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

Bump pytorch to 2.5.0 #3663

Merged
merged 1 commit into from
Oct 21, 2024
Merged

Bump pytorch to 2.5.0 #3663

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
54 changes: 27 additions & 27 deletions .github/workflows/daily.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,11 +20,6 @@ jobs:
strategy:
matrix:
include:
- name: cpu-3.11-2.2
container: mosaicml/pytorch:2.2.1_cpu-python3.11-ubuntu20.04
markers: not daily and (remote or not remote) and not gpu and not doctest
pytest_command: coverage run -m pytest
composer_package_name: mosaicml
- name: cpu-3.11-2.3
container: mosaicml/pytorch:2.3.1_cpu-python3.11-ubuntu20.04
markers: not daily and (remote or not remote) and not gpu and not doctest
Expand All @@ -35,8 +30,13 @@ jobs:
markers: not daily and (remote or not remote) and not gpu and not doctest
pytest_command: coverage run -m pytest
composer_package_name: mosaicml
- name: cpu-3.11-2.4-composer
container: mosaicml/pytorch:2.4.1_cpu-python3.11-ubuntu20.04
- name: cpu-3.11-2.5
container: mosaicml/pytorch:2.5.0_cpu-python3.11-ubuntu20.04
markers: not daily and (remote or not remote) and not gpu and not doctest
pytest_command: coverage run -m pytest
composer_package_name: mosaicml
- name: cpu-3.11-2.5-composer
container: mosaicml/pytorch:2.5.0_cpu-python3.11-ubuntu20.04
markers: not daily and (remote or not remote) and not gpu and not doctest
pytest_command: coverage run -m pytest
composer_package_name: composer
Expand All @@ -45,11 +45,6 @@ jobs:
markers: not daily and (remote or not remote) and not gpu and doctest
pytest_command: coverage run -m pytest tests/test_docs.py
composer_package_name: mosaicml
- name: daily-cpu-3.11-2.2
container: mosaicml/pytorch:2.2.1_cpu-python3.11-ubuntu20.04
markers: daily and (remote or not remote) and not gpu and not doctest
pytest_command: coverage run -m pytest
composer_package_name: mosaicml
- name: daily-cpu-3.11-2.3
container: mosaicml/pytorch:2.3.1_cpu-python3.11-ubuntu20.04
markers: daily and (remote or not remote) and not gpu and not doctest
Expand All @@ -60,13 +55,18 @@ jobs:
markers: daily and (remote or not remote) and not gpu and not doctest
pytest_command: coverage run -m pytest
composer_package_name: mosaicml
- name: daily-cpu-3.11-2.4-composer
container: mosaicml/pytorch:2.4.1_cpu-python3.11-ubuntu20.04
- name: daily-cpu-3.11-2.5
container: mosaicml/pytorch:2.5.0_cpu-python3.11-ubuntu20.04
markers: daily and (remote or not remote) and not gpu and not doctest
pytest_command: coverage run -m pytest
composer_package_name: mosaicml
- name: daily-cpu-3.11-2.5-composer
container: mosaicml/pytorch:2.5.0_cpu-python3.11-ubuntu20.04
markers: daily and (remote or not remote) and not gpu and not doctest
pytest_command: coverage run -m pytest
composer_package_name: composer
- name: daily-cpu-doctest
container: mosaicml/pytorch:2.4.1_cpu-python3.11-ubuntu20.04
container: mosaicml/pytorch:2.5.0_cpu-python3.11-ubuntu20.04
markers: daily and (remote or not remote) and not gpu and doctest
pytest_command: coverage run -m pytest tests/test_docs.py
composer_package_name: mosaicml
Expand Down Expand Up @@ -107,12 +107,6 @@ jobs:
include:
# Unlike CPU tests, we run daily tests together with GPU tests to minimize launch time
# on MCLOUD and not eat up all GPUs at once
- name: "gpu-3.11-2.2-1-gpu"
container: mosaicml/pytorch:2.2.1_cu121-python3.11-ubuntu20.04
markers: "(daily or not daily) and (remote or not remote) and gpu and (doctest or not doctest)"
pytest_command: "coverage run -m pytest"
composer_package_name: "mosaicml"
gpu_num: 1
- name: "gpu-3.11-2.3-1-gpu"
container: mosaicml/pytorch:2.3.1_cu121-python3.11-ubuntu20.04
markers: "(daily or not daily) and (remote or not remote) and gpu and (doctest or not doctest)"
Expand All @@ -125,12 +119,12 @@ jobs:
pytest_command: "coverage run -m pytest"
composer_package_name: "mosaicml"
gpu_num: 1
- name: "gpu-3.11-2.2-2-gpu"
container: mosaicml/pytorch:2.2.1_cu121-python3.11-ubuntu20.04
- name: "gpu-3.11-2.5-1-gpu"
container: mosaicml/pytorch:2.5.0_cu124-python3.11-ubuntu20.04
markers: "(daily or not daily) and (remote or not remote) and gpu and (doctest or not doctest)"
pytest_command: "coverage run -m pytest"
composer_package_name: "mosaicml"
gpu_num: 2
gpu_num: 1
- name: "gpu-3.11-2.3-2-gpu"
container: mosaicml/pytorch:2.3.1_cu121-python3.11-ubuntu20.04
markers: "(daily or not daily) and (remote or not remote) and gpu and (doctest or not doctest)"
Expand All @@ -143,12 +137,12 @@ jobs:
pytest_command: "coverage run -m pytest"
composer_package_name: "mosaicml"
gpu_num: 2
- name: "gpu-3.11-2.2-4-gpu"
container: mosaicml/pytorch:2.2.1_cu121-python3.11-ubuntu20.04
- name: "gpu-3.11-2.5-2-gpu"
container: mosaicml/pytorch:2.5.0_cu124-python3.11-ubuntu20.04
markers: "(daily or not daily) and (remote or not remote) and gpu and (doctest or not doctest)"
pytest_command: "coverage run -m pytest"
composer_package_name: "mosaicml"
gpu_num: 4
gpu_num: 2
- name: "gpu-3.11-2.3-4-gpu"
container: mosaicml/pytorch:2.3.1_cu121-python3.11-ubuntu20.04
markers: "(daily or not daily) and (remote or not remote) and gpu and (doctest or not doctest)"
Expand All @@ -161,6 +155,12 @@ jobs:
pytest_command: "coverage run -m pytest"
composer_package_name: "mosaicml"
gpu_num: 4
- name: "gpu-3.11-2.5-4-gpu"
container: mosaicml/pytorch:2.5.0_cu124-python3.11-ubuntu20.04
markers: "(daily or not daily) and (remote or not remote) and gpu and (doctest or not doctest)"
pytest_command: "coverage run -m pytest"
composer_package_name: "mosaicml"
gpu_num: 4
steps:
- name: Checkout code
uses: actions/checkout@v3
Expand Down
10 changes: 5 additions & 5 deletions .github/workflows/pr-cpu.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,10 +16,6 @@ jobs:
strategy:
matrix:
include:
- name: cpu-3.11-2.2
container: mosaicml/pytorch:2.2.1_cpu-python3.11-ubuntu20.04
markers: not daily and not remote and not gpu and not doctest
pytest_command: coverage run -m pytest
- name: cpu-3.11-2.3
container: mosaicml/pytorch:2.3.1_cpu-python3.11-ubuntu20.04
markers: not daily and not remote and not gpu and not doctest
Expand All @@ -28,8 +24,12 @@ jobs:
container: mosaicml/pytorch:2.4.1_cpu-python3.11-ubuntu20.04
markers: not daily and not remote and not gpu and not doctest
pytest_command: coverage run -m pytest
- name: cpu-3.11-2.5
container: mosaicml/pytorch:2.5.0_cpu-python3.11-ubuntu20.04
markers: not daily and not remote and not gpu and not doctest
pytest_command: coverage run -m pytest
- name: cpu-doctest
container: mosaicml/pytorch:2.4.1_cpu-python3.11-ubuntu20.04
container: mosaicml/pytorch:2.5.0_cpu-python3.11-ubuntu20.04
markers: not daily and not remote and not gpu and doctest
pytest_command: coverage run -m pytest tests/test_docs.py
steps:
Expand Down
14 changes: 7 additions & 7 deletions .github/workflows/pr-gpu.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
name: PR GPU tests
on:
pull_request_target:
pull_request:
workflow_dispatch:
# Cancel old runs when a new commit is pushed to the same branch if not on main
# or dev
Expand All @@ -15,8 +15,8 @@ jobs:
strategy:
matrix:
include:
- name: gpu-3.11-2.4-1
container: mosaicml/pytorch:2.4.1_cu124-python3.11-ubuntu20.04
- name: gpu-3.11-2.5-1
container: mosaicml/pytorch:2.5.0_cu124-python3.11-ubuntu20.04
markers: not daily and not remote and gpu and (doctest or not doctest)
pytest_command: coverage run -m pytest
composer_package_name: mosaicml
Expand Down Expand Up @@ -44,8 +44,8 @@ jobs:
strategy:
matrix:
include:
- name: gpu-3.11-2.4-2
container: mosaicml/pytorch:2.4.1_cu124-python3.11-ubuntu20.04
- name: gpu-3.11-2.5-2
container: mosaicml/pytorch:2.5.0_cu124-python3.11-ubuntu20.04
markers: not daily and not remote and gpu and (doctest or not doctest)
pytest_command: coverage run -m pytest
composer_package_name: mosaicml
Expand Down Expand Up @@ -74,8 +74,8 @@ jobs:
strategy:
matrix:
include:
- name: gpu-3.11-2.4-4
container: mosaicml/pytorch:2.4.1_cu124-python3.11-ubuntu20.04
- name: gpu-3.11-2.5-4
container: mosaicml/pytorch:2.5.0_cu124-python3.11-ubuntu20.04
markers: not daily and not remote and gpu and (doctest or not doctest)
pytest_command: coverage run -m pytest
composer_package_name: mosaicml
Expand Down
2 changes: 1 addition & 1 deletion composer/checkpoint/state_dict.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -88,7 +88,7 @@ def get_model_state_dict(
log.debug('Calling model.state_dict() for non-FSDP model...')
model_state_dict = model.state_dict()
if isinstance(model, DistributedDataParallel):
nn.modules.utils.consume_prefix_in_state_dict_if_present(model_state_dict, 'module.')
nn.modules.utils.consume_prefix_in_state_dict_if_present(model_state_dict, 'module.') # type: ignore

if include_keys is not None:
model_state_dict = _extract_keys_from_state_dict(model_state_dict, include_keys)
Expand Down
2 changes: 1 addition & 1 deletion composer/models/huggingface.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -916,7 +916,7 @@ def write_huggingface_pretrained_from_composer_checkpoint(
peft_config.save_pretrained(str(output_folder))

weights_state_dict = composer_state_dict['state']['model']
torch.nn.modules.utils.consume_prefix_in_state_dict_if_present(weights_state_dict, prefix='model.')
torch.nn.modules.utils.consume_prefix_in_state_dict_if_present(weights_state_dict, prefix='model.') # type: ignore

# NOTE: This only works for default adapter name, not multiple adapters
if peft_config is not None:
Expand Down
57 changes: 56 additions & 1 deletion composer/trainer/_patch_pytorch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -106,7 +106,13 @@ def patch_pytorch():
elif version.parse(torch.__version__) < version.parse('2.4.1'):
# Monkey patch for torch < 2.4.1 ie torch == 2.4.0

# No monkeypatches!
# No monkeypatches besides unshard (below)!
pass

elif version.parse(torch.__version__) < version.parse('2.5.1'):
# Monkey patch for torch < 2.5.1 ie torch == 2.5.0

# No monkeypatches besides unshard (below)!
pass


Expand Down Expand Up @@ -1046,3 +1052,52 @@ def unshard_with_sync(self):
raise RuntimeError('CUDA out of memory encountered on a different rank')
padded_unsharded_flat_param = self._all_gather_flat_param(unsharded_flat_param)
self._use_unsharded_flat_param(padded_unsharded_flat_param)

if version.parse(torch.__version__) >= version.parse('2.5.0') and version.parse(
torch.__version__,
) < version.parse('2.5.1'):

# Save original FlatParamHandle.unshard to revert back to when dropping automicrobatching hooks
from torch.distributed.fsdp._flat_param import FlatParamHandle
original_unshard = FlatParamHandle.unshard

@no_type_check
def unshard_with_sync(self):
"""Run the unshard logic, but with a sync after a :meth:`_alloc_padded_unsharded_flat_param`.

This prevents deadlocks when some ranks OOM after the alloc call and others do not.
This is a patched method from pytorch, meant to be called when automicrobatching
turns on hooks in its search process for the optimal non-OOMing microbatch size.
This includes all-gathering the flat parameter
and switching to using the unsharded flat parameter. If the handle does
not need unsharding, then this only switches to using the unsharded
flat parameter. For ``NO_SHARD``, this is a no-op.
If FSDP is in :meth:`summon_full_params` and the handle uses parameter
mixed precision, then the parameter is forced to full precision.
"""
if not self.needs_unshard():
# Even when not needing an unshard, we should switch to using
# the unsharded flat parameter
unsharded_flat_param = (
self._get_padded_unsharded_flat_param()
if self.uses_sharded_strategy
else self.flat_param
)
self._use_unsharded_flat_param(unsharded_flat_param)
return
unsharded_flat_param = self._alloc_padded_unsharded_flat_param()

# Check if any other rank hit an OOM
found_cuda_oom_tensor = torch.tensor([0], dtype=torch.uint8).to(self.device, non_blocking=True)

dist.all_reduce(found_cuda_oom_tensor, reduce_operation='MAX')
found_cuda_oom = found_cuda_oom_tensor.item()
# Signal current rank is still in batch
all_ranks_finished_tensor = torch.tensor([0], dtype=torch.uint8).to(self.device, non_blocking=True)

dist.all_reduce(all_ranks_finished_tensor, reduce_operation='MIN')

if found_cuda_oom == 1:
raise RuntimeError('CUDA out of memory encountered on a different rank')
padded_unsharded_flat_param = self._all_gather_flat_param(unsharded_flat_param)
self._use_unsharded_flat_param(padded_unsharded_flat_param)
8 changes: 5 additions & 3 deletions composer/trainer/trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2317,9 +2317,11 @@ def fit(
self.state.max_duration = duration + self.state.timestamp.get(duration.unit)

# Raise error if callig fit with SGD
if type(
self.state.optimizers[0],
) == torch.optim.SGD and version.parse(torch.__version__) >= version.parse('2.4.0'):
if (
type(self.state.optimizers[0]) == torch.optim.SGD and
version.parse(torch.__version__) >= version.parse('2.4.0') and
version.parse(torch.__version__) < version.parse('2.5.0')
):
raise ValueError(
'PyTorch 2.4 breaks (distributed) checkpointing with SGD. '
'Please use a different optimizer, e.g. composer.optim.DecoupledSGDW, '
Expand Down
12 changes: 6 additions & 6 deletions docker/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,15 +30,15 @@ To install composer, once inside the image, run `pip install mosaicml`.
<!-- BEGIN_PYTORCH_BUILD_MATRIX -->
| Linux Distro | Flavor | PyTorch Version | CUDA Version | Python Version | Docker Tags |
|----------------|----------|-------------------|---------------------|------------------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| Ubuntu 20.04 | Base | 2.4.1 | 12.4.1 (Infiniband) | 3.11 | `mosaicml/pytorch:latest`, `mosaicml/pytorch:2.4.1_cu124-python3.11-ubuntu20.04` |
| Ubuntu 20.04 | Base | 2.4.1 | 12.4.1 (EFA) | 3.11 | `mosaicml/pytorch:latest-aws`, `mosaicml/pytorch:2.4.1_cu124-python3.11-ubuntu20.04-aws` |
| Ubuntu 20.04 | Base | 2.4.1 | cpu | 3.11 | `mosaicml/pytorch:latest_cpu`, `mosaicml/pytorch:2.4.1_cpu-python3.11-ubuntu20.04` |
| Ubuntu 20.04 | Base | 2.5.0 | 12.4.1 (Infiniband) | 3.11 | `mosaicml/pytorch:latest`, `mosaicml/pytorch:2.5.0_cu124-python3.11-ubuntu20.04` |
| Ubuntu 20.04 | Base | 2.5.0 | 12.4.1 (EFA) | 3.11 | `mosaicml/pytorch:latest-aws`, `mosaicml/pytorch:2.5.0_cu124-python3.11-ubuntu20.04-aws` |
| Ubuntu 20.04 | Base | 2.5.0 | cpu | 3.11 | `mosaicml/pytorch:latest_cpu`, `mosaicml/pytorch:2.5.0_cpu-python3.11-ubuntu20.04` |
| Ubuntu 20.04 | Base | 2.4.1 | 12.4.1 (Infiniband) | 3.11 | `mosaicml/pytorch:2.4.1_cu124-python3.11-ubuntu20.04` |
| Ubuntu 20.04 | Base | 2.4.1 | 12.4.1 (EFA) | 3.11 | `mosaicml/pytorch:2.4.1_cu124-python3.11-ubuntu20.04-aws` |
| Ubuntu 20.04 | Base | 2.4.1 | cpu | 3.11 | `mosaicml/pytorch:2.4.1_cpu-python3.11-ubuntu20.04` |
| Ubuntu 20.04 | Base | 2.3.1 | 12.1.1 (Infiniband) | 3.11 | `mosaicml/pytorch:2.3.1_cu121-python3.11-ubuntu20.04` |
| Ubuntu 20.04 | Base | 2.3.1 | 12.1.1 (EFA) | 3.11 | `mosaicml/pytorch:2.3.1_cu121-python3.11-ubuntu20.04-aws` |
| Ubuntu 20.04 | Base | 2.3.1 | cpu | 3.11 | `mosaicml/pytorch:2.3.1_cpu-python3.11-ubuntu20.04` |
| Ubuntu 20.04 | Base | 2.2.2 | 12.1.1 (Infiniband) | 3.11 | `mosaicml/pytorch:2.2.2_cu121-python3.11-ubuntu20.04` |
| Ubuntu 20.04 | Base | 2.2.2 | 12.1.1 (EFA) | 3.11 | `mosaicml/pytorch:2.2.2_cu121-python3.11-ubuntu20.04-aws` |
| Ubuntu 20.04 | Base | 2.2.2 | cpu | 3.11 | `mosaicml/pytorch:2.2.2_cpu-python3.11-ubuntu20.04` |
<!-- END_PYTORCH_BUILD_MATRIX -->

**Note**: The `mosaicml/pytorch:latest`, `mosaicml/pytorch:latest_cpu`, and `mosaicml/pytorch:latest-aws`
Expand Down
Loading
Loading