[TorchAcc][Experimental] Integrate TorchAcc. #647
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| NNODES=4 \ | ||
| NPROC_PER_NODE=8 \ | ||
| swift sft \ | ||
| --model_type qwen-72b-chat \ |
| --gradient_accumulation_steps 1 \ | ||
| --gradient_checkpointing no \ | ||
| --tuner_backend 'peft' \ | ||
| --eval_steps 2000000 \ |
| --save_steps 2000000 \ | ||
| --logging_steps 10 \ | ||
| --preprocess_num_proc 1 \ | ||
| --dataloader_num_workers 0 \ |
| @@ -0,0 +1,31 @@ | |||
| # Experimental environment: 4 * A800 | |||
| # wrapper the model and make these properties wrong. | ||
| label_names = find_labels(model) | ||
| return_loss = can_return_loss(model) | ||
| model = ta.patch_qwen_model(model) |
There was a problem hiding this comment.
目前仅支持qwen吗，这个方法名是否有点特异化了
swift/llm/sft.py
Outdated
| sft_main = get_main(SftArguments, llm_sft) | ||
| def get_sft_main(args, llm): | ||
| if use_torchacc(): | ||
| logger.warning('TorchAcc is currently only available internally.') |
There was a problem hiding this comment.
internally建议改为更具体的场景，否则用户会疑问什么是内部场景
swift/llm/utils/template.py
Outdated
| return [max_length // 4 * (i + 1) for i in range(4)] | ||
|
|
||
|
|
||
| def _get_bucket(bucket_sizes, data_length): |
There was a problem hiding this comment.
建议给torchacc单独建一个py文件
| @@ -428,6 +453,32 @@ def data_collator(self, | |||
| loss_scale, batch_first=True, padding_value=0.) | |||
| labels = pad_sequence(labels, batch_first=True, padding_value=-100) | |||
|
|
|||
| if use_torchacc(): | |||
| rank, _, world_size, _ = get_dist_setting() | |||
There was a problem hiding this comment.
封装成方法，单独放入torchacc单独的py中，以免用户阅读这里的代码的时候有疑问
swift/trainers/mixin.py
Outdated
| if not use_torchacc(): | ||
| return super()._save_tpu(output_dir) | ||
|
|
||
| import torch_xla.core.xla_model as xm |
There was a problem hiding this comment.
同样，这里也建议封装单独的方法，放入单独的py中
swift/trainers/trainers.py
Outdated
| return super().get_train_dataloader() | ||
| else: | ||
| # patch skip_first_batches for customized dataloader. | ||
| def acc_skip_first_batches(dataloader, num_batches=0): |
swift/llm/sft.py
Outdated
| @@ -181,8 +203,21 @@ def llm_sft(args: SftArguments) -> Dict[str, Union[str, Any]]: | |||
| if val_dataset is not None: | |||
| val_dataset = LazyLLMDataset(val_dataset, template) | |||
|
|
|||
| bucket_sizes = get_bucket_sizes( | |||
baoleai
force-pushed
the
features/rebase_0401
branch
from
e0fe1d4
to
dd5b360
Compare
baoleai
force-pushed
the
features/rebase_0401
branch
from
dd5b360
to
30ad8c8
Compare
baoleai
force-pushed
the
features/rebase_0401
branch
from
2a849fa
to
48ff067
Compare
tastelikefeet
approved these changes
Apr 9, 2024
tastelikefeet
added a commit
to tastelikefeet/swift
that referenced
this pull request
Apr 10, 2024
* main: update Agent best practice with Modelscope-Agent (modelscope#676) [TorchAcc][Experimental] Integrate TorchAcc. (modelscope#647)
4 tasks
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
PR type
PR information
TorchAcc is a framework developed by Alibaba PAI to accelerate PyTorch model training, providing computational acceleration based on compilation optimization and distributed strategies such as FSDP and TP+SP. This PR uses TorchAcc to accelerate the training of Swift SFT LoRA and Full scenarios and provides examples of qwen-72b-chat. Users can enable TorchAcc acceleration by setting
export USE_TORCHACC=1. Currently, this feature is still in the experimental stage and only available internally.Experiment results
Test with 4*80G A100 on qwen-72b-chat lora with script:
sh examples/pytorch/llm/scripts/qwen_72b_chat/torchacc/lora_fsdp_sft.sh{"eval_loss": 0.35520425, "eval_acc": 0.87996558, "eval_runtime": 753.7452, "eval_samples_per_second": 0.362, "eval_steps_per_second": 0.004, "epoch": 1.0, "global_step": 1123}
{"train_runtime": 13049.8116, "train_samples_per_second": 2.065, "train_steps_per_second": 0.086, "total_flos": 0.0, "train_loss": 0.36302515, "epoch": 1.0, "global_step": 1123}