"""CPU smoke test with two 224x224 RGB images and three attention parity checks.""" import hashlib, json, math, sys from pathlib import Path import torch import torch.nn.functional as F from PIL import Image from transformers import Qwen2VLImageProcessor from component import load_vision_model def sha(path): h=hashlib.sha256() with open(path,'rb') as f: for b in iter(lambda:f.read(8*1024*1024),b''):h.update(b) return h.hexdigest() def main(): root=Path(__file__).resolve().parent;images=[Path(x) for x in sys.argv[1:]];assert len(images)>=2 torch.set_num_threads(4);torch.set_grad_enabled(False) model=load_vision_model(root);print('STRICT LOAD OK',flush=True) p=json.loads((root/'preprocessor_config.json').read_text()) processor=Qwen2VLImageProcessor(patch_size=p['patch_size'],temporal_patch_size=p['temporal_patch_size'],merge_size=p['merge_size'],image_mean=p['image_mean'],image_std=p['image_std'],size={'shortest_edge':1024,'longest_edge':224*224}) original=F.scaled_dot_product_attention;rows=[] for path in images: image=Image.open(path).convert('RGB').resize((224,224));data=processor(images=image,return_tensors='pt');grid=data['image_grid_thw'];assert grid.tolist()==[[1,14,14]],grid calls=[0];metrics={} def wrapped(q,k,v,*args,**kwargs): out=original(q,k,v,*args,**kwargs);layer=calls[0];calls[0]+=1 if layer in (0,13,26): mask=kwargs.get('attn_mask',args[0] if args else None);assert mask is None and not kwargs.get('is_causal',False) and kwargs.get('dropout_p',0.0)==0 scale=kwargs.get('scale',None);scale=q.shape[-1]**-.5 if scale is None else scale assert q.shape==(1,16,196,72),q.shape probs=(q@k.transpose(-2,-1)*scale).softmax(-1);calc=probs@v;err=(calc-out).abs().max().item();assert torch.allclose(calc,out,rtol=2e-4,atol=2e-5),err metrics[str(layer)]={'sdpa_max_absolute_error':err,'row_sum_max_error':float((probs.sum(-1)-1).abs().max()),'mean_normalized_entropy':float((-(probs*probs.clamp_min(1e-30).log()).sum(-1)).mean()/math.log(196))} return out try: F.scaled_dot_product_attention=wrapped with torch.inference_mode():out=model(data['pixel_values'],grid_thw=grid) finally:F.scaled_dot_product_attention=original assert calls[0]==27 and len(metrics)==3;assert out.last_hidden_state.shape==(196,1152) and out.pooler_output.shape==(49,4096) assert torch.isfinite(out.last_hidden_state).all() and torch.isfinite(out.pooler_output).all() rows.append({'image':path.name,'image_sha256':sha(path),'grid_thw':grid.tolist(),'last_hidden_state_shape':list(out.last_hidden_state.shape),'pooler_output_shape':list(out.pooler_output.shape),'pooler_mean':float(out.pooler_output.mean()),'attention_metrics':metrics}) print('TESTED',path.name,flush=True) assert abs(rows[0]['pooler_mean']-rows[1]['pooler_mean'])>1e-8 import transformers result={'passed':True,'strict_load':True,'model_sha256':sha(root/'model.safetensors'),'device':'cpu','dtype':'float32','torch':torch.__version__,'transformers':transformers.__version__,'processor':'Qwen2VLImageProcessor with source normalization and patch settings; testing capped at 224x224','attention_method':'Compare explicit softmax(QK^T/sqrt(d))V to native SDPA after rotary embeddings, layers 0,13,26','images':rows} (root/'validation.json').write_text(json.dumps(result,indent=2)+'\n');print('VALIDATION PASSED',flush=True) if __name__=='__main__':main()