Spaces:
Running on Zero
Running on Zero
| # SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved. | |
| # SPDX-License-Identifier: Apache-2.0 | |
| # | |
| # Licensed under the Apache License, Version 2.0 (the "License"); | |
| # you may not use this file except in compliance with the License. | |
| # You may obtain a copy of the License at | |
| # | |
| # http://www.apache.org/licenses/LICENSE-2.0 | |
| # | |
| # Unless required by applicable law or agreed to in writing, software | |
| # distributed under the License is distributed on an "AS IS" BASIS, | |
| # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | |
| # See the License for the specific language governing permissions and | |
| # limitations under the License. | |
| import torch | |
| import torch.utils.data | |
| from diffusion.training.utils.checkpointer import MultiRankCheckpointer | |
| from utils import misc | |
| from utils.fsdp_checkpointer import FSDPCheckpointer | |
| from utils.model import Model | |
| from utils.trainer import Trainer as BaseTrainer | |
| class Trainer(BaseTrainer): | |
| def __init__(self, config): | |
| super(Trainer, self).__init__(config) | |
| if config.trainer.distributed_parallelism == "ddp": | |
| self.checkpointer = MultiRankCheckpointer(config.checkpoint, config.job, callbacks=self.callbacks) | |
| elif config.trainer.distributed_parallelism == "fsdp": | |
| self.checkpointer = FSDPCheckpointer(config.checkpoint, config.job, callbacks=self.callbacks) | |
| else: | |
| raise ValueError(f"Unsupported distributed parallelism: {config.trainer.distributed_parallelism}") | |
| def validate( | |
| self, | |
| model: Model, | |
| dataloader_val: torch.utils.data.DataLoader, | |
| iteration: int = 0, | |
| ) -> None: | |
| """Validation without ema_scope — FSDP models manage EMA internally.""" | |
| self.callbacks.on_validation_start(model, dataloader_val, iteration=iteration) | |
| model.eval() | |
| for val_iter, data_batch in enumerate(dataloader_val): | |
| if self.config.trainer.max_val_iter is not None and val_iter >= self.config.trainer.max_val_iter: | |
| break | |
| data_batch = misc.to(data_batch, device="cuda") | |
| self.callbacks.on_validation_step_start(model, data_batch, iteration=iteration) | |
| output_batch, loss = model.validation_step(data_batch, iteration) | |
| self.callbacks.on_validation_step_end( | |
| model, data_batch, output_batch, loss, iteration=iteration | |
| ) | |
| self.callbacks.on_validation_end(model, iteration=iteration) | |