"""Reference-expression pilot alongside background generation on one shared GPU.""" import hashlib import io import json import os from pathlib import Path import tarfile import time from datetime import datetime,timezone import torch from PIL import Image from native_flux_inference import create_native_pipeline, generate_native_image, encode_text from background_production import add_member,dump,BASE_REV,LORA_REV ROOT=Path(__file__).resolve().parents[1] STATE=ROOT/'experiments/character-production' BUFFER=Path('/tmp/vivarium-character-production') LORA_SHA='977642f4d7427fa3203b5b3b3c3a4ef9cb72186c59bd71b63e870ff53bfa6451' def now():return datetime.now(timezone.utc).isoformat() def metadata(person,expression,image_bytes,reference_sha,seconds,encode_seconds,shard,profile): record={k:v for k,v in person.items() if k!='expressions'} record.update(expression) record.update(dict(schema_version='1.0',generated_utc=now(),image_sha256=hashlib.sha256(image_bytes).hexdigest(), reference_image_sha256=reference_sha,reference_resize=dict(width=416 if reference_sha else 0,height=624 if reference_sha else 0,resample='PIL BICUBIC' if reference_sha else ''), inference=dict(model='black-forest-labs/FLUX.2-klein-base-9B',revision=BASE_REV, lora='laion/FLUX.2-klein-base-9B-vivarium-loras',lora_revision=LORA_REV,lora_subfolder='characters', lora_sha256=LORA_SHA,lora_rank=64,lora_alpha=64,lora_training_step=1500,lora_scale=1.0, steps=50,guidance_scale=4.0,seed=expression['seed'],width=832,height=1248,dtype='bfloat16', quantization=False,distillation=False,max_sequence_length=256,cfg_mode='batched',lora_fused=True, qkv_fused=True,compile_mode='default',negative_embedding_cached=True,positive_embedding_cached=True), runtime=dict(profile=profile,gpu=torch.cuda.get_device_name(0),physical_gpu=int(os.environ.get('CUDA_VISIBLE_DEVICES','0')), batch_size=1,generation_seconds=seconds,positive_text_encoding_seconds=encode_seconds, torch=torch.__version__,cuda=torch.version.cuda,includes_compile_warmup=False, gpu_shared_with_backgrounds=profile.startswith('shared'),model_cpu_offload=profile.startswith('shared')), storage=dict(shard=shard,image_member=expression['id']+'.png',json_member=expression['id']+'.json',prompt_member=expression['id']+'.txt',encoding='PNG RGB'))) return record def main(): BUFFER.mkdir(parents=True,exist_ok=True) plan=json.loads((STATE/'person-plan.json').read_text()) ids=['person-0002','person-0120','person-0253'] people=[next(p for p in plan if p['personality_id']==i) for i in ids] dump(STATE/'pilot-state.json',dict(status='loading CPU-offloaded pipeline',ids=ids,utc=now())) pipe=create_native_pipeline('characters',device='cpu',compile_mode='default',cfg_mode='batched') pipe.enable_model_cpu_offload(gpu_id=0) cache={};encode_times={} for p in people: for e in p['expressions']: start=time.perf_counter();cache[e['id']]=encode_text(pipe,e['prompt']).cpu();torch.cuda.synchronize() encode_times[e['id']]=time.perf_counter()-start neg=p['expressions'][0]['negative_prompt'] if neg not in pipe._native_text_cache:pipe._native_text_cache[neg]=encode_text(pipe,neg) dump(STATE/'pilot-state.json',dict(status='generating',ids=ids,preencoded_prompts=len(cache),utc=now())) all_records=[];references={};reference_shas={};archives={} for p in people: name='pilot-'+p['personality_id']+'.tar' archives[p['personality_id']]=tarfile.open(BUFFER/(name+'.partial'),'w',format=tarfile.USTAR_FORMAT) # Inspect every age group's neutral image early, then prioritize visible # expressions. Only accepted neutral pixels are reused from the first pilot. order=['neutral','angry','joyful','sad','afraid','disgust','pain','annoyed','grateful','playful'] first_generated=True for emotion in order: for p in people: ident=p['personality_id'];e=next(x for x in p['expressions'] if x['emotion']==emotion) name='pilot-'+ident+'.tar';tar=archives[ident];bootstrap=None if ident=='person-0002' and emotion=='neutral': prior=list((BUFFER/'pilot-first-pass').glob('pilot-person-0002.tar*')) if prior: with tarfile.open(prior[0],'r:') as old: png_member=old.next();png=old.extractfile(png_member).read() json_member=old.next();bootstrap=json.loads(old.extractfile(json_member).read()) assert bootstrap['id']==e['id'] and bootstrap['prompt']==e['prompt'] with Image.open(io.BytesIO(png)) as im:image=im.convert('RGB') m=bootstrap if bootstrap is None: torch.cuda.synchronize();start=time.perf_counter() image=generate_native_image(pipe,prompt_embeds=cache[e['id']].to(pipe._execution_device),seed=e['seed'], steps=50,guidance=4.,negative_prompt=e['negative_prompt'],image=references.get(ident) if emotion!='neutral' else None, width=832,height=1248)[0] torch.cuda.synchronize();seconds=time.perf_counter()-start encoded=io.BytesIO();image.save(encoded,format='PNG',compress_level=4);png=encoded.getvalue() m=metadata(p,e,png,reference_shas.get(ident,''),seconds,encode_times[e['id']],name,'shared-GPU7-CPU-offload-preencoded-pilot') m['runtime']['includes_compile_warmup']=first_generated first_generated=False add_member(tar,e['id']+'.png',png);add_member(tar,e['id']+'.json',json.dumps(m,indent=2,ensure_ascii=False).encode());add_member(tar,e['id']+'.txt',e['prompt'].encode()) tar.fileobj.flush() all_records.append(m) if emotion=='neutral': references[ident]=image.resize((416,624),Image.Resampling.BICUBIC);reference_shas[ident]=hashlib.sha256(png).hexdigest() print(json.dumps(dict(id=e['id'],seconds=round(m['runtime']['generation_seconds'],2))),flush=True) dump(STATE/'pilot-records.json',all_records) dump(STATE/'pilot-state.json',dict(status='generating',images_completed=len(all_records),latest=e['id'],utc=now())) for ident,tar in archives.items(): tar.close();name='pilot-'+ident+'.tar';(BUFFER/(name+'.partial')).replace(BUFFER/name) dump(STATE/'pilot-state.json',dict(status='complete',images_completed=len(all_records),ids=ids,utc=now())) if __name__=='__main__':main()