[FSDP] Limit all gather after pre-unshard #89057
Closed
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
[ghstack-poisoned]
awgu
requested review from
mrshenli,
zhaojuanmao,
pritamdamania87,
rohan-varma,
H-Huang and
kwen2501
as code owners
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/89057
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 47af6d0: This comment was automatically generated by Dr. CI and updates every 15 minutes. |
pytorch-bot
bot
added
the
release notes: distributed (fsdp)
awgu
added a commit
to awgu/pytorch
that referenced
this pull request
Nov 23, 2022
ghstack-source-id: 72e1b5f3d523e6fde6ca183af492cee947adbb43 Pull Request resolved: pytorch#89057
awgu
added a commit
to awgu/pytorch
that referenced
this pull request
Nov 29, 2022
ghstack-source-id: 72e1b5f3d523e6fde6ca183af492cee947adbb43 Pull Request resolved: pytorch#89057
To reuse memory when allocating the unsharded `FlatParameter` in the unshard stream, we only need to block the CPU thread on the preceding free event (i.e. `event.synchronize()`) before allocating the unsharded memory, which happens in `handle.unshard()`. Notably, this can be done after the pre-unshard logic, which at most performs _sharded_ allocations (low precision shard or H2D sharded `FlatParameter` copy) in its own pre-unshard stream. This enables the pre-unshard to overlap with any pending ops. With this change, I believe that we should use `limit_all_gathers=True` all the time to stay true to FSDP's proposed memory semantics. If a user wants to set `limit_all_gathers=False`, that would mean that he/she wants to overlap ops that are issued after the unshard logic's all-gather with ops that are pending at the time when FSDP _would_ block the CPU thread via `event.synchronize()`. - If the user is willing to not reuse memory for that all-gather, then the user may as well have applied `NO_SHARD` and optionally ZeRO-1 (if this niche is important, then maybe we should consider hardening ZeRO-1). This is because now the unsharded memory for the all-gather additionally contributes to peak memory since it cannot reuse memory. - If the user wanted to reuse memory for that all-gather, then we needed to block the CPU thread. There is no way around that given the caching allocator semantics. [ghstack-poisoned]
kulinseth
pushed a commit
to kulinseth/pytorch
that referenced
this pull request
Dec 10, 2022
To reuse memory when allocating the unsharded `FlatParameter` in the unshard stream, we only need to block the CPU thread on the preceding free event (i.e. `event.synchronize()`) before allocating the unsharded memory, which happens in `handle.unshard()`. Notably, this can be done after the pre-unshard logic, which at most performs _sharded_ allocations (low precision shard or H2D sharded `FlatParameter` copy) in its own pre-unshard stream. This enables the pre-unshard to overlap with any pending ops. With this change, I believe that we should use `limit_all_gathers=True` all the time to stay true to FSDP's proposed memory semantics. If a user wants to set `limit_all_gathers=False`, that would mean that he/she wants to overlap ops that are issued after the unshard logic's all-gather with ops that are pending at the time when FSDP _would_ block the CPU thread via `event.synchronize()`. - If the user is willing to not reuse memory for that all-gather, then the user may as well have applied `NO_SHARD` and optionally ZeRO-1 (if this niche is important, then maybe we should consider hardening ZeRO-1). This is because now the unsharded memory for the all-gather additionally contributes to peak memory since it cannot reuse memory. - If the user wanted to reuse memory for that all-gather, then we needed to block the CPU thread. There is no way around that given the caching allocator semantics. Pull Request resolved: pytorch#89057 Approved by: https://github.com/mrshenli
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.
Stack from ghstack:
ModuleWrapPolicy.__repr__#89058 [FSDP] Include module classes inModuleWrapPolicy.__repr__To reuse memory when allocating the unsharded
FlatParameterin the unshard stream, we only need to block the CPU thread on the preceding free event (i.e.event.synchronize()) before allocating the unsharded memory, which happens inhandle.unshard(). Notably, this can be done after the pre-unshard logic, which at most performs sharded allocations (low precision shard or H2D shardedFlatParametercopy) in its own pre-unshard stream. This enables the pre-unshard to overlap with any pending ops.With this change, I believe that we should use
limit_all_gathers=Trueall the time to stay true to FSDP's proposed memory semantics.If a user wants to set
limit_all_gathers=False, that would mean that he/she wants to overlap ops that are issued after the unshard logic's all-gather with ops that are pending at the time when FSDP would block the CPU thread viaevent.synchronize().NO_SHARDand optionally ZeRO-1 (if this niche is important, then maybe we should consider hardening ZeRO-1). This is because now the unsharded memory for the all-gather additionally contributes to peak memory since it cannot reuse memory.