Pull Request resolved: https://github.com/pytorch/pytorch/pull/75753
As per the design in https://github.com/pytorch/pytorch/issues/72138,
convert DDP parameters to ReplicatedTensor during its forward pass. Concretely,
this is done as follows:
1) Create a separate `_replicated_tensor_module` which is a copy of self.module
without creating copies of the Tensors themselves.
2) Use `_replicated_tensor_module` instead of `self.module` during the forward
pass.
3) Have a context manager `_ddp_replicated_tensor` to enable this, since
certain edge cases can fail where self.module is changed out of band resulting
in discrepancy between self.module and `_replicated_tensor_module`.
Differential Revision: [D35533736](https://our.internmc.facebook.com/intern/diff/D35533736/)
Approved by: https://github.com/wanchaol, https://github.com/rohan-varma
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/73842
**Overview**
This cleans up the `ZeroRedundancyOptimizer` tests. I apologize for strong formatting changes mixed in with actually-beneficial changes. It was convenient to unify the formatting while doing a deep comb through the full test file.
The main non-formatting changes include:
- Using `parametrize` instead of manually including `for` loops over possible argument values
- Removing the `DEVICE` global variable, which was used only for the `TestZeroRedundancyOptimizerSingleRank` tests, in favor of consistent usage of `self.device` in both `TestZeroRedundancyOptimizerSingleRank` and `TestZeroRedundancyOptimizerDistributed`
- Moving `assert ... == ...` to `self.assertEqual(..., ...)` when the assert is part of the test's correctness
- Removing the `if self.rank >= self.world_size or (torch.cuda.is_available() and torch.cuda.device_count() < 2):` conditional guards in favor of `common_distributed.skip_if_no_gpu` for `TestZeroRedundancyOptimizerDistributed`
- For `TestZeroRedundancyOptimizerDistributed`, `self.device` is `torch.device(self.rank)` if CUDA is available, while `self.world_size` is at least 2, even if `torch.cuda.device_count() == 1`.
- The problematic case is exactly when `torch.cuda.device_count() == 1` but `self.world_size == 2` since then calling `self.device` on rank 1 will error. The existing conditional guard prevented this case for some tests, but it was not used consistently (e.g. `test_multiple_groups()`), which is most likely the reason for the hangs and resulting test flakiness. (From my experience landing the recent ZeRO constructor changes, the Windows environment uses a world size of 2 but only has 1 device available.)
- A more robust solution is to always use the `skip_if_no_gpu` decorator as long as the test uses `self.device` and CUDA is available. This is in line with the recommended SPSD usage of ZeRO.
- Renaming `test_multiple_groups()` to `test_nondefault_process_group()`
- The existing `test_multiple_groups()` was slightly misnamed. Also, it is only nontrivial for a world size of (at least) 4 since it tests using a process group including only even ranks. It was marked as flaky on Windows, and I believe this is because of the world size and `torch.cuda.device_count()` mismatch. Now, the test only uses GPU if there are enough available and falls back to CPU otherwise, which is safe since the test uses Gloo backend.
- There was also a duplicated section, which I was unsure how to non-naively de-duplicate. The top half and bottom half are identical even though they claim to target fitting into the broadcast bucket and not fitting into the broadcast bucket:
1d497114e7/test/distributed/optim/test_zero_redundancy_optimizer.py (L658-L684)
- Changing `_test_zero_model_parallel()` to not use CPU
- This is my own fault, having introduced this inefficiency last summer. It makes more sense to simply designate one of the two GPUs for a process to be its default device rather than routing through CPU.
**Questions**
- How might we limit the runs for `test_ddp_zero_overlap()`? Because it parameterizes over many values, it contributes significantly to the time-to-signal. However, it is an experimental feature, so it is not critical that the tests run every time.
Test Plan: Imported from OSS
Reviewed By: rohan-varma
Differential Revision: D34675709
Pulled By: awgu
fbshipit-source-id: 71ce9ac968fb34415cd65206855b4bb5e67754fb
(cherry picked from commit 34e3dd0a184318ea9f63a1ee20cd14b111af3501)
Summary:
Reland of https://github.com/pytorch/pytorch/pull/72578.
**Overview**
Windows CI was failing due to the multi-rank single-GPU case (see [here](https://github.com/pytorch/pytorch/runs/5204906995?check_suite_focus=true)).
To address this, I
- added `common_distributed.skip_if_no_gpu` for `test_multiple_param_groups()` to ensure that each rank can safely call `to(self.device)` -- this targets the expected SPSD use case where each rank has its own GPU;
- moved `test_constructor()` back to `TestZeroRedundancyOptimizerSingleRank` to check that the multiple parameter group method for construction works even on a single rank.
**Test Plan**
- I checked both tests for CPU, 1 GPU, 2 GPUs, 4 GPUs, and 8 GPUs.
- I added the `ciflow/win` label to run the failing Windows CI test.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/72932
Reviewed By: rohan-varma
Differential Revision: D34281482
Pulled By: awgu
fbshipit-source-id: c4fe604ddd9d2c123c3071249741e6b8a6454b6e
(cherry picked from commit 6bea9bcc63)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/72578
**Overview**
This adds `ZeroRedundancyOptimizer` constructor support for multiple parameter groups (i.e. passing an `iterable` of `dict`s instead of an `iterable` of `torch.Tensor` as the `parameters` argument) to mirror the API for non-sharded optimizers.
Fixes https://github.com/pytorch/pytorch/issues/71347 and https://github.com/pytorch/pytorch/issues/59973.
This modifies `test_collect_shards()` to skip if ROCm.
**Test Plan**
I adjusted the existing constructor test, and I added a test for parity between constructing with two parameter groups up front versus constructor with one parameter group and adding the second parameter group after (via `add_param_group()`) versus a non-sharded optimizer.
Test Plan: Imported from OSS
Reviewed By: rohan-varma
Differential Revision: D34106940
Pulled By: awgu
fbshipit-source-id: 7e70fc0b3cec891646e0698eaedf02ff4354c128
(cherry picked from commit 40f2d45172)
Summary:
Solves the next most important use case in https://github.com/pytorch/pytorch/issues/68052.
I have kept the style as close to that in SGD as seemed reasonable, given the slight differences in their internal implementations.
All feedback welcome!
cc pietern mrshenli pritamdamania87 zhaojuanmao satgera rohan-varma gqchen aazzolini osalpekar jiayisuse SciPioneer H-Huang
Pull Request resolved: https://github.com/pytorch/pytorch/pull/68164
Reviewed By: VitalyFedyunin
Differential Revision: D32994129
Pulled By: albanD
fbshipit-source-id: 65c57c3f3dbbd3e3e5338d51def54482503e8850
Summary:
Fixes https://github.com/pytorch/pytorch/issues/46480 -- for SGD.
## Notes:
- I have modified the existing tests to take a new `constructor_accepts_maximize` flag. When this is set to true, the ` _test_basic_cases_template` function will test both maximizing and minimizing the sample function.
- This was the clearest way I could think of testing the changes -- I would appreciate feedback on this strategy.
## Work to be done:
[] I need to update the docs.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/67847
Reviewed By: H-Huang
Differential Revision: D32252631
Pulled By: albanD
fbshipit-source-id: 27915a3cc2d18b7e4d17bfc2d666fe7d2cfdf9a4
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/65519
Adds buck target so we can run this internally.
ghstack-source-id: 139009957
Test Plan: CI
Reviewed By: SciPioneer
Differential Revision: D31072784
fbshipit-source-id: 7185cc1e6f9df3d79251eb017270471942a9d7dd
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/65385
Enables the ZeRO tests to run on windows. Closes
https://github.com/pytorch/pytorch/issues/63086.
Backend == NCCL was used as a proxy to see if we were running under CUDA, but Windows GPU tests uses Gloo. In this case use Gloo on GPU.
For some reason these tests don't seem to test Gloo on GPU with ZeRO in general (picks NCCL backend when GPU is available), so kept that behavior for now.
ghstack-source-id: 139003920
Test Plan: CI
Reviewed By: mrshenli
Differential Revision: D31071181
fbshipit-source-id: 45a76309ac5e882f5aa6c4b130118a68800754bb
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/62937
reland due to windows + cuda failure, fix by running it on gloo on windows even with cuda.
ghstack-source-id: 135306176
Test Plan: ci
Reviewed By: mrshenli
Differential Revision: D30177734
fbshipit-source-id: 7625746984c8f858648c1b3632394b98bd4518d2
Summary:
**Overview:**
This removes the preceding `_` from `_Join`, `_Joinable`, and `_JoinHook` in preparation for adding the generic join context manager tutorial (see [here](https://github.com/pytorch/tutorials/pull/1610)). This also adds a docs page, which can be linked from the tutorial. [Here](https://github.com/pytorch/pytorch/files/6919475/render.pdf) is a render of the docs page.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/62605
Test Plan:
`DistributedDataParallel.join()`:
```
touch /tmp/barrier && TEMP_DIR="/tmp" BACKEND="nccl" WORLD_SIZE="2" gpurun python test/distributed/test_distributed_fork.py -- TestDistBackendWithFork.test_ddp_uneven_inputs TestDistBackendWithFork.test_ddp_uneven_inputs_stop_iteration_sync_bn TestDistBackendWithFork.test_ddp_grad_div_uneven_inputs TestDistBackendWithFork.test_ddp_uneven_input_join_disable TestDistBackendWithFork.test_ddp_uneven_input_exception
```
`ZeroRedundancyOptimizer`:
```
gpurun4 python test/distributed/optim/test_zero_redundancy_optimizer.py
```
NOTE: DDP overlap tests are failing due to a landing race. See https://github.com/pytorch/pytorch/pull/62592. Once the fix is landed, I will rebase, and tests should be passing.
`Join`:
```
gpurun4 python test/distributed/algorithms/test_join.py
```
Reviewed By: mrshenli
Differential Revision: D30055544
Pulled By: andwgu
fbshipit-source-id: a5ce1f1d9f1904de3bdd4edd0b31b0a612d87026
Summary:
**Overview:**
This adds two approaches to overlapping `DistributedDataParallel.backward()` with `ZeroRedundancyOptimizer.step()` by providing two hook constructors: `hook_with_zero_step()` and `hook_with_zero_step_interleaved()`. The former waits for all backward computation to finish before starting optimizer computation, while the latter launches a partial optimizer computation using the contents of a gradient bucket once that bucket's all-reduce completes. The two approaches each suffer from their own weaknesses, and which one to use depends on the specific hardware configuration.
Both approaches can share changes to `ZeroRedundancyOptimizer`. A user should pass `overlap_with_ddp=True` to `ZeroRedundancyOptimizer`, construct a DDP communication hook using either `hook_with_zero_step()` or `hook_with_zero_step_interleaved()`, and register that communication hook. `ZeroRedundancyOptimizer.step()` should still be called in the training loop, though the optimizer computation and communication will be offloaded to originate from the communication hook. Currently, the first two iterations are vacuous, meaning they do not result in parameter updates and the inputs are ignored. This is required to finalize the DDP bucket strategy and to then initialize the `ZeroRedundancyOptimizer`'s local optimizer based on that bucketing.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/62157
Test Plan:
The existing `ZeroRedundancyOptimizer` tests pass, and new unit tests for both hooks pass:
- ~~`test_ddp_with_zero_step_parity_cpu`~~ (removed for now due to flakiness in CI -- under investigation, could possibly be similar Gloo issue as with `hook_with_zero_step_interleaved()`)
- `test_ddp_with_zero_step_parity_gpu`
- `test_ddp_with_zero_step_interleaved_parity_gpu`
These were tested on the AI AWS cluster.
An analogous `test_ddp_with_zero_step_interleaved_parity_cpu` is missing due to existing bugs with Gloo. See https://github.com/pytorch/pytorch/pull/62302.
Both approaches have been verified using an internal accuracy benchmark.
Reviewed By: mrshenli
Differential Revision: D29971046
Pulled By: andwgu
fbshipit-source-id: a7234c23c7ea253f144a698fd7e3c0fe039de5e8
Summary:
Revised version of https://github.com/pytorch/pytorch/issues/60573.
**Overview:**
This makes two changes:
- It introduces a `map_location` argument to `broadcast_object_list()`. The argument specifies the device to load tensors contained in objects received from the broadcast. This change requires modifying the implementation of `_object_to_tensor()` and `_tensor_to_object()` to use `torch.save()` and torch.load()` respectively.
- It removes all calls to `_broadcast_object()` in `ZeroRedundancyOptimizer` and the corresponding test file in favor of `broadcast_object_list()`.
The default value of `map_location` is `None`, in which case `_object_to_tensor()` and hence `broadcast_object_list()` preserve their original behavior. Namely, contained tensors are loaded to their original device.
In `consolidate_state_dict()`, I specify `map_location=torch.device("cpu")` instead of `self._default_device`. This slightly changes the behavior from before when using `_broadcast_object()`. The reason I do so is that it saves one GPU to CPU data transfer since the action immediately after receiving the broadcasted `local_state_dict` is to copy it to CPU.
Explicitly, if `map_location=self._default_device`, then the data transfer path assuming NCCL backend is as follows:
`source GPU --[before serialize]--> source CPU --[before broadcast]--> source GPU --[broadcast]--> destination GPU --[before deserialize]--> destination CPU --[deserialize]--> destination GPU --[copy]--> destination CPU`
Hence, by setting `map_location=torch.device("cpu")` instead, the suffix becomes:
`destination CPU --[deserialize]--> destination CPU --[copy]--> destination CPU`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/61539
Test Plan:
I added a test `test_broadcast_object_list_map_location()` that checks for both `map_location` as CPU and GPU that (1) tensors contained in broadcasted objects are appropriately loaded onto the specified device and (2) that the contents of the tensors are correct.
The existing `ZeroRedundancyOptimizer` tests pass.
```
gpurun4 python test/distributed/optim/test_zero_redundancy_optimizer.py
```
The existing `broadcast_object_list()` test passes:
```
touch /tmp/barrier && TEMP_DIR="/tmp" BACKEND="nccl" WORLD_SIZE="2" gpurun python test/distributed/test_distributed_fork.py -- TestDistBackendWithFork.test_broadcast_object_list
```
Reviewed By: zou3519
Differential Revision: D29701479
Pulled By: andwgu
fbshipit-source-id: c8d5f9057b32e5e9f40e8edc5b2cc25fb21414a9
Summary:
**Overview:**
This refactors the computation on non-joined processes relating to the join context manager. The concept was inspired by a comment from pritamdamania.
**Changes:**
This introduces a `_Joinable` abstract base class, which requires a `_join_hook()` method and `_join_device()` and `_join_process_group()` property methods. Any class that we want to be compatible with the generic join context manager should inherit from `_Joinable` and implement `_join_hook()`, `_join_device()`, and `_join_process_group()`. (The `device` and `process_group` information has been moved from `_JoinHook` to `_Joinable`.)
The generic join context manager now takes in a `List[_Joinable]` instead of `List[_JoinHook]`. The motivation for this is that previously, by passing the `_JoinHook`s into the context manager, the class providing a `_JoinHook` can modify the context manager's behavior, but the context manager cannot modify the class's behavior. This is solved by giving the context manager a reference to the class's instance.
This implementation reserves the field `_join_config` in every `_Joinable` to store a `_JoinConfig` instance, which holds all dynamic fields needed from the `_Joinable` for the join context manager: `enable`, `throw_on_early_termination`, and `is_first_joinable`. ("dynamic" here means that for a given `_Joinable` instance, the values for those fields may change across different join context usages.) In particular, these fields are needed to implement a method `notify_join_context()`, which encapsulates the computation performed on non-joined processes relating to the join context manager --- (1) the all-reduce to indicate that the process has not yet joined and (2) the all-reduce to check whether to throw an exception if `throw_on_uneven_inputs=True`. The idea is that every `_Joinable` class only needs to make a call to `notify_join_context()` before its per-iteration collective communications; it is a simple one-line addition.
Only the first `_Joinable` instance passed into the context manager actually performs the collective communications in `notify_join_context()`. In that case, the method returns an async work handle for the initial all-reduce indicating that the process not yet joined. Otherwise, the method returns `None`. This conditional logic is handled internally without additional input from the user.
**New API:**
Now, the example usage would look like:
```
ddp_model = DistributedDataParallel(...)
zero_optim = ZeroRedundancyOptimizer(ddp_model.parameters(), ...)
with _Join([ddp_model, zero_optim]):
...
```
Any arguments meant for a join hook (e.g. `divide_by_initial_world_size`) must be specified as keyword arguments. For example:
```
with _Join([ddp_model, zero_optim], divide_by_initial_world_size=False):
...
```
They will be forwarded to every `_join_hook()` function via `**kwargs`. This creates a clear separation between the variables needed by the context manager (`enable` and `throw_on_early_termination`) and those needed by the `_Joinable` class (e.g. `divide_by_initial_world_size`).
**Recap:**
After this change, the relevant information to use the generic join context manager looks like the following (omitting prefix `_` from names):
- Suppose we have a class `C` (e.g. `DistributedDataParallel`) that we want to be able to use the `Join` context.
- We make `C` inherit from `Joinable` and implement `join_hook() -> JoinHook`, `join_device()`, and `join_process_group()`.
- To implement `join_hook()`, we define a `CJoinHook` class inheriting from `JoinHook` and implement `main_hook()` and `post_hook()` as needed.
- We locate a place before `C`'s per-iteration collective communications and add a call to `Join.notify_join_context()`.
- We call `Joinable.__init__(self)` in `C`'s constructor.
- The `C.join_config` field will be used internally by the context manager. This does not affect `C`'s serializability.
- Run time arguments for `C`'s join hook can be passed in as keyword arguments to the context manager: `with Join([C()], arg1=..., arg2=...):`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/61555
Test Plan:
I ran the existing DDP join tests:
```
touch /tmp/barrier && TEMP_DIR="/tmp" BACKEND="nccl" WORLD_SIZE="2" gpurun python test/distributed/test_distributed_fork.py -- TestDistBackendWithFork.test_ddp_uneven_inputs TestDistBackendWithFork.test_ddp_uneven_inputs_stop_iteration_sync_bn TestDistBackendWithFork.test_ddp_grad_div_uneven_inputs TestDistBackendWithFork.test_ddp_uneven_input_join_disable TestDistBackendWithFork.test_ddp_uneven_input_exception
```
I ran the ZeRO join tests:
```
gpurun4 python test/distributed/optim/test_zero_redundancy_optimizer.py TestZeroRedundancyOptimizerDistributed.test_zero_join_gpu TestZeroRedundancyOptimizerDistributed.test_zero_join_cpu
```
Reviewed By: zou3519
Differential Revision: D29690359
Pulled By: andwgu
fbshipit-source-id: 2950f78de755eb5fb13b95b803dd7c705879a9c7
Summary:
**Overview:**
The existing `ZeroRedundancyOptimizer` implementation assumes that all model parameters are stored on the same device (due to the recent [refactor](https://github.com/pytorch/pytorch/pull/59834)). This change allows model parameters to be sharded across multiple devices, as in the DDP with Model Parallelism example [here](https://pytorch.org/tutorials/intermediate/ddp_tutorial.html).
The only logic affected is the bucketing strategy used when `parameters_as_bucket_view=True`. Let `n` denote the world size and `k` denote the number of devices per process.
- Previously, `k = 1`, and `self._buckets` was a `List[torch.Tensor]`, where `self._buckets[j]` is a tensor (i.e. bucket) containing the parameters assigned to rank `j` for `j = 0, ..., n - 1`.
- Now, `self._buckets` is a `List[List[torch.Tensor]]`, where `self._buckets[i][j]` is a tensor containing the parameters stored on device `i` assigned to rank `j` for `i = 0, ..., k - 1` and `j = 0, ..., n - 1`.
This bucket construction uses an auxiliary data structure `self._device_to_per_rank_params`, which is a `Dict[torch.device, List[List[torch.Tensor]]]`. It maps:
- `dev_0` to `[rank 0's assigned parameters on dev_0, rank 1's assigned parameters on dev_1, ...]`,
- `...`
- `dev_{k-1}` to `[rank 0's assigned parameters on dev_{k-1}, rank 1's assigned parameters on dev_{k-1}, ...]`
I removed the invariant checker `_verify_same_param_device()` and its corresponding test since it is no longer an invariant.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/61370
Test Plan: I added a new test `test_zero_model_parallel()` that checks for parity between a DDP model with model parallelism using `ZeroRedundancyOptimizer` and a local model with the same architecture using a local optimizer. I also verified that the existing tests still pass.
Reviewed By: soulitzer
Differential Revision: D29637132
Pulled By: andwgu
fbshipit-source-id: 07112959fa4e94a3f40e67e88cbb58ce3cd1e033
Summary:
Targets https://github.com/pytorch/pytorch/issues/54318.
**Overview:**
DDP offers a `join()` context manager to accommodate training on uneven inputs. This creates a new generic `_Join()` API permitting custom hooks, refactors DDP `join()` to call this generic `_Join()`, and implements a hook for ZeRO. (For now, the generic `_Join()` is implemented as private, but this may change after design discussions are cleared.)
There are two classes introduced: `_JoinHook`, the class defining the customizable join hook, and `_Join`, the generic join context manager.
The `_JoinHook` provides two entry points: `main_hook()`, which is called repeatedly while there exists a non-joined process, and `post_hook()`, which is called once all process have joined with the additional `bool` argument `is_last_joiner`. The class also requires `process_group` and `device` information by defining corresponding abstract property methods. Thus, to implement a join hook, (1) inherit from `_JoinHook`, (2) override `main_hook()` and `post_hook()` as appropriate, and (3) override `process_group()` and `device()` to provide process group and device information to be used by the join context manager implementation for collective communications.
The `_Join` constructor requires `join_hooks: List[_JoinHook]` and optionally `enable: bool = True` and `throw_on_early_termination: bool = False`. A training loop only needs to be wrapped with `with _Join(join_hooks):` (using the appropriate `join_hooks`) to be able to train on uneven inputs without hanging/erroring. The context manager requires a `dist.all_reduce(torch.ones(1))` to be called on every non-joined process each time before it performs its collective communications in order to indicate that the process has not yet joined. It also requires that all `process_group` attributes in the `_JoinHook` objects are the same.
**Notes:**
- The argument `is_last_joiner` to `post_hook()` may be useful for finding an authoritative rank when synchronizing.
- `enable` is a flag that can be set to `False` if the user knows the current training loop will not have uneven inputs. This may be used to disable join-related computation in the classes providing join hooks.
- `throw_on_early_termination` is a flag that can be set to `True` to notify processes to terminate upon detecting uneven inputs (i.e. upon the first process joining when there exists a non-joined process). Notably, the notification requires an all-reduce, so to prevent hanging/erroring, non-joined process must participate in the all-reduce. The first-joining process raises a `RuntimeError`, and the other processes are expected (but not required) to do the same. This may be used to implement training on uneven inputs in cases that do not conform to the generic join context manager (e.g. `SyncBatchNorm`).
- Classes providing a join hook should do so via a `_join_hook()` method that returns a `_JoinHook` instance with the methods appropriately overridden.
- If there are multiple join hooks, the device specified by the first is used by the join context manager implementation to perform its collective communications.
- If there are multiple join hooks, both the main and post-hooks are iterated in the order in which the `_JoinHook` objects are passed into the context manager constructor.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/60757
Test Plan:
The current implementation preserves backward compatibility by not changing the existing DDP `join()` API at all. To check this, I ran through the uneven input tests (`test_ddp_grad_div_uneven_inputs`, `test_ddp_uneven_inputs_stop_iteration_sync_bn`, `test_ddp_uneven_inputs`, `test_ddp_uneven_input_join_disable`, `test_ddp_uneven_input_exception`) on the AI AWS cluster:
```
touch /tmp/barrier && TEMP_DIR="/tmp" BACKEND="nccl" WORLD_SIZE="2" gpurun python test/distributed/test_distributed_fork.py --
```
Because the existing DDP join logic does not provide correct gradients to the joined processes if `gradient_as_bucket_view=False` and a joined process requires those gradients to correctly update its shard of the parameters in `ZeroRedundancyOptimizer.step()`, DDP and ZeRO are not fully compatible at the moment. To work around this and to test ZeRO's join hook separately, I added a test `_test_zero_join()` (with `test_zero_join_gpu()` and `test_zero_join_cpu()` flavors), which compares DDP with a local optimizer on uneven inputs against ZeRO on uneven inputs with the gradients set manually.
Reviewed By: iramazanli, mrshenli
Differential Revision: D29624636
Pulled By: andwgu
fbshipit-source-id: ec70a290e02518b0d8b683f9fed2126705b896c7
Summary:
**Overview:**
Being relatively new to PyTorch and ZeRO, I found parts of the code slightly hard to follow. This change strives to clean up the `ZeroRedundancyOptimizer` code in `zero_redundancy_optimizer.py` by reorganizing some computations, making variable names more explicit and consistent, and unifying terminology in the documentation. The goal is for the code to be easier to extend afterwards.
**Changes:**
1) `state_dict()`: The [logic](85517a2b70/torch/distributed/optim/zero_redundancy_optimizer.py (L510)) for updating the global `state_dict` with each rank's local `state_dict` is simplified and made more explicit. Notably, the `dict` [`local_index_to_param_id`](85517a2b70/torch/distributed/optim/zero_redundancy_optimizer.py (L513)) is unneeded. It maps `local_pg["params"][i]` to `id(global_pg["params"][i])`, so it is equivalent to make a single pass over both lists in tandem, effectively iterating over `i`, without a need for the explicit `dict`.
2) `_update_trainable()`: The function [initializes](85517a2b70/torch/distributed/optim/zero_redundancy_optimizer.py (L597)) the local optimizer if it does not exist. I am unaware of any reason for the local optimizer to be destroyed after initialization, so I moved that logic to its own function `_init_local_optimizer()`, which is called once in the constructor.
After [discussion](https://github.com/pytorch/pytorch/pull/60285#discussion_r654706728), I removed the function `_update_trainable()` itself in favor of adding a check for `parameters_as_bucket_view` in `build_param_buckets()` directly.
3) `rank_local_state_dict()`: This [function](85517a2b70/torch/distributed/optim/zero_redundancy_optimizer.py (L528)) is currently broken. It appears to be legacy and relies on the input `state_dict` to have the key `"partitions"`. For now, I have removed it and added an [issue](https://github.com/pytorch/pytorch/issues/60284). Is it a notable use case to want to access another rank's `state_dict` in particular (as opposed to consolidating the entire state and then accessing)?
4) `local_state_dict():` After [discussion](https://github.com/pytorch/pytorch/pull/60285#discussion_r655571043), I removed the function.
5) `partition_parameters()`: After [discussion](https://github.com/pytorch/pytorch/pull/60285#discussion_r654708183), I renamed the function to `_partition_parameters()` to mark it as private.
6) `_param_to_index`: After [discussion](https://github.com/pytorch/pytorch/pull/60285#discussion_r654828100), I changed the key to be the parameter itself rather than its integer ID.
7) `buckets`: I renamed the data structure to `_buckets` to mark it as private.
8) Terminology: I tried to reduce the set of terms being used instead of juggling a number of synonyms. In particular, I made an effort to distinguish between "local" and "global" and to make names more indicative of typing.
9) Style: Per the [PyTorch contributing guide](https://github.com/pytorch/pytorch/blob/master/CONTRIBUTING.md#writing-documentation), I made all docstrings abide by the 80 character limit, except for the one [line](554891f6fa/torch/distributed/optim/zero_redundancy_optimizer.py (L142)) showing the example ZeRO usage. Some code lines violate the limit for readability. Also, I unified some of the minor stylistic usages out of habit.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/60285
Test Plan:
The test suite passes as expected (on the AI AWS cluster):
```
gpurun python test/distributed/optim/test_zero_redundancy_optimizer.py
```
I visually inspected the generated HTML doc (as generated following [this](https://github.com/pytorch/pytorch/blob/master/CONTRIBUTING.md#writing-documentation)).
Reviewed By: mrshenli
Differential Revision: D29320726
Pulled By: andwgu
fbshipit-source-id: 23f69a19ecc5e877a38fe1df0da11329428311dd
Summary:
**Overview:**
This refactors the `ZeroRedundancyOptimizer` implementation to assume single-process single-device (SPSD) instead of accommodating single-process multiple-device (SPMD). `DistributedDataParallel` [retired SPMD recently](https://github.com/pytorch/pytorch/issues/47012), so this change follows the same spirit.
**Changes:**
The parent-class `Optimizer` constructor permits the input argument `params` to be both an `iterable` of `torch.Tensor` and an `iterable` of `dict`. The latter usage is for initializing the optimizer with multiple `param_group`s to start. However, currently, `ZeroRedundancyOptimizer` only supports the former usage, requiring explicit calls to `add_param_group()` for multiple `param_group`s. Given the existing implementation, the type error would be silent and not manifest until much later (e.g. since `super().__init__()` would have no issue). Hence, I added a series of checks to begin the `__init__()` function (encapsulated in `_verify_and_init_params()`). A postcondition of this validation is that `self._all_params` is a non-empty list of all model parameters.
Additionally, I added a check for SPSD usage assuming that all model parameters exist on the same device. This logic is included in `_verify_same_param_device()` and is called immediately after the `params` type-checking. Support for SPSD with model parameters sharded across devices may be added in the future.
Related to that aforementioned post-condition on `self._all_params`, previously there was undefined behavior resulting from different typing of the passed in `params` input argument. If `params` was a `List`, then the usage of `self._reference_is_trainable_mask` was as expected. However, if `params` was a generator (e.g. as in the canonical usage of passing `model.parameters()`), then the ensuing behavior was divergent. This is because after a generator is iterated over, it is empty. As a result, when we set `self._all_params = params` [in the old code](68d690ffbd/torch/distributed/optim/zero_redundancy_optimizer.py (L165)), `self._all_params` is empty, reducing `training_mask` to always be the empty list. This causes missed calls to `_update_trainable()` in `step()`. (A consequence of this is that `test_pytorch_parity()`, which is renamed to `test_local_optimizer_parity()`, now outputs warnings about the trainable parameters changing.)
The existing implementation assumes that all parameters share the same dense type when allocating the bucket buffers. This change preserves this assumption, which may be removed in the future. I added a check for this in `_verify_same_dense_param_type()` to avoid erroring silently later on. Note that it is insufficient to simply check for the same `dtype` since dense and sparse tensors may share the same `dtype` but require differing storage sizes. One solution is to use `torch.typename()` as the means for comparison.
---
The primary change in this refactor is with respect to `self._per_device_params` and `self.buckets`. `self._per_device_params` mapped `torch.device` to `List[List[Parameter]]`. The keys were the devices that the model parameters exist on, and the values designated which ranks are assigned to updating those parameters. `self.buckets` mapped `torch.device` to `List[torch.Tensor]`. The keys were the same as `self._per_device_params`, and the values were the buckets for that device. The usage of these two data structures were confined to each other only. Hence, because the notions of device and rank are now in 1:1 correspondence, we can eliminate the former completely and only use rank. As such, I removed `self._per_device_params` and made `self.buckets` directly a list of buckets (i.e. `torch.Tensor`s).
Iteration over the parameters of a rank for a given device could be simplified to just iteration over the parameters of a rank. Hence, I relied on `self.partition_parameters()` now for that iteration. Refer to `_setup_flat_buffers()` and `step()` for these changes.
One convenient side effect of removing `self._per_device_params` is that there is no longer the re-computation of the parameter partitions mentioned at the end of this [PR](https://github.com/pytorch/pytorch/pull/59410).
---
I changed the data structure `self._index_to_param_cache` from a `dict` to a `List` because the domain is `0`, `1`, ..., `k-1` where `k` is the number of parameters. This should yield marginal improvements in memory usage and access speed.
`_sync_param_groups()` is a static method, meaning it can be called either via `self._sync_param_groups()` or `ZeroRedundancyOptimizer._sync_param_groups()` when inside the class. I made the usage consistently `self._sync_param_groups()` rather than have instances of both.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/59834
Test Plan:
I ran through the existing test suite on an AI AWS cluster:
```
srun -p $DEV_QUEUE --cpus-per-task=16 -t 5:00:00 --gpus-per-node=4 python test/distributed/optim/test_zero_redundancy_optimizer.py
```
Note: The only test where `parameters_as_bucket_view` is `True` is `test_step_with_closure()`, meaning that that is the test that exercises the core changes of removing `self._per_device_params` and changing `self.buckets`.
Also, I added tests for the `ZeroRedundancyOptimizer` constructor changes and the assumption checks.
Reviewed By: mrshenli
Differential Revision: D29177065
Pulled By: andwgu
fbshipit-source-id: 0ff004ae3959d6d3b521024028c7156bfddc93d8
Summary:
Addresses https://github.com/pytorch/pytorch/issues/59548
**Overview:**
Recently, we changed ZeRO's partitioning algorithm to first sort the parameters by decreasing size and then greedily allocate to shards. See [here](ea1de87f4b).
The current tests `test_sharding()` and `test_add_param_group()` check for a uniform partitioning, which is not achieved with the old naive greedy partitioning algorithm for general world sizes but is achieved with the new sorted-greedy algorithm. This reliance is not ideal, but for now, we opt to simply add comments to document the dependency.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/59713
Test Plan:
I tested for world sizes of 1, 2, 3, and 4 via the AI AWS cluster:
```
srun -p $DEV_QUEUE --cpus-per-task=16 -t 5:00:00 --gpus-per-node=1 python test/distributed/optim/test_zero_redundancy_optimizer.py -- TestZeroRedundancyOptimizerDistributed.test_sharding
srun -p $DEV_QUEUE --cpus-per-task=16 -t 5:00:00 --gpus-per-node=2 python test/distributed/optim/test_zero_redundancy_optimizer.py -- TestZeroRedundancyOptimizerDistributed.test_sharding
srun -p $DEV_QUEUE --cpus-per-task=16 -t 5:00:00 --gpus-per-node=3 python test/distributed/optim/test_zero_redundancy_optimizer.py -- TestZeroRedundancyOptimizerDistributed.test_sharding
srun -p $DEV_QUEUE --cpus-per-task=16 -t 5:00:00 --gpus-per-node=4 python test/distributed/optim/test_zero_redundancy_optimizer.py -- TestZeroRedundancyOptimizerDistributed.test_sharding
srun -p $DEV_QUEUE --cpus-per-task=16 -t 5:00:00 --gpus-per-node=1 python test/distributed/optim/test_zero_redundancy_optimizer.py -- TestZeroRedundancyOptimizerDistributed.test_add_param_group
srun -p $DEV_QUEUE --cpus-per-task=16 -t 5:00:00 --gpus-per-node=2 python test/distributed/optim/test_zero_redundancy_optimizer.py -- TestZeroRedundancyOptimizerDistributed.test_add_param_group
srun -p $DEV_QUEUE --cpus-per-task=16 -t 5:00:00 --gpus-per-node=3 python test/distributed/optim/test_zero_redundancy_optimizer.py -- TestZeroRedundancyOptimizerDistributed.test_add_param_group
srun -p $DEV_QUEUE --cpus-per-task=16 -t 5:00:00 --gpus-per-node=4 python test/distributed/optim/test_zero_redundancy_optimizer.py -- TestZeroRedundancyOptimizerDistributed.test_add_param_group
```
However, because the train queue (which offers instances with 8 GPUs) is not working at the moment, I was unable to test for world sizes of 5+. Nonetheless, I believe that they should still work.
First, consider `test_sharding()`. Given the sorted-greedy algorithm, each shard will be assigned one of the parameters with size `9`, then one of the parameters with size `7`, then `5`, and finally `3`. Hence, each will have a uniform partition. Now, consider `test_add_param_group()`. Similarly, the same allocation behavior occurs, only the last shard is not assigned the final parameter with size `3` to begin. However, after adding the new `param_group` with the parameter with size `3`, a re-partitioning occurs. The first `param_group` is partitioned as before, and the parameter with size `3` in the new `param_group` is assigned to the last shard since it has the minimal total size. Thus, in the end, all shards have a uniform partition.
Reviewed By: mrshenli
Differential Revision: D28996460
Pulled By: andwgu
fbshipit-source-id: 22bdc638d8569ed9a20836812eac046d628d6df2
Summary:
Fixes https://github.com/pytorch/pytorch/issues/53322, the test has some hardcoded values to check that the sharding works as expected, and was not used beyond 4 gpus prior
Pull Request resolved: https://github.com/pytorch/pytorch/pull/54788
Reviewed By: mrshenli
Differential Revision: D27483078
Pulled By: blefaudeux
fbshipit-source-id: 63fe072c41e1601925af23d8fb1ea3f4729b2044
Summary:
Currently there is some code that intends to skip distributed tests if
the distributed module is not built. However, they are missing in some
test files; and in some other test files they are checked after
distributed module is imported, which leads to failure. This is
generating a lot of headaches when testing minimal builds locally.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/52945
Reviewed By: anjali411
Differential Revision: D26848241
Pulled By: ezyang
fbshipit-source-id: 983a848844add40869a86f3c9413503a3659b115
Summary:
Updated version following https://github.com/pytorch/pytorch/issues/52764 (including comments from Shen), but this one I expect to be able to land.
ZeroRedundancyOptimizer:
- bucket as tensor views, optional
- make a lot of attributes private
- minor unit test refactor
- adding coverage in the unit test for with and without bucket views
Pull Request resolved: https://github.com/pytorch/pytorch/pull/52987
Reviewed By: mrshenli
Differential Revision: D26728851
Pulled By: blefaudeux
fbshipit-source-id: f8c745966719c9076c20a554ef56198fb838856c
Summary:
Same as https://github.com/pytorch/pytorch/issues/52760 which I could not get to land. I just could not live with ghstack/ghimport/randomly broken things, I break enough of them myself, so this is a fresh copy without ghstack shenanigans. I'm hopeful that this can land relatively bug free, and am sorry for the duplications..
What this does:
- call the common_utils test runner instead of unittest, because it seems that it's how it should be done
- change the returned state from ZeroRedundancyOptimizer to be PyTorch compliant, which has the added benefit of being elastic (world size independent)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/52960
Reviewed By: mrshenli
Differential Revision: D26710932
Pulled By: blefaudeux
fbshipit-source-id: 1d914bc9221442ba1bb2b48f5df10c313e674ece
Summary:
- Makes it possible to use non-sharded optimizer checkpoints (as long as the model/param groups are the same, of course)
- Makes it possible to save with a given world size, and load with another world size
- Use Torch Distributed built-in broadcast object list instead of a ad-hoc version
Pull Request resolved: https://github.com/pytorch/pytorch/pull/50956
Reviewed By: malfet
Differential Revision: D26113953
Pulled By: blefaudeux
fbshipit-source-id: 030bfeee2c34c2d987590d45dc8efe05515f2e5c
Summary:
Implement the first stage of ZeRO, sharding of the optimizer state, as described in [this blog post](https://www.microsoft.com/en-us/research/blog/zero-2-deepspeed-shattering-barriers-of-deep-learning-speed-scale/) and [this paper](https://arxiv.org/abs/1910.02054). This implementation is completely independent from the [DeepSpeed](https://github.com/microsoft/DeepSpeed) framework, and aims at providing ZeRO-compliant building blocks within the PyTorch scheme of things.
This works by:
- acting as a wrapper to a pytorch optimizer. ZeROptimizer does not optimize anything by itself, it only shards optimizers for distributed jobs
- each rank distributes parameters according to a given partitioning scheme (could be updated), and owns the update of a given shard only
- the .step() is called on each rank as expected, the fact that the optimizer actually works on a shard of the model is not visible from the outside
- when the update is completed, each rank broadcasts the updated model shard to all the other ranks
This can be used with DDP, although some communications are wasted in that case (gradients are all-reduced to all ranks). This implementation was initially developed in [Fairscale](https://github.com/facebookresearch/fairscale), and can also be used with an optimized DDP which only reduces to the relevant ranks. More context on ZeRO and PyTorch can be found in [this RFC](https://github.com/pytorch/pytorch/issues/42849)
The API with respect to loading and saving the state is a known pain point and should probably be discussed an updated. Other possible follow ups include integrating more closely to a [modularized DDP](https://github.com/pytorch/pytorch/issues/37002), [making the checkpoints partition-agnostic](https://github.com/facebookresearch/fairscale/issues/164), [exposing a gradient clipping option](https://github.com/facebookresearch/fairscale/issues/98) and making sure that mixed precision states are properly handled.
original authors include msbaines, min-xu-ai and myself
Pull Request resolved: https://github.com/pytorch/pytorch/pull/46750
Reviewed By: mruberry
Differential Revision: D25958918
Pulled By: blefaudeux
fbshipit-source-id: 14280f2fd90cf251eee8ef9ac0f1fa6025ae9c50
