import json,pathlib
r=pathlib.Path('/home/lab/bench/ornith/modular/vllm')
base=json.loads((r/'cases/nf4-kv8-262k-c3.json').read_text())
for kv in ('auto','int8_per_token_head','int4_per_token_head'):
 c=json.loads(json.dumps(base));c['id']='full-int8-'+{'auto':'bf16','int8_per_token_head':'kv8','int4_per_token_head':'kv4'}[kv]
 c['env'].update(ORNITH_MODEL=str(r/'models/full-int8-text'),ORNITH_QUANT='compressed-tensors',ORNITH_LINEAR='marlin',ORNITH_LOAD_FORMAT='auto',ORNITH_KV=kv,ORNITH_CONTEXT='32768',ORNITH_SSM_DTYPE='float32')
 (r/'cases'/f"{c['id']}.json").write_text(json.dumps(c,indent=2))
print('Full INT8 cases prepared, no service changes.')
c=json.loads(json.dumps(base));c['id']='nf4-kv8-ssm32-32k';c['env'].update(ORNITH_SSM_DTYPE='float32',ORNITH_CONTEXT='32768')
(r/'cases'/f"{c['id']}.json").write_text(json.dumps(c,indent=2))
c=json.loads((r/'cases/full-int8-kv8.json').read_text());c['id']='full-int8-e5m2-flashinfer';c['env'].update(ORNITH_KV='fp8_e5m2',ORNITH_ATTN='FLASHINFER')
(r/'cases'/f"{c['id']}.json").write_text(json.dumps(c,indent=2))
for source,name,ctx in [('full-int8-bf16','full-int8-bf16-128k',131072),('full-int8-e5m2-flashinfer','full-int8-e5m2-262k',262144)]:
 c=json.loads((r/'cases'/f'{source}.json').read_text());c['id']=name;c['env'].update(ORNITH_CONTEXT=str(ctx),ORNITH_MEMORY='0.97')
 (r/'cases'/f"{c['id']}.json").write_text(json.dumps(c,indent=2))
c=json.loads((r/'cases/full-int8-bf16.json').read_text());c['id']='full-int8-bf16-eager';c['env']['ORNITH_GRAPH']='eager'
(r/'cases'/f"{c['id']}.json").write_text(json.dumps(c,indent=2))
