|
| 1 | +# Copyright 2025 The HuggingFace Inc. team. All rights reserved. |
| 2 | +# |
| 3 | +# Licensed under the Apache License, Version 2.0 (the "License"); |
| 4 | +# you may not use this file except in compliance with the License. |
| 5 | +# You may obtain a copy of the License at |
| 6 | +# |
| 7 | +# http://www.apache.org/licenses/LICENSE-2.0 |
| 8 | +# |
| 9 | +# Unless required by applicable law or agreed to in writing, software |
| 10 | +# distributed under the License is distributed on an "AS IS" BASIS, |
| 11 | +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 12 | +# See the License for the specific language governing permissions and |
| 13 | +# limitations under the License. |
| 14 | + |
| 15 | +""" |
| 16 | +This script tests to ensure that `accelerate` performs at the same level as raw `torchao`. |
| 17 | +
|
| 18 | +This particular script verifies this for DDP training. |
| 19 | +""" |
| 20 | + |
| 21 | +from functools import partial |
| 22 | + |
| 23 | +import evaluate |
| 24 | +import torch |
| 25 | +from fp8_utils import get_training_utilities |
| 26 | +from torch.nn.parallel import DistributedDataParallel as DDP |
| 27 | +from torchao.float8 import convert_to_float8_training |
| 28 | + |
| 29 | +from accelerate import Accelerator |
| 30 | +from accelerate.state import AcceleratorState |
| 31 | +from accelerate.utils import AORecipeKwargs, set_seed |
| 32 | + |
| 33 | + |
| 34 | +MODEL_NAME = "bert-base-cased" |
| 35 | +METRIC = evaluate.load("glue", "mrpc") |
| 36 | + |
| 37 | + |
| 38 | +def evaluate_model(model, dataloader, metric, accelerator=None): |
| 39 | + "Turns model to .eval(), runs dataloader, calculates metric, then turns eval back on" |
| 40 | + model.eval() |
| 41 | + for step, batch in enumerate(dataloader): |
| 42 | + with torch.no_grad(): |
| 43 | + outputs = model(**batch) |
| 44 | + predictions = outputs.logits.argmax(dim=-1) |
| 45 | + references = batch["labels"] |
| 46 | + if accelerator is not None and accelerator.num_processes > 1: |
| 47 | + predictions, references = accelerator.gather_for_metrics((predictions, references)) |
| 48 | + metric.add_batch(predictions=predictions, references=references) |
| 49 | + return metric.compute() |
| 50 | + |
| 51 | + |
| 52 | +def filter_linear_layers(module, fqn, first_layer_name=None, last_layer_name=None): |
| 53 | + if isinstance(module, torch.nn.Linear): |
| 54 | + if module.in_features % 16 != 0 or module.out_features % 16 != 0: |
| 55 | + return False |
| 56 | + # For stability reasons, we skip the first and last linear layers |
| 57 | + # Otherwise can lead to the model not training or converging properly |
| 58 | + if fqn in (first_layer_name, last_layer_name): |
| 59 | + return False |
| 60 | + return True |
| 61 | + |
| 62 | + |
| 63 | +def train_baseline(): |
| 64 | + set_seed(42) |
| 65 | + model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = get_training_utilities(MODEL_NAME) |
| 66 | + first_linear = None |
| 67 | + last_linear = None |
| 68 | + for name, module in model.named_modules(): |
| 69 | + if isinstance(module, torch.nn.Linear): |
| 70 | + if first_linear is None: |
| 71 | + first_linear = name |
| 72 | + last_linear = name |
| 73 | + func = partial(filter_linear_layers, first_layer_name=first_linear, last_layer_name=last_linear) |
| 74 | + accelerator = Accelerator() |
| 75 | + device = accelerator.device |
| 76 | + model.to(device) |
| 77 | + |
| 78 | + convert_to_float8_training(model, module_filter_fn=func) |
| 79 | + |
| 80 | + # Convert the model to DDP |
| 81 | + device_ids, output_device = [accelerator.local_process_index], accelerator.local_process_index |
| 82 | + model = DDP(model, device_ids=device_ids, output_device=output_device) |
| 83 | + |
| 84 | + base_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator) |
| 85 | + model.train() |
| 86 | + |
| 87 | + for batch in train_dataloader: |
| 88 | + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): |
| 89 | + batch = batch.to(device) |
| 90 | + outputs = model(**batch) |
| 91 | + loss = outputs.loss |
| 92 | + loss.backward() |
| 93 | + optimizer.step() |
| 94 | + optimizer.zero_grad() |
| 95 | + lr_scheduler.step() |
| 96 | + |
| 97 | + trained_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator) |
| 98 | + |
| 99 | + assert ( |
| 100 | + trained_model_results["accuracy"] > base_model_results["accuracy"] |
| 101 | + ), f'Accuracy should be higher for the trained model: {trained_model_results["accuracy"]} > {base_model_results["accuracy"]}' |
| 102 | + assert ( |
| 103 | + trained_model_results["f1"] > base_model_results["f1"] |
| 104 | + ), f'F1 score should be higher for the trained model: {trained_model_results["f1"]} > {base_model_results["f1"]}' |
| 105 | + |
| 106 | + return base_model_results, trained_model_results |
| 107 | + |
| 108 | + |
| 109 | +def train_integration(): |
| 110 | + AcceleratorState()._reset_state(True) |
| 111 | + accelerator = Accelerator(mixed_precision="fp8", kwargs_handlers=[AORecipeKwargs()]) |
| 112 | + set_seed(42) |
| 113 | + model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = get_training_utilities( |
| 114 | + MODEL_NAME, accelerator=accelerator |
| 115 | + ) |
| 116 | + |
| 117 | + model, optimizer = accelerator.prepare(model, optimizer) |
| 118 | + base_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator) |
| 119 | + model.train() |
| 120 | + |
| 121 | + for batch in train_dataloader: |
| 122 | + outputs = model(**batch) |
| 123 | + loss = outputs.loss |
| 124 | + accelerator.backward(loss) |
| 125 | + optimizer.step() |
| 126 | + optimizer.zero_grad() |
| 127 | + lr_scheduler.step() |
| 128 | + |
| 129 | + trained_model_results = evaluate_model(model, eval_dataloader, METRIC, accelerator=accelerator) |
| 130 | + |
| 131 | + assert ( |
| 132 | + trained_model_results["accuracy"] > base_model_results["accuracy"] |
| 133 | + ), f'Accuracy should be higher for the trained model: {trained_model_results["accuracy"]} > {base_model_results["accuracy"]}' |
| 134 | + assert ( |
| 135 | + trained_model_results["f1"] > base_model_results["f1"] |
| 136 | + ), f'F1 score should be higher for the trained model: {trained_model_results["f1"]} > {base_model_results["f1"]}' |
| 137 | + |
| 138 | + return base_model_results, trained_model_results |
| 139 | + |
| 140 | + |
| 141 | +if __name__ == "__main__": |
| 142 | + baseline_not_trained, baseline_trained = train_baseline() |
| 143 | + accelerator_not_trained, accelerator_trained = train_integration() |
| 144 | + |
| 145 | + assert ( |
| 146 | + baseline_not_trained["accuracy"] == accelerator_not_trained["accuracy"] |
| 147 | + ), f'Accuracy should be the same for the baseline and accelerator: {baseline_not_trained["accuracy"]} == {accelerator_not_trained["accuracy"]}' |
| 148 | + assert ( |
| 149 | + baseline_not_trained["f1"] == accelerator_not_trained["f1"] |
| 150 | + ), f'F1 score should be the same for the baseline and accelerator: {baseline_not_trained["f1"]} == {accelerator_not_trained["f1"]}' |
| 151 | + assert ( |
| 152 | + baseline_trained["accuracy"] == accelerator_trained["accuracy"] |
| 153 | + ), f'Accuracy should be the same for the baseline and accelerator: {baseline_trained["accuracy"]} == {accelerator_trained["accuracy"]}' |
| 154 | + assert ( |
| 155 | + baseline_trained["f1"] == accelerator_trained["f1"] |
| 156 | + ), f'F1 score should be the same for the baseline and accelerator: {baseline_trained["f1"]} == {accelerator_trained["f1"]}' |
| 157 | + |
| 158 | + torch.distributed.destroy_process_group() |
0 commit comments