Fix TP embedding layers #152
Assignees
Comments
3 tasks
hyunwoongko
pushed a commit
that referenced
this issue
Apr 18, 2023
## 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 #152
yhna940
added a commit
to yhna940/oslo
that referenced
this issue
Apr 18, 2023
* 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 issue
Jun 8, 2023
## 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 #152
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Describe a TODO feature
https://discord.com/channels/729741769192767510/1012603449910759504/1083785802930192434
Assignees
The text was updated successfully, but these errors were encountered: