Write full tensors out at once in HF consolidation script#159394
Closed
ankitageorge wants to merge 6 commits intogh/ankitageorge/16/basefrom
Closed
Write full tensors out at once in HF consolidation script#159394ankitageorge wants to merge 6 commits intogh/ankitageorge/16/basefrom
ankitageorge wants to merge 6 commits intogh/ankitageorge/16/basefrom
Conversation
Not all storage systems support writing at random offsets. This PR changes the writes of the consolidation script to write each tensor to a buffer, and then write out the buffer, sequentially going through every tensor in the output file. This will also help in the case where the sharded files weren't just sharded in the row-wise dimension. The reason is because small writes are expensive and we were writing each write for every chunk that was the largest number of contiguous bytes in the final tensor, but this could be a small amount of bytes for col-wise sharding. Now the full tensor is needed for the write, making the number of small writes smaller. Differential Revision: [D78684452](https://our.internmc.facebook.com/intern/diff/D78684452/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/159394
Note: Links to docs will display an error until the docs builds have been completed. ⏳ No Failures, 1 PendingAs of commit 48753cb with merge base eed9dbf ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This was referenced Jul 29, 2025
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D78684452 |
Not all storage systems support writing at random offsets. This PR changes the writes of the consolidation script to write each tensor to a buffer, and then write out the buffer, sequentially going through every tensor in the output file. This will also help in the case where the sharded files weren't just sharded in the row-wise dimension. The reason is because small writes are expensive and we were writing each write for every chunk that was the largest number of contiguous bytes in the final tensor, but this could be a small amount of bytes for col-wise sharding. Now the full tensor is needed for the write, making the number of small writes smaller. Differential Revision: [D78684452](https://our.internmc.facebook.com/intern/diff/D78684452/) cc H-Huang awgu wanchaol fegin fduwjj wz337 wconstab d4l3k pragupta [ghstack-poisoned]
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D78684452 |
Not all storage systems support writing at random offsets. This PR changes the writes of the consolidation script to write each tensor to a buffer, and then write out the buffer, sequentially going through every tensor in the output file. This will also help in the case where the sharded files weren't just sharded in the row-wise dimension. The reason is because small writes are expensive and we were writing each write for every chunk that was the largest number of contiguous bytes in the final tensor, but this could be a small amount of bytes for col-wise sharding. Now the full tensor is needed for the write, making the number of small writes smaller. Differential Revision: [D78684452](https://our.internmc.facebook.com/intern/diff/D78684452/) cc H-Huang awgu wanchaol fegin fduwjj wz337 wconstab d4l3k pragupta [ghstack-poisoned]
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D78684452 |
Not all storage systems support writing at random offsets. This PR changes the writes of the consolidation script to write each tensor to a buffer, and then write out the buffer, sequentially going through every tensor in the output file. This will also help in the case where the sharded files weren't just sharded in the row-wise dimension. The reason is because small writes are expensive and we were writing each write for every chunk that was the largest number of contiguous bytes in the final tensor, but this could be a small amount of bytes for col-wise sharding. Now the full tensor is needed for the write, making the number of small writes smaller. Differential Revision: [D78684452](https://our.internmc.facebook.com/intern/diff/D78684452/) cc H-Huang awgu wanchaol fegin fduwjj wz337 wconstab d4l3k pragupta [ghstack-poisoned]
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D78684452 |
Not all storage systems support writing at random offsets. This PR changes the writes of the consolidation script to write each tensor to a buffer, and then write out the buffer, sequentially going through every tensor in the output file. This will also help in the case where the sharded files weren't just sharded in the row-wise dimension. The reason is because small writes are expensive and we were writing each write for every chunk that was the largest number of contiguous bytes in the final tensor, but this could be a small amount of bytes for col-wise sharding. Now the full tensor is needed for the write, making the number of small writes smaller. Differential Revision: [D78684452](https://our.internmc.facebook.com/intern/diff/D78684452/) cc H-Huang awgu wanchaol fegin fduwjj wz337 wconstab d4l3k pragupta [ghstack-poisoned]
ankitageorge
added a commit
that referenced
this pull request
Aug 1, 2025
Pull Request resolved: #159394 Not all storage systems support writing at random offsets. This PR changes the writes of the consolidation script to write each tensor to a buffer, and then write out the buffer, sequentially going through every tensor in the output file. This will also help in the case where the sharded files weren't just sharded in the row-wise dimension. The reason is because small writes are expensive and we were writing each write for every chunk that was the largest number of contiguous bytes in the final tensor, but this could be a small amount of bytes for col-wise sharding. Now the full tensor is needed for the write, making the number of small writes smaller. ghstack-source-id: 300234774 @exported-using-ghexport Differential Revision: [D78684452](https://our.internmc.facebook.com/intern/diff/D78684452/)
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D78684452 |
saumishr
reviewed
Aug 12, 2025
| - Optimized chunks for other patterns | ||
|
|
||
| Args: | ||
| full_tensor_mv: Buffer to write the full tensor to |
Contributor
There was a problem hiding this comment.
Can the buffered approach increase the chances of OOMs compared to the earlier file stream write?
Contributor
Author
There was a problem hiding this comment.
ya it probably can have OOMs if you have many parallel threads at the same time, but I don't think a single tensor would be more than 10GB (as a guess), so then in that case any more than 8 threads could cause OOMs. I think it would be up to users to manage this. But there isn't really an alternative for the remote storage case since many of them don't support random offset writes
| @@ -329,13 +341,12 @@ def _write_data( | |||
|
|
|||
|
|
|||
| def _write_sub_tensor_to_file_optimized( | |||
Contributor
There was a problem hiding this comment.
It may be good to add a test for this method.
Not all storage systems support writing at random offsets. This PR changes the writes of the consolidation script to write each tensor to a buffer, and then write out the buffer, sequentially going through every tensor in the output file. This will also help in the case where the sharded files weren't just sharded in the row-wise dimension. The reason is because small writes are expensive and we were writing each write for every chunk that was the largest number of contiguous bytes in the final tensor, but this could be a small amount of bytes for col-wise sharding. Now the full tensor is needed for the write, making the number of small writes smaller. Differential Revision: [D78684452](https://our.internmc.facebook.com/intern/diff/D78684452/) cc H-Huang awgu wanchaol fegin fduwjj wz337 wconstab d4l3k pragupta [ghstack-poisoned]
ankitageorge
added a commit
that referenced
this pull request
Aug 12, 2025
Pull Request resolved: #159394 Not all storage systems support writing at random offsets. This PR changes the writes of the consolidation script to write each tensor to a buffer, and then write out the buffer, sequentially going through every tensor in the output file. This will also help in the case where the sharded files weren't just sharded in the row-wise dimension. The reason is because small writes are expensive and we were writing each write for every chunk that was the largest number of contiguous bytes in the final tensor, but this could be a small amount of bytes for col-wise sharding. Now the full tensor is needed for the write, making the number of small writes smaller. ghstack-source-id: 302555412 @exported-using-ghexport Differential Revision: [D78684452](https://our.internmc.facebook.com/intern/diff/D78684452/)
Contributor
|
This pull request was exported from Phabricator. Differential Revision: D78684452 |
saumishr
approved these changes
Aug 12, 2025
Contributor
Author
|
@pytorchmergebot merge |
Collaborator
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
chuanhaozhuge
pushed a commit
that referenced
this pull request
Aug 14, 2025
Not all storage systems support writing at random offsets. This PR changes the writes of the consolidation script to write each tensor to a buffer, and then write out the buffer, sequentially going through every tensor in the output file. This will also help in the case where the sharded files weren't just sharded in the row-wise dimension. The reason is because small writes are expensive and we were writing each write for every chunk that was the largest number of contiguous bytes in the final tensor, but this could be a small amount of bytes for col-wise sharding. Now the full tensor is needed for the write, making the number of small writes smaller. Differential Revision: [D78684452](https://our.internmc.facebook.com/intern/diff/D78684452/) Pull Request resolved: #159394 Approved by: https://github.com/saumishr ghstack dependencies: #159392, #159393
chuanhaozhuge
pushed a commit
that referenced
this pull request
Aug 18, 2025
Not all storage systems support writing at random offsets. This PR changes the writes of the consolidation script to write each tensor to a buffer, and then write out the buffer, sequentially going through every tensor in the output file. This will also help in the case where the sharded files weren't just sharded in the row-wise dimension. The reason is because small writes are expensive and we were writing each write for every chunk that was the largest number of contiguous bytes in the final tensor, but this could be a small amount of bytes for col-wise sharding. Now the full tensor is needed for the write, making the number of small writes smaller. Differential Revision: [D78684452](https://our.internmc.facebook.com/intern/diff/D78684452/) Pull Request resolved: #159394 Approved by: https://github.com/saumishr ghstack dependencies: #159392, #159393
can-gaa-hou
pushed a commit
to can-gaa-hou/pytorch
that referenced
this pull request
Aug 22, 2025
…9394) Not all storage systems support writing at random offsets. This PR changes the writes of the consolidation script to write each tensor to a buffer, and then write out the buffer, sequentially going through every tensor in the output file. This will also help in the case where the sharded files weren't just sharded in the row-wise dimension. The reason is because small writes are expensive and we were writing each write for every chunk that was the largest number of contiguous bytes in the final tensor, but this could be a small amount of bytes for col-wise sharding. Now the full tensor is needed for the write, making the number of small writes smaller. Differential Revision: [D78684452](https://our.internmc.facebook.com/intern/diff/D78684452/) Pull Request resolved: pytorch#159394 Approved by: https://github.com/saumishr ghstack dependencies: pytorch#159392, pytorch#159393
markc-614
pushed a commit
to markc-614/pytorch
that referenced
this pull request
Sep 17, 2025
…9394) Not all storage systems support writing at random offsets. This PR changes the writes of the consolidation script to write each tensor to a buffer, and then write out the buffer, sequentially going through every tensor in the output file. This will also help in the case where the sharded files weren't just sharded in the row-wise dimension. The reason is because small writes are expensive and we were writing each write for every chunk that was the largest number of contiguous bytes in the final tensor, but this could be a small amount of bytes for col-wise sharding. Now the full tensor is needed for the write, making the number of small writes smaller. Differential Revision: [D78684452](https://our.internmc.facebook.com/intern/diff/D78684452/) Pull Request resolved: pytorch#159394 Approved by: https://github.com/saumishr ghstack dependencies: pytorch#159392, pytorch#159393
Khanaksahu
pushed a commit
to Khanaksahu/pytorch-fork
that referenced
this pull request
Nov 17, 2025
Pull Request resolved: pytorch/pytorch#159394 Not all storage systems support writing at random offsets. This PR changes the writes of the consolidation script to write each tensor to a buffer, and then write out the buffer, sequentially going through every tensor in the output file. This will also help in the case where the sharded files weren't just sharded in the row-wise dimension. The reason is because small writes are expensive and we were writing each write for every chunk that was the largest number of contiguous bytes in the final tensor, but this could be a small amount of bytes for col-wise sharding. Now the full tensor is needed for the write, making the number of small writes smaller. ghstack-source-id: 300120717 @exported-using-ghexport Differential Revision: [D78684452](https://our.internmc.facebook.com/intern/diff/D78684452/)
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
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 (oldest at bottom):
Not all storage systems support writing at random offsets. This PR changes the writes of the consolidation script to write each tensor to a buffer, and then write out the buffer, sequentially going through every tensor in the output file. This will also help in the case where the sharded files weren't just sharded in the row-wise dimension. The reason is because small writes are expensive and we were writing each write for every chunk that was the largest number of contiguous bytes in the final tensor, but this could be a small amount of bytes for col-wise sharding. Now the full tensor is needed for the write, making the number of small writes smaller.
Differential Revision: D78684452
cc @H-Huang @awgu @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @pragupta