OneAlign / model_runtime.py
loveaa's picture
Deploy Home Pulse: grounded setup, isolated workspaces and shared HF visual inference
90d354b verified
Raw History Blame Contribute Delete
3.16 kB
"""HF-hosted imajev, using the same adapter revision as the verified local app."""
from pathlib import Path
import base64
import io
import json
import os
import sys
from time import perf_counter
ROOT = Path(__file__).resolve().parent
for folder in (ROOT/'upstream/src', ROOT/'upstream/scripts', ROOT/'upstream/scripts/playground'):
sys.path.insert(0, str(folder))
def load_model():
from huggingface_hub import snapshot_download
from server import TorchBackend
from vision_decision.calibration import TemperatureCalibrator
base = snapshot_download('Qwen/Qwen3.5-4B', revision='851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a', allow_patterns=['*.json','*.safetensors','*.txt','*.jinja'], token=False)
adapter = snapshot_download('mohit67890/imajev-4b', revision='f8d8234cebc6c99065c07731e59716dc0a6e27ab', allow_patterns=['*.json','*.safetensors'], ignore_patterns=['mlx/*','assets/*'], token=False)
runtime = Path(os.getenv('HOME_PULSE_DATA_DIR', '/tmp/home-pulse'))
runtime.mkdir(parents=True, exist_ok=True)
bundle = runtime/'model-bundle.json'
bundle.write_text(json.dumps({'path':base}))
# ZeroGPU emulates CUDA at module startup and materializes it inside spaces.GPU.
backend = TorchBackend(bundle=bundle, adapter=Path(adapter), device='cuda', rotations=1, max_input_tokens=4096, fast=False, merge_lora=False)
backend.model = 'imajev-4b-hf-bf16'
backend.engine.processor._vision()
calibration = TemperatureCalibrator.load(Path(adapter)/'calibration.json')
return backend, calibration
def predict(backend, calibration, payload):
from PIL import Image
import torch
from server import compile_payload
from vision_decision.jev_api import to_response
start=perf_counter()
supplied=payload.get('images',[])
if not 1<=len(supplied)<=2: raise ValueError('Provide one or two images')
images=[]
for value in supplied:
if not isinstance(value,str) or not value.startswith('data:image/') or ';base64,' not in value: raise ValueError('Invalid image')
blob=base64.b64decode(value.split(',',1)[1],validate=True)
if len(blob)>3_000_000: raise ValueError('Image too large')
with Image.open(io.BytesIO(blob)) as original:
if original.width*original.height>16_000_000: raise ValueError('Image dimensions too large')
image=original.convert('RGB');image.thumbnail((640,640));images.append(image)
if not 1<=len(payload.get('questions',{}))<=4: raise ValueError('Provide 1–4 questions')
request, plan=compile_payload({k:v for k,v in payload.items() if k not in ('images','model','thinking')},backend.max_options)
with torch.inference_mode(): results,usage=backend.score(images,request)
torch.cuda.synchronize()
results=[calibration.calibrate_result(result,field.type,len(result.scores)-1,image=True,photo_only=not request.state) for field,result in zip(request.fields,results)]
response=to_response(request,results,model=backend.model,plan=plan)
response['usage']={**usage,'total_ms':round((perf_counter()-start)*1000,1),'processed_image_sizes':[list(im.size) for im in images]}
return response