Closed
Conversation
[ghstack-poisoned]
liangel-02
requested review from
albanD,
jbschlosser and
mikaylagawarecki
as code owners
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/164502
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 59331ad with merge base ffe3cb2 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
[ghstack-poisoned]
[ghstack-poisoned]
jbschlosser
reviewed
Oct 3, 2025
drisspg
reviewed
Oct 7, 2025
drisspg
reviewed
Oct 7, 2025
**Summary** Today, the only way to have variable sequence length support in PyTorch attention is through nested tensors [here](https://docs.pytorch.org/tutorials/intermediate/scaled_dot_product_attention_tutorial.html#nestedtensor-and-dense-tensor-support). We also want to add an explicit lower-level API that provides variable sequence length support without padding/masking in SDPA. This PR builds out `varlen_attn`, the public API that users can call for the forward method, and `_varlen_attn`, the private API that calls into the Flash Attention/cuDNN backend. **Benchmarking** To benchmark, we compare runtime and TFLOPs against the current SDPA approach with padding. Settings: - 1 H100 machine - `batch_size=8`, `max_seq_len=2048`, `embed_dim=1024`, `num_heads=16` - dtype `torch.bfloat16` - `is_causal=False` - for variable length, we set sequences to be random multiples of 64 up to `max_seq_len` - 100 runs | | Variable Length API | SDPA | |--------|--------------------|----------| | Runtime | 0.21750560760498047 ms | 0.43171775817871094 ms | | TFLOPs | 231.812 | 320.840 | The sparsity is 0.453 which we can see matches the speedup we get from Varlen (approx 50%). TFLOPs remains around the same, with SDPA slightly larger due to potential higher overhead and total flops scaling with sequence length. **Testing** Run `python test/test_varlen_attention.py` for unit tests where we verify basic functionality and confirm numerical match between varlen outputs vs SDPA. **Next steps** Next steps from this PR (higher in the stack) include registering the private API `_varlen_attn` as a custom op, implementing backward support, and enabling cuDNN with correct numerics. (This stack builds on top of #162326) [ghstack-poisoned]
**Summary** Today, the only way to have variable sequence length support in PyTorch attention is through nested tensors [here](https://docs.pytorch.org/tutorials/intermediate/scaled_dot_product_attention_tutorial.html#nestedtensor-and-dense-tensor-support). We also want to add an explicit lower-level API that provides variable sequence length support without padding/masking in SDPA. This PR builds out `varlen_attn`, the public API that users can call for the forward method, and `_varlen_attn`, the private API that calls into the Flash Attention/cuDNN backend. **Benchmarking** To benchmark, we compare runtime and TFLOPs against the current SDPA approach with padding. Settings: - 1 H100 machine - `batch_size=8`, `max_seq_len=2048`, `embed_dim=1024`, `num_heads=16` - dtype `torch.bfloat16` - `is_causal=False` - for variable length, we set sequences to be random multiples of 64 up to `max_seq_len` - 100 runs | | Variable Length API | SDPA | |--------|--------------------|----------| | Runtime | 0.21750560760498047 ms | 0.43171775817871094 ms | | TFLOPs | 231.812 | 320.840 | The sparsity is 0.453 which we can see matches the speedup we get from Varlen (approx 50%). TFLOPs remains around the same, with SDPA slightly larger due to potential higher overhead and total flops scaling with sequence length. **Testing** Run `python test/test_varlen_attention.py` for unit tests where we verify basic functionality and confirm numerical match between varlen outputs vs SDPA. **Next steps** Next steps from this PR (higher in the stack) include registering the private API `_varlen_attn` as a custom op, implementing backward support, and enabling cuDNN with correct numerics. (This stack builds on top of #162326) [ghstack-poisoned]
drisspg
added
release notes: nn
[ghstack-poisoned]
drisspg
reviewed
Oct 10, 2025
v0i0
approved these changes
Oct 10, 2025
**Summary** Today, the only way to have variable sequence length support in PyTorch attention is through nested tensors [here](https://docs.pytorch.org/tutorials/intermediate/scaled_dot_product_attention_tutorial.html#nestedtensor-and-dense-tensor-support). We also want to add an explicit lower-level API that provides variable sequence length support without padding/masking in SDPA. This PR builds out `varlen_attn`, the public API that users can call for the forward method, and `_varlen_attn`, the private API that calls into the Flash Attention/cuDNN backend. **Benchmarking** To benchmark, we compare runtime and TFLOPs against the current SDPA approach with padding. Settings: - 1 H100 machine - `batch_size=8`, `max_seq_len=2048`, `embed_dim=1024`, `num_heads=16` - dtype `torch.bfloat16` - `is_causal=False` - for variable length, we set sequences to be random multiples of 64 up to `max_seq_len` - 100 runs | | Variable Length API | SDPA | |--------|--------------------|----------| | Runtime | 0.21750560760498047 ms | 0.43171775817871094 ms | | TFLOPs | 231.812 | 320.840 | The sparsity is 0.453 which we can see matches the speedup we get from Varlen (approx 50%). TFLOPs remains around the same, with SDPA slightly larger due to potential higher overhead and total flops scaling with sequence length. **Testing** Run `python test/test_varlen_attention.py` for unit tests where we verify basic functionality and confirm numerical match between varlen outputs vs SDPA. **Next steps** Next steps from this PR (higher in the stack) include registering the private API `_varlen_attn` as a custom op, implementing backward support, and enabling cuDNN with correct numerics. (This stack builds on top of #162326) [ghstack-poisoned]
zhudada0120
pushed a commit
to zhudada0120/pytorch
that referenced
this pull request
Oct 15, 2025
**Summary** Today, the only way to have variable sequence length support in PyTorch attention is through nested tensors [here](https://docs.pytorch.org/tutorials/intermediate/scaled_dot_product_attention_tutorial.html#nestedtensor-and-dense-tensor-support). We also want to add an explicit lower-level API that provides variable sequence length support without padding/masking in SDPA. This PR builds out `varlen_attn`, the public API that users can call for the forward method, and `_varlen_attn`, the private API that calls into the Flash Attention/cuDNN backend. **Benchmarking** To benchmark, we compare runtime and TFLOPs against the current SDPA approach with padding. Settings: - 1 H100 machine - `batch_size=8`, `max_seq_len=2048`, `embed_dim=1024`, `num_heads=16` - dtype `torch.bfloat16` - `is_causal=False` - for variable length, we set sequences to be random multiples of 64 up to `max_seq_len` - 100 runs | | Variable Length API | SDPA | |--------|--------------------|----------| | Runtime | 0.21750560760498047 ms | 0.43171775817871094 ms | | TFLOPs | 231.812 | 320.840 | The sparsity is 0.453 which we can see matches the speedup we get from Varlen (approx 50%). TFLOPs remains around the same, with SDPA slightly larger due to potential higher overhead and total flops scaling with sequence length. **Testing** Run `python test/test_varlen_attention.py` for unit tests where we verify basic functionality and confirm numerical match between varlen outputs vs SDPA. **Next steps** Next steps from this PR (higher in the stack) include registering the private API `_varlen_attn` as a custom op, implementing backward support, and enabling cuDNN with correct numerics. (This stack builds on top of pytorch#162326) Pull Request resolved: pytorch#164502 Approved by: https://github.com/v0i0, https://github.com/drisspg
Contributor
|
@pytorchbot revert -m 'Sorry for reverting your change, but the doctests failure is legit' -c weird Due to a macOS outage we had during the weekend, |
Collaborator
|
@pytorchbot successfully started a revert job. Check the current status here. |
Collaborator
|
@liangel-02 your PR has been successfully reverted. |
pytorchmergebot
added a commit
that referenced
this pull request
Oct 15, 2025
This reverts commit 3681312. Reverted #164502 on behalf of https://github.com/huydhn due to Sorry for reverting your change, but the doctests failure is legit ([comment](#164502 (comment)))
|
@pytorchbot revert -m "Reverted automatically by pytorch's autorevert, to avoid this behaviour add the tag autorevert: disable" -c autorevert This PR is attributed to have caused regression in:
Please investigate and fix the issues. |
Collaborator
|
@pytorchbot successfully started a revert job. Check the current status here. |
Collaborator
Reverting PR 164502 failedReason: list index out of range Details for Dev Infra teamRaised by workflow job |
**Summary** Today, the only way to have variable sequence length support in PyTorch attention is through nested tensors [here](https://docs.pytorch.org/tutorials/intermediate/scaled_dot_product_attention_tutorial.html#nestedtensor-and-dense-tensor-support). We also want to add an explicit lower-level API that provides variable sequence length support without padding/masking in SDPA. This PR builds out `varlen_attn`, the public API that users can call for the forward method, and `_varlen_attn`, the private API that calls into the Flash Attention/cuDNN backend. **Benchmarking** To benchmark, we compare runtime and TFLOPs against the current SDPA approach with padding. Settings: - 1 H100 machine - `batch_size=8`, `max_seq_len=2048`, `embed_dim=1024`, `num_heads=16` - dtype `torch.bfloat16` - `is_causal=False` - for variable length, we set sequences to be random multiples of 64 up to `max_seq_len` - 100 runs | | Variable Length API | SDPA | |--------|--------------------|----------| | Runtime | 0.21750560760498047 ms | 0.43171775817871094 ms | | TFLOPs | 231.812 | 320.840 | The sparsity is 0.453 which we can see matches the speedup we get from Varlen (approx 50%). TFLOPs remains around the same, with SDPA slightly larger due to potential higher overhead and total flops scaling with sequence length. **Testing** Run `python test/test_varlen_attention.py` for unit tests where we verify basic functionality and confirm numerical match between varlen outputs vs SDPA. **Next steps** Next steps from this PR (higher in the stack) include registering the private API `_varlen_attn` as a custom op, implementing backward support, and enabling cuDNN with correct numerics. (This stack builds on top of #162326) [ghstack-poisoned]
**Summary** Today, the only way to have variable sequence length support in PyTorch attention is through nested tensors [here](https://docs.pytorch.org/tutorials/intermediate/scaled_dot_product_attention_tutorial.html#nestedtensor-and-dense-tensor-support). We also want to add an explicit lower-level API that provides variable sequence length support without padding/masking in SDPA. This PR builds out `varlen_attn`, the public API that users can call for the forward method, and `_varlen_attn`, the private API that calls into the Flash Attention/cuDNN backend. **Benchmarking** To benchmark, we compare runtime and TFLOPs against the current SDPA approach with padding. Settings: - 1 H100 machine - `batch_size=8`, `max_seq_len=2048`, `embed_dim=1024`, `num_heads=16` - dtype `torch.bfloat16` - `is_causal=False` - for variable length, we set sequences to be random multiples of 64 up to `max_seq_len` - 100 runs | | Variable Length API | SDPA | |--------|--------------------|----------| | Runtime | 0.21750560760498047 ms | 0.43171775817871094 ms | | TFLOPs | 231.812 | 320.840 | The sparsity is 0.453 which we can see matches the speedup we get from Varlen (approx 50%). TFLOPs remains around the same, with SDPA slightly larger due to potential higher overhead and total flops scaling with sequence length. **Testing** Run `python test/test_varlen_attention.py` for unit tests where we verify basic functionality and confirm numerical match between varlen outputs vs SDPA. **Next steps** Next steps from this PR (higher in the stack) include registering the private API `_varlen_attn` as a custom op, implementing backward support, and enabling cuDNN with correct numerics. (This stack builds on top of #162326) [ghstack-poisoned]
**Summary** Today, the only way to have variable sequence length support in PyTorch attention is through nested tensors [here](https://docs.pytorch.org/tutorials/intermediate/scaled_dot_product_attention_tutorial.html#nestedtensor-and-dense-tensor-support). We also want to add an explicit lower-level API that provides variable sequence length support without padding/masking in SDPA. This PR builds out `varlen_attn`, the public API that users can call for the forward method, and `_varlen_attn`, the private API that calls into the Flash Attention/cuDNN backend. **Benchmarking** To benchmark, we compare runtime and TFLOPs against the current SDPA approach with padding. Settings: - 1 H100 machine - `batch_size=8`, `max_seq_len=2048`, `embed_dim=1024`, `num_heads=16` - dtype `torch.bfloat16` - `is_causal=False` - for variable length, we set sequences to be random multiples of 64 up to `max_seq_len` - 100 runs | | Variable Length API | SDPA | |--------|--------------------|----------| | Runtime | 0.21750560760498047 ms | 0.43171775817871094 ms | | TFLOPs | 231.812 | 320.840 | The sparsity is 0.453 which we can see matches the speedup we get from Varlen (approx 50%). TFLOPs remains around the same, with SDPA slightly larger due to potential higher overhead and total flops scaling with sequence length. **Testing** Run `python test/test_varlen_attention.py` for unit tests where we verify basic functionality and confirm numerical match between varlen outputs vs SDPA. **Next steps** Next steps from this PR (higher in the stack) include registering the private API `_varlen_attn` as a custom op, implementing backward support, and enabling cuDNN with correct numerics. (This stack builds on top of #162326) [ghstack-poisoned]
**Summary** Today, the only way to have variable sequence length support in PyTorch attention is through nested tensors [here](https://docs.pytorch.org/tutorials/intermediate/scaled_dot_product_attention_tutorial.html#nestedtensor-and-dense-tensor-support). We also want to add an explicit lower-level API that provides variable sequence length support without padding/masking in SDPA. This PR builds out `varlen_attn`, the public API that users can call for the forward method, and `_varlen_attn`, the private API that calls into the Flash Attention/cuDNN backend. **Benchmarking** To benchmark, we compare runtime and TFLOPs against the current SDPA approach with padding. Settings: - 1 H100 machine - `batch_size=8`, `max_seq_len=2048`, `embed_dim=1024`, `num_heads=16` - dtype `torch.bfloat16` - `is_causal=False` - for variable length, we set sequences to be random multiples of 64 up to `max_seq_len` - 100 runs | | Variable Length API | SDPA | |--------|--------------------|----------| | Runtime | 0.21750560760498047 ms | 0.43171775817871094 ms | | TFLOPs | 231.812 | 320.840 | The sparsity is 0.453 which we can see matches the speedup we get from Varlen (approx 50%). TFLOPs remains around the same, with SDPA slightly larger due to potential higher overhead and total flops scaling with sequence length. **Testing** Run `python test/test_varlen_attention.py` for unit tests where we verify basic functionality and confirm numerical match between varlen outputs vs SDPA. **Next steps** Next steps from this PR (higher in the stack) include registering the private API `_varlen_attn` as a custom op, implementing backward support, and enabling cuDNN with correct numerics. (This stack builds on top of #162326) [ghstack-poisoned]
Contributor
Author
|
@pytorchbot 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 |
Chao1Han
pushed a commit
to Chao1Han/pytorch
that referenced
this pull request
Oct 21, 2025
**Summary** Today, the only way to have variable sequence length support in PyTorch attention is through nested tensors [here](https://docs.pytorch.org/tutorials/intermediate/scaled_dot_product_attention_tutorial.html#nestedtensor-and-dense-tensor-support). We also want to add an explicit lower-level API that provides variable sequence length support without padding/masking in SDPA. This PR builds out `varlen_attn`, the public API that users can call for the forward method, and `_varlen_attn`, the private API that calls into the Flash Attention/cuDNN backend. **Benchmarking** To benchmark, we compare runtime and TFLOPs against the current SDPA approach with padding. Settings: - 1 H100 machine - `batch_size=8`, `max_seq_len=2048`, `embed_dim=1024`, `num_heads=16` - dtype `torch.bfloat16` - `is_causal=False` - for variable length, we set sequences to be random multiples of 64 up to `max_seq_len` - 100 runs | | Variable Length API | SDPA | |--------|--------------------|----------| | Runtime | 0.21750560760498047 ms | 0.43171775817871094 ms | | TFLOPs | 231.812 | 320.840 | The sparsity is 0.453 which we can see matches the speedup we get from Varlen (approx 50%). TFLOPs remains around the same, with SDPA slightly larger due to potential higher overhead and total flops scaling with sequence length. **Testing** Run `python test/test_varlen_attention.py` for unit tests where we verify basic functionality and confirm numerical match between varlen outputs vs SDPA. **Next steps** Next steps from this PR (higher in the stack) include registering the private API `_varlen_attn` as a custom op, implementing backward support, and enabling cuDNN with correct numerics. (This stack builds on top of pytorch#162326) Pull Request resolved: pytorch#164502 Approved by: https://github.com/v0i0, https://github.com/drisspg
Chao1Han
pushed a commit
to Chao1Han/pytorch
that referenced
this pull request
Oct 21, 2025
This reverts commit 3681312. Reverted pytorch#164502 on behalf of https://github.com/huydhn due to Sorry for reverting your change, but the doctests failure is legit ([comment](pytorch#164502 (comment)))
Chao1Han
pushed a commit
to Chao1Han/pytorch
that referenced
this pull request
Oct 21, 2025
**Summary** Today, the only way to have variable sequence length support in PyTorch attention is through nested tensors [here](https://docs.pytorch.org/tutorials/intermediate/scaled_dot_product_attention_tutorial.html#nestedtensor-and-dense-tensor-support). We also want to add an explicit lower-level API that provides variable sequence length support without padding/masking in SDPA. This PR builds out `varlen_attn`, the public API that users can call for the forward method, and `_varlen_attn`, the private API that calls into the Flash Attention/cuDNN backend. **Benchmarking** To benchmark, we compare runtime and TFLOPs against the current SDPA approach with padding. Settings: - 1 H100 machine - `batch_size=8`, `max_seq_len=2048`, `embed_dim=1024`, `num_heads=16` - dtype `torch.bfloat16` - `is_causal=False` - for variable length, we set sequences to be random multiples of 64 up to `max_seq_len` - 100 runs | | Variable Length API | SDPA | |--------|--------------------|----------| | Runtime | 0.21750560760498047 ms | 0.43171775817871094 ms | | TFLOPs | 231.812 | 320.840 | The sparsity is 0.453 which we can see matches the speedup we get from Varlen (approx 50%). TFLOPs remains around the same, with SDPA slightly larger due to potential higher overhead and total flops scaling with sequence length. **Testing** Run `python test/test_varlen_attention.py` for unit tests where we verify basic functionality and confirm numerical match between varlen outputs vs SDPA. **Next steps** Next steps from this PR (higher in the stack) include registering the private API `_varlen_attn` as a custom op, implementing backward support, and enabling cuDNN with correct numerics. (This stack builds on top of pytorch#162326) Pull Request resolved: pytorch#164502 Approved by: https://github.com/v0i0, https://github.com/drisspg
Khanaksahu
pushed a commit
to Khanaksahu/pytorch
that referenced
this pull request
Nov 17, 2025
ghstack-source-id: 19289b7 Pull Request resolved: pytorch/pytorch#164502
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
9 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.
Summary
Today, the only way to have variable sequence length support in PyTorch attention is through nested tensors here. We also want to add an explicit lower-level API that provides variable sequence length support without padding/masking in SDPA.
This PR builds out
varlen_attn, the public API that users can call for the forward method, and_varlen_attn, the private API that calls into the Flash Attention/cuDNN backend.Benchmarking
To benchmark, we compare runtime and TFLOPs against the current SDPA approach with padding.
Settings:
batch_size=8,max_seq_len=2048,embed_dim=1024,num_heads=16torch.bfloat16is_causal=Falsemax_seq_lenThe sparsity is 0.453 which we can see matches the speedup we get from Varlen (approx 50%). TFLOPs remains around the same, with SDPA slightly larger due to potential higher overhead and total flops scaling with sequence length.
Testing
Run
python test/test_varlen_attention.pyfor unit tests where we verify basic functionality and confirm numerical match between varlen outputs vs SDPA.Next steps
Next steps from this PR (higher in the stack) include registering the private API
_varlen_attnas a custom op, implementing backward support, and enabling cuDNN with correct numerics.Stack from ghstack (oldest at bottom):
(This stack builds on top of #162326)