container['Config'].get('Labels',{}).get('de.casaderoll.athena-deck.standalone')!=str(base) or container['State']['Running']:
+   raise RuntimeError('Nur bei gestopptem, dieser Installation zugeordnetem Deck-Container anwenden.')
+  if path.read_bytes()!=original:raise RuntimeError('Profile wurden zwischenzeitlich geändert.')
+  backup=base/'backups'/str(time.time_ns());backup.mkdir(parents=True,mode=0o700)
+  (backup/'profiles.json').write_bytes(original)
+  temporary=path.with_suffix('.sampling.tmp');temporary.write_text(json.dumps(configured))
+  metadata=path.stat();os.chmod(temporary,metadata.st_mode);os.chown(temporary,metadata.st_uid,metadata.st_gid)
+  temporary.replace(path)
+ print(json.dumps(dict(applied=args.apply,profiles=report),ensure_ascii=False,indent=2))
+
+if __name__=='__main__':main()