| """Stable, name-ordered FP32 gradient mean independent of DDP bucket rebuilds.""" |
| import hashlib |
| import torch |
| import torch.distributed as dist |
|
|
| def tensor_hash(value): |
| raw=value.detach().cpu().contiguous().reshape(-1).view(torch.uint8).numpy().tobytes() |
| return hashlib.sha256(raw).hexdigest() |
|
|
| def gradient_mean(model): |
| selected=sorted((n,p) for n,p in model.named_parameters() if p.requires_grad) |
| if not selected:return |
| flat=torch.cat([(p.grad if p.grad is not None else torch.zeros_like(p)).detach().float().reshape(-1) for _,p in selected]) |
| if dist.is_initialized(): |
| dist.all_reduce(flat,op=dist.ReduceOp.SUM);flat.div_(dist.get_world_size()) |
| offset=0 |
| for _,p in selected: |
| value=flat[offset:offset+p.numel()].view_as(p).to(p.dtype) |
| if p.grad is None:p.grad=value.clone() |
| else:p.grad.copy_(value) |
| offset+=p.numel() |
|
|
| def gradients_hash(model): |
| return tensor_hash(torch.cat([p.grad.detach().float().reshape(-1) for n,p in sorted(model.named_parameters()) if p.requires_grad and p.grad is not None])) |
|
|
| def tree_hash(value): |
| h=hashlib.sha256() |
| def visit(v): |
| if isinstance(v,torch.Tensor):h.update(str((str(v.dtype),tuple(v.shape))).encode());h.update(tensor_hash(v).encode()) |
| elif isinstance(v,dict): |
| for k in sorted(v,key=str):h.update(str(k).encode());visit(v[k]) |
| elif isinstance(v,(list,tuple)): |
| for x in v:visit(x) |
| else:h.update(repr(v).encode()) |
| visit(value);return h.hexdigest() |
|
|
