| 249 | |
| 250 | |
| 251 | class NativeScalerWithGradNormCount: |
| 252 | state_dict_key = "amp_scaler" |
| 253 | |
| 254 | def __init__(self): |
| 255 | self._scaler = torch.cuda.amp.GradScaler() |
| 256 | |
| 257 | def __call__(self, loss, optimizer, clip_grad=None, parameters=None, create_graph=False, update_grad=True): |
| 258 | self._scaler.scale(loss).backward(create_graph=create_graph) |
| 259 | if update_grad: |
| 260 | if clip_grad is not None: |
| 261 | assert parameters is not None |
| 262 | self._scaler.unscale_(optimizer) # unscale the gradients of optimizer's assigned params in-place |
| 263 | norm = torch.nn.utils.clip_grad_norm_(parameters, clip_grad) |
| 264 | else: |
| 265 | self._scaler.unscale_(optimizer) |
| 266 | norm = get_grad_norm_(parameters) |
| 267 | self._scaler.step(optimizer) |
| 268 | self._scaler.update() |
| 269 | else: |
| 270 | norm = None |
| 271 | return norm |
| 272 | |
| 273 | def state_dict(self): |
| 274 | return self._scaler.state_dict() |
| 275 | |
| 276 | def load_state_dict(self, state_dict): |
| 277 | self._scaler.load_state_dict(state_dict) |
| 278 | |
| 279 | |
| 280 | def get_grad_norm_(parameters, norm_type: float = 2.0) -> torch.Tensor: |
nothing calls this directly
no outgoing calls
no test coverage detected