"""Bounded GPU equivalence validation, not a behavioral pilot.""" import os import json from pathlib import Path import modal app = modal.App('modelclinic-qwen-validation') volume = modal.Volume.from_name(os.environ.get("MODELCLINIC_MODAL_VOLUME", "modelclinic-cache"), create_if_missing=False) image = (modal.Image.debian_slim(python_version='3.11') .pip_install('torch==2.8.0') .pip_install('transformers==5.20.2','accelerate==1.12.0','huggingface-hub==2.4.0','safetensors==0.8.0') .add_local_file('code/src/interventions.py', '/root/project/interventions.py')) REPOS = [('Qwen/Qwen3.5-0.8B','2fc06364715b967f1860aea9cf38778875588b17','base','04b1c301231dd422b8860db31311ab2721511346a32cb1e079c4c4e5f1fe4696'), ('RikiyaT/Qwen3.5-0.8B-weight-space-programming','d1671a178705bcab7644806d23c535517f35723b','edited','fd1e42acdd118ac1b300d37500fd2885e75f5a09de3a33010d802bf7f737b553')] @app.function(image=image, volumes={'/cache':volume}, cpu=2, memory=8192, timeout=900, scaledown_window=2) def prepare(): import hashlib,time from huggingface_hub import snapshot_download start=time.time(); result=[] for repo,rev,name,expected in REPOS: dst=f'/cache/modelclinic/{name}-{rev}' snapshot_download(repo_id=repo,revision=rev,local_dir=dst,allow_patterns=['*.json','*.jinja','*.txt','*.safetensors'],token=False,max_workers=4) p=Path(dst)/'model.safetensors-00001-of-00001.safetensors';h=hashlib.sha256() with p.open('rb') as f: for chunk in iter(lambda:f.read(4*1024*1024),b''):h.update(chunk) digest=h.hexdigest() if digest != expected: raise ValueError(f'{name} checkpoint SHA-256 mismatch') result.append({'repo':repo,'revision':rev,'sha256':digest,'size':p.stat().st_size,'path':dst}) volume.commit() return {'checkpoints':result,'seconds':time.time()-start} @app.function(image=image, gpu=['L4','A10'], volumes={'/cache':volume}, cpu=2, memory=8192, timeout=600, scaledown_window=2) def validate(): import sys,time,gc,torch from contextlib import ExitStack from transformers import AutoTokenizer,Qwen3_5ForConditionalGeneration sys.path.insert(0,'/root/project') from interventions import additive_hook,replacement_hook volume.reload(); torch.manual_seed(123); torch.set_grad_enabled(False) assert torch.cuda.is_available() and torch.cuda.is_bf16_supported() start=time.time(); base=f'/cache/modelclinic/base-{REPOS[0][1]}'; edited=f'/cache/modelclinic/edited-{REPOS[1][1]}' tok=AutoTokenizer.from_pretrained(base,local_files_only=True) prompts=['What is 2 + 2?','Give the next number: 1, 4, 9,','A rectangle has sides 3 and 5. Find its area.'] inputs=[] for prompt in prompts: v=tok.apply_chat_template([{'role':'user','content':prompt}],tokenize=True,add_generation_prompt=True,enable_thinking=True,return_tensors='pt') inputs.append((v['input_ids'] if hasattr(v,'keys') else v).to('cuda')) def load(path): return Qwen3_5ForConditionalGeneration.from_pretrained(path,dtype=torch.bfloat16,local_files_only=True).eval().to('cuda') def layers(model):return [model.model.language_model.layers[i].mlp.down_proj for i in (15,19)] def forward(model,ids): captures=[]; handles=[m.register_forward_hook(lambda _m,_a,out: captures.append(out.detach().float().cpu())) for m in layers(model)] try: with torch.inference_mode(): out=model(input_ids=ids,use_cache=False,output_hidden_states=True) return captures,[v.detach().float().cpu() for v in out.hidden_states],out.logits.detach().float().cpu() finally: for h in handles:h.remove() edited_model=load(edited); edited_weights=[x.weight.detach().cpu().clone() for x in layers(edited_model)] reference=[forward(edited_model,x) for x in inputs] del edited_model;gc.collect();torch.cuda.empty_cache() model=load(base); mods=layers(model); deltas=[(e.float()-m.weight.detach().float().cpu()).to('cuda') for e,m in zip(edited_weights,mods)] def compare(a,b): d=(a-b).abs(); denom=torch.linalg.vector_norm(a).item() return {'max_abs':float(d.max()),'mean_abs':float(d.mean()),'relative_l2':float(torch.linalg.vector_norm(d))/max(denom,1e-30),'exact':bool(torch.equal(a,b))} data=[] for prompt,ids,ref in zip(prompts,inputs,reference): base_out=forward(model,ids) with ExitStack() as s: for m,d in zip(mods,deltas):s.enter_context(replacement_hook(m,delta=d)) rec=forward(model,ids) with ExitStack() as s: for m,d in zip(mods,deltas):s.enter_context(additive_hook(m,'delta',delta=d)) add=forward(model,ids) def summ(x):return {'layers':[compare(a,b) for a,b in zip(ref[0],x[0])],'hidden':[compare(a,b) for a,b in zip(ref[1],x[1])],'logits':compare(ref[2],x[2])} data.append({'prompt':prompt,'input_ids':ids[0].tolist(),'edited_logits_shape':list(ref[2].shape),'base_logits_shape':list(base_out[2].shape),'replacement':summ(rec),'additive':summ(add)}) with ExitStack() as s: for m,d in zip(mods,deltas):s.enter_context(replacement_hook(m,delta=d)) with torch.inference_mode(): generated=model.generate(input_ids=inputs[0],max_new_tokens=4,do_sample=False,use_cache=True,pad_token_id=tok.eos_token_id) tail=generated[0,inputs[0].shape[-1]:].tolist();torch.cuda.synchronize(); free,total=torch.cuda.mem_get_info() return {'hardware':{'gpu':torch.cuda.get_device_name(),'total_vram_bytes':total,'free_vram_bytes':free,'peak_allocated_bytes':torch.cuda.max_memory_allocated(),'torch':torch.__version__,'cuda':torch.version.cuda,'bf16':torch.cuda.is_bf16_supported()},'precision':'bfloat16','model_class':type(model).__name__,'tokenizer_class':type(tok).__name__,'generation':{'max_new_tokens':4,'token_ids':tail,'text':tok.decode(tail)},'prefixes':data,'seconds':time.time()-start} @app.local_entrypoint() def main(): prep=prepare.remote();Path('results/checkpoint_analysis/modal_checkpoint_preparation.json').write_text(json.dumps(prep,indent=2)); print('Pinned checkpoint preparation complete',flush=True) result=validate.remote();Path('results/checkpoint_analysis/modal_qwen_validation.json').write_text(json.dumps(result,indent=2));print('Validation results saved',flush=True)