"""Bounded real-mesh tensor-operator probe, NOT a model benchmark or training run.""" import argparse import json import os from pathlib import Path import resource import sys import time import numpy as np import torch def main(): p=argparse.ArgumentParser();p.add_argument("sample",type=Path);p.add_argument("output",type=Path);a=p.parse_args() torch.set_num_threads(2);torch.manual_seed(1729) if not torch.cuda.is_available():raise RuntimeError("CUDA unavailable") if torch.cuda.get_device_capability()!=(7,0):raise RuntimeError("Expected V100 sm_70") data=np.load(a.sample,allow_pickle=False);n=len(data["points_raw"]) edges=torch.as_tensor(data["edges"],device="cuda",dtype=torch.long) src=torch.cat((edges[:,0],edges[:,1]));dst=torch.cat((edges[:,1],edges[:,0])) x=torch.randn(n,96,device="cuda") w=(torch.randn(96,96,device="cuda")/96**.5).requires_grad_() degree=torch.bincount(dst,minlength=n).float().clamp_min(1)[:,None] volumes=torch.as_tensor(data["node_volume_raw"],device="cuda",dtype=torch.float32) volumes=volumes/volumes.sum() def operation(amp): with torch.autocast("cuda",dtype=torch.float16,enabled=amp): h=x@w summed=torch.zeros(n,96,device="cuda",dtype=torch.float32) summed.index_add_(0,dst,h.float()[src]) local=summed/degree global_mean=(local*volumes[:,None]).sum(0) return local,global_mean torch.cuda.reset_peak_memory_stats() reference,ref_global=operation(False) candidate,amp_global=operation(True) error=(torch.linalg.vector_norm(candidate-reference)/torch.linalg.vector_norm(reference)).item() if not torch.isfinite(candidate).all() or error>0.01:raise ValueError(f"AMP comparison failed {error}") del reference,candidate,ref_global,amp_global timings=[] for i in range(25): torch.cuda.synchronize();start=time.perf_counter() local,pooled=operation(True) loss=local.square().mean()+pooled.square().mean();loss.backward() if not torch.isfinite(w.grad).all():raise ValueError("Nonfinite gradient") w.grad=None del local,pooled,loss torch.cuda.synchronize() if i>=5:timings.append(time.perf_counter()-start) result=dict(status="pass",purpose="operator compatibility/AMP/autograd only; no optimizer, labels, model or training", python=sys.version,executable=sys.executable,torch=torch.__version__,cuda=torch.version.cuda, gpu=torch.cuda.get_device_name(),capability=list(torch.cuda.get_device_capability()), physical_gpu=os.environ.get("CUDA_VISIBLE_DEVICES"),nodes=n,undirected_edges=len(edges),width=96, warmup_iterations=5,measured_iterations=20,fp16_vs_fp32_relative_l2=error, forward_backward_seconds_p50=float(np.median(timings)),forward_backward_seconds_p90=float(np.quantile(timings,.9)), peak_allocated_bytes=torch.cuda.max_memory_allocated(),peak_reserved_bytes=torch.cuda.max_memory_reserved(), peak_rss_kib=resource.getrusage(resource.RUSAGE_SELF).ru_maxrss) a.output.write_text(json.dumps(result,indent=2)+"\n");print(json.dumps(result,indent=2)) if __name__=="__main__":main()