| import hashlib,json,time |
| from pathlib import Path |
| import torch |
| torch.set_num_threads(4) |
| from safetensors import safe_open |
| from component import load_model |
| r=Path(__file__).parent;checks=[];start=time.monotonic() |
| def check(name,ok,**details): |
| assert ok,(name,details) |
| checks.append({'name':name,'passed':True,**details});print(name,details,flush=True) |
| expected=json.loads((r/'tensor_sha256.json').read_text()) |
| with safe_open(r/'model.safetensors',framework='pt') as f: |
| assert set(f.keys())==set(expected) |
| for k in f.keys(): |
| v=f.get_tensor(k) |
| assert hashlib.sha256(v.contiguous().reshape(-1).view(torch.uint8).numpy().tobytes()).hexdigest()==expected[k],k |
| check('1582_original_tensor_payloads',len(expected)==1582) |
| check('only_active_visual_prefixes',all(k.startswith(('model.vision_tower.','model.mm_projector.','model.diffusion_connector.','model.dit.','model.vae.')) for k in expected)) |
| m=load_model(r) |
| check('native_1582entry_strict_load',len(m.state_dict())==1582) |
| check('original_DCAE_scaling',m.model.vae.config.scaling_factor==0.41407) |
| with torch.inference_mode(): |
| for size in [128,1024]: |
| pixels=torch.linspace(0,1,3*size*size).reshape(1,3,size,size) |
| features=m.model.vision_tower(pixels) |
| check(f'{size}px_FastViT_geometry',tuple(features.shape)==(1,(size//64)**2,3072) and torch.isfinite(features).all().item()) |
| direct=m.model.vision_tower.vision_tower(pixels,return_image_embeddings=True)['image_embeddings'].flatten(2).transpose(1,2) |
| check(f'{size}px_native_features_exact',torch.equal(features,direct)) |
| projected=m.model.mm_projector(features) |
| check(f'{size}px_VLM_projector',tuple(projected.shape)==(1,(size//64)**2,896) and torch.isfinite(projected).all().item()) |
| hidden=tuple(torch.full((1,8,896),i/20) for i in range(5)) |
| cond=m.condition(hidden) |
| check('MCP_preserves_sequence',tuple(cond.shape)==(1,8,2304) and torch.isfinite(cond).all().item()) |
| changed=(hidden[0]+100,)+hidden[1:] |
| check('MCP_selects_final_four_layers',torch.equal(cond,m.condition(changed))) |
| fusion=m.model.diffusion_connector.fusion |
| weights=(fusion.layer_weights/fusion.temperature).softmax(0) |
| fused=sum(weights[i]*hidden[-4+i] for i in range(4)) |
| direct=fusion.input_compress(fused) |
| for refiner in fusion.sequence_refiners:direct=refiner(direct) |
| direct=fusion.output_proj(direct) |
| check('native_weighted_fusion_exact',torch.equal(cond,direct)) |
| for h,w in [(64,64),(64,96)]: |
| rgb=torch.linspace(-1,1,3*h*w).reshape(1,3,h,w) |
| z=m.encode_latent(rgb) |
| check(f'{h}x{w}_native_encode_exact',tuple(z.shape)==(1,32,h//32,w//32) and torch.equal(z,m.model.vae.encode(rgb).latent*0.41407)) |
| decoded=m.decode_latent(z) |
| check(f'{h}x{w}_native_decode_exact',torch.equal(decoded,m.model.vae.decode(z/0.41407).sample)) |
| check(f'{h}x{w}_finite_RGB',tuple(decoded.shape)==tuple(rgb.shape) and torch.isfinite(decoded).all().item(),mse=(decoded-rgb).square().mean().item()) |
| z=torch.linspace(-1,1,32*4*4).reshape(1,32,4,4) |
| flow=m.model.dit(z,timestep=torch.tensor([500.]),encoder_hidden_states=cond,encoder_attention_mask=torch.ones(1,8),return_dict=False)[0] |
| check('complete_28layer_SANA_prediction',tuple(flow.shape)==tuple(z.shape) and torch.isfinite(flow).all().item()) |
| report={'strict_load':True,'source_tensor_bytes_equal':True,'checks':checks,'smoke':{'passed':True,'device':'cpu','dtype':'float32'},'elapsed_seconds':time.monotonic()-start,'limitations':['Original Qwen language model is excluded; MCP/SANA smoke fixtures use synthetic VLM hidden states.','FastViT is checked at 128px and native 1024px; DC-AE reconstruction fixtures are 64px square and rectangular; no downstream quality benchmark is claimed.','Legacy base_model.model.* checkpoint entries are excluded because the pinned official inference constructor and loader use model.*; all selected active visual branches load strictly.','Native code is retained without changing sampler equations or model math. CPU float32 inference does not establish CoreML/MLX/CUDA parity.']} |
| (r/'validation.json').write_text(json.dumps(report,indent=2)+'\n') |
| print('MOBILE VALIDATED',len(checks),flush=True) |
|
|
