Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Make decoder-only models to be able to generate with inputs_embeds #172

Merged
merged 5 commits into from
Mar 30, 2023
Merged

Make decoder-only models to be able to generate with inputs_embeds #172

merged 5 commits into from
Mar 30, 2023

Conversation

ingyuseong
Copy link
Contributor

Title

Make decoder-only models to be able to generate with inputs_embeds

Description

Synchronize GPT2 code with Hugging Face transformers—GPT2 can generate with input_embeds.

Accepting .generate() calls with inputs_embeds on decoder-only models is a long-standing request (huggingface/transformers#6535) -- see huggingface/transformers#6535 (comment) particular and its reacts.

It has to be added on a per-model basis, and this PR adds the necessary changes for GPT2. Other models will throw an informative exception if the user passes inputs_embeds, asking them to check this PR and implement the same pattern on the model they want to use it with 🤗

Please note that it is still expected that the user passes input_ids, i.e.

outputs = model.generate(input_ids, inputs_embeds=inputs_embeds)

This is because decoder-only models expect the prompt to be present in the output, and this is the only way to preserve it! input_ids can also be omitted and, in that case, the output won't contain the prompt.

For more details, please check out this PR.

@hyunwoongko
Copy link
Member

Code looks good. could you run precommit run --all-files? you can install precommit using pip install precommit. it's our formatting tool.

@hyunwoongko hyunwoongko merged commit fc2175f into main Mar 30, 2023
@ingyuseong ingyuseong deleted the feature/decoder_only_input_embeds branch April 4, 2023 23:43
yhna940 added a commit to yhna940/oslo that referenced this pull request Apr 18, 2023
@yhna940 @bzantium
@tree-park @ingyuseong @jinwonkim93 @KKIEEK
Merge (#1)
1987953 
* import ParallelMode (EleutherAI#166)

## fix typo on tensor parallel tutorial

- `from oslo import ParallelContext, ParallelMode`

* [Fix] zero param check (EleutherAI#164)

## Title

- [Fix] zero param check

## Description

- ZeRO checks the redundancy of parameters to calculate the norm. There
is a minor bug in checking the TP and needs to be fixed.

## Linked Issues

- N/A

* [Fix] zero optimizer w/ tensor parallel test (EleutherAI#167)

## Title

- [Fix] zero optimizer w/ tensor parallel test

## Description

- ZeRO was not running in tensor parallel mode, so I fixed this by
switching to a model from `transformers`.

## Linked Issues

- N/A

* Add restarting model from saved model and fix bug (EleutherAI#171)

## Description

- load a model
- start training again from a saved point
- fix bug that training_arg not saved with nccl error. It was because of
parallel_context, and it was removed before saving training_arg and
re-attached again
- test load and restart with oslo TP

* Make decoder-only models to be able to generate with `inputs_embeds` (EleutherAI#172)

## Title
Make decoder-only models to be able to generate with `inputs_embeds`

## Description
Synchronize GPT2 code with Hugging Face transformers—GPT2 can generate
with `input_embeds`.

>Accepting `.generate()` calls with `inputs_embeds` on decoder-only
models is a long-standing request
(huggingface/transformers#6535) -- see
huggingface/transformers#6535 (comment)
particular and its reacts.
>
>It has to be added on a per-model basis, and this PR adds the necessary
changes for GPT2. Other models will throw an informative exception if
the user passes `inputs_embeds`, asking them to check this PR and
implement the same pattern on the model they want to use it with 🤗
>
>Please note that it is still expected that the user passes `input_ids`,
i.e.

```python
outputs = model.generate(input_ids, inputs_embeds=inputs_embeds)
```

>This is because decoder-only models expect the prompt to be present in
the output, and this is the only way to preserve it! input_ids can also
be omitted and, in that case, the output won't contain the prompt.

For more details, please check out [this
PR](huggingface/transformers#21405).

* Wrong import in zero (EleutherAI#169)

## Title

Prevent from using torch 2.0

## Description

- Some of feature have changed in torch 2.0. and oslo has dependency on
torch._six which no longer support by torch 2.0.

olso Dependency
-
https://github.com/EleutherAI/oslo/blob/910c789e7f46d2876b964c221d31984b7924974f/oslo/torch/nn/parallel/data_parallel/zero/sharded_optim/_utils.py#L19

other issues
- microsoft/DeepSpeed#2845

## Linked Issues

- resolved #00

* [Fix] Support gradient accumulation for DDP (EleutherAI#173)

## Description

In order to support gradient accumulation, I removed `free_storage`
function that can cause `CUDA error: an illegal memory access was
encountered` in many case. (but this change may lead to an increase in
memory consumption)
What do you guys think about this PR? @nijkah @jinwonkim93

* [Fix] minor bug for single output in _DistributedDataParallel (EleutherAI#177)

## Title

- Fix minor bug for single output in _DistributedDataParallel

## Description

- This PR addresses a minor bug in the `_DistributedDataParallel` class
when handling single output tensors. The changes include:

1. Update the `forward` method in `_DistributedDataParallel` to
correctly handle single output tensors.
2. Add new test cases in
`tests_deprecated/torch/nn/parallel/data_parallel/data_parallel.py` to
ensure the correct behavior for models with various output types (single
tensor, multiple tensors, and dictionary of tensors).

These updates will ensure that the `_DistributedDataParallel` class
works correctly with various output types, providing a more robust
solution for users.

## Linked Issues

- N/A

* [Enhance] Support ViT for TensorParallel (EleutherAI#155)

## Description

I added support for ViT in TensorParallel by appending config to
`_TensorParallelMapping`.
`PatchEmbed` layer in ViT does not have the `weight` parameter unlike
`Embedding` layer, so I replaced the `weight` parameter with a dummy
value to prevent an `AttributeError`.

Any feedback is welcome.

### Memory usage
mode | world_size=1 | world_size=2 | world_size=4 | world_size=8
-|-|-|-|-
1D | 1760MiB | 1126MiB | 789MiB |
2D | | | 589MiB |
2.5D (d=1) | | | 589MiB |
2.5D (d=2) | | | | 586MiB
3D | | | |

### TODO
- [ ] Benchmark with `world_size=8`
- [ ] Refactor slicing patch embedding
- [ ] Fix slicing logic to return the same value as `TensorParallel1D`

<details><summary>code for testing</summary>
<p>

```python
import os
import torch.multiprocessing as mp

import torch
from torch import nn
from torch import optim
import torch.distributed as dist
from transformers import ViTModel, ViTForImageClassification, ViTConfig

import oslo
from oslo.torch.distributed.parallel_context import ParallelContext
from oslo.torch.distributed.parallel_mode import ParallelMode
from oslo.torch.nn.parallel import TensorParallel


def setup(rank, world_size):
    os.environ["MASTER_ADDR"] = "localhost"
    os.environ["MASTER_PORT"] = "12340"
    os.environ["RANK"] = str(rank)
    os.environ["LOCAL_RANK"] = str(rank)
    os.environ["WORLD_SIZE"] = str(world_size)
    os.environ["LOCAL_WORLD_SIZE"] = str(world_size)


def cleanup():
    dist.destroy_process_group()


def train(rank, world_size):
    print(f"Running oslo TP example on rank {rank}.")
    setup(rank, world_size)
    parallel_context = ParallelContext.from_torch(
        tensor_parallel_size=world_size,
        tensor_parallel_mode=ParallelMode.TENSOR_1D,
    )  # TENSOR2D or TENSOR_2P5D

    model = ViTForImageClassification(ViTConfig(num_labels=1000)).to(rank)
    model = TensorParallel(model, parallel_context)
    optimizer = optim.SGD(model.parameters(), lr=1e-4)
    loss_fn = nn.MSELoss()

    oslo.ready(model, parallel_context)

    for _ in range(100):
        model.zero_grad()
        logits = model(pixel_values=torch.ones(8, 3, 224, 224).to(rank)).logits
        labels = torch.ones(8, 1000).to(rank) * 100
        loss = loss_fn(logits, labels)
        loss.backward()
        optimizer.step()
        print(logits)
        print(torch.cuda.max_memory_allocated() / 1024**2)  # MB

    cleanup()


def main(world_size):
    mp.spawn(train, args=(world_size,), nprocs=world_size, join=True)


if __name__ == "__main__":
    main(4)
```

</p>
</details> 

## Linked Issues

Related to EleutherAI#152

---------

Co-authored-by: Minho Ryu <[email protected]>
Co-authored-by: Hansol Park <[email protected]>
Co-authored-by: Ingyu Seong <[email protected]>
Co-authored-by: whooray <[email protected]>
Co-authored-by: Junhwa Song <[email protected]>
dyanos pushed a commit that referenced this pull request Jun 8, 2023
@ingyuseong @dyanos
Make decoder-only models to be able to generate with inputs_embeds (#…
d62bc18 
…172)

## Title
Make decoder-only models to be able to generate with `inputs_embeds`

## Description
Synchronize GPT2 code with Hugging Face transformers—GPT2 can generate
with `input_embeds`.

>Accepting `.generate()` calls with `inputs_embeds` on decoder-only
models is a long-standing request
(huggingface/transformers#6535) -- see
huggingface/transformers#6535 (comment)
particular and its reacts.
>
>It has to be added on a per-model basis, and this PR adds the necessary
changes for GPT2. Other models will throw an informative exception if
the user passes `inputs_embeds`, asking them to check this PR and
implement the same pattern on the model they want to use it with 🤗
>
>Please note that it is still expected that the user passes `input_ids`,
i.e.

```python
outputs = model.generate(input_ids, inputs_embeds=inputs_embeds)
```

>This is because decoder-only models expect the prompt to be present in
the output, and this is the only way to preserve it! input_ids can also
be omitted and, in that case, the output won't contain the prompt.

For more details, please check out [this
PR](huggingface/transformers#21405).
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@hyunwoongko hyunwoongko Awaiting requested review from hyunwoongko hyunwoongko is a code owner

Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
2 participants
@ingyuseong @hyunwoongko