Add an attention bias subclass for a lower right causal masking#114823
Closed
drisspg wants to merge 26 commits intogh/drisspg/12/basefrom
Closed
Add an attention bias subclass for a lower right causal masking#114823drisspg wants to merge 26 commits intogh/drisspg/12/basefrom
drisspg wants to merge 26 commits intogh/drisspg/12/basefrom
Conversation
[ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/114823
Note: Links to docs will display an error until the docs builds have been completed. ✅ You can merge normally! (1 Unrelated Failure)As of commit 80b03db with merge base 597d3fb ( FLAKY - The following job failed but was likely due to flakiness present on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This was referenced Nov 30, 2023
Results from benchmarking ```Shell +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Type | Speedup | batch_size | num_heads | q_seq_len | k_seq_len | embed_dim | dtype | head_dim | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Average | 1.4320973389745877 | | | | | | | | | Max | 2.577393980674173 | 128 | 32 | 512 | 4097 | 2048 | torch.bfloat16 | 64 | | Min | 0.942474845104863 | 1 | 16 | 256 | 416 | 2048 | torch.bfloat16 | 128 | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ ``` [ghstack-poisoned]
Results from benchmarking ```Shell +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Type | Speedup | batch_size | num_heads | q_seq_len | k_seq_len | embed_dim | dtype | head_dim | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Average | 1.4320973389745877 | | | | | | | | | Max | 2.577393980674173 | 128 | 32 | 512 | 4097 | 2048 | torch.bfloat16 | 64 | | Min | 0.942474845104863 | 1 | 16 | 256 | 416 | 2048 | torch.bfloat16 | 128 | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ ``` [ghstack-poisoned]
cpuhrsch
approved these changes
Dec 1, 2023
Results from benchmarking ```Shell +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Type | Speedup | batch_size | num_heads | q_seq_len | k_seq_len | embed_dim | dtype | head_dim | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Average | 1.4320973389745877 | | | | | | | | | Max | 2.577393980674173 | 128 | 32 | 512 | 4097 | 2048 | torch.bfloat16 | 64 | | Min | 0.942474845104863 | 1 | 16 | 256 | 416 | 2048 | torch.bfloat16 | 128 | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ ``` [ghstack-poisoned]
cpuhrsch
reviewed
Dec 1, 2023
Results from benchmarking ```Shell +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Type | Speedup | batch_size | num_heads | q_seq_len | k_seq_len | embed_dim | dtype | head_dim | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Average | 1.4320973389745877 | | | | | | | | | Max | 2.577393980674173 | 128 | 32 | 512 | 4097 | 2048 | torch.bfloat16 | 64 | | Min | 0.942474845104863 | 1 | 16 | 256 | 416 | 2048 | torch.bfloat16 | 128 | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ ``` [ghstack-poisoned]
drisspg
commented
Dec 2, 2023
Results from benchmarking ```Shell +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Type | Speedup | batch_size | num_heads | q_seq_len | k_seq_len | embed_dim | dtype | head_dim | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Average | 1.4320973389745877 | | | | | | | | | Max | 2.577393980674173 | 128 | 32 | 512 | 4097 | 2048 | torch.bfloat16 | 64 | | Min | 0.942474845104863 | 1 | 16 | 256 | 416 | 2048 | torch.bfloat16 | 128 | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ ``` [ghstack-poisoned]
Results from benchmarking ```Shell +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Type | Speedup | batch_size | num_heads | q_seq_len | k_seq_len | embed_dim | dtype | head_dim | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Average | 1.4320973389745877 | | | | | | | | | Max | 2.577393980674173 | 128 | 32 | 512 | 4097 | 2048 | torch.bfloat16 | 64 | | Min | 0.942474845104863 | 1 | 16 | 256 | 416 | 2048 | torch.bfloat16 | 128 | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ ``` [ghstack-poisoned]
Results from benchmarking ```Shell +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Type | Speedup | batch_size | num_heads | q_seq_len | k_seq_len | embed_dim | dtype | head_dim | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Average | 1.4320973389745877 | | | | | | | | | Max | 2.577393980674173 | 128 | 32 | 512 | 4097 | 2048 | torch.bfloat16 | 64 | | Min | 0.942474845104863 | 1 | 16 | 256 | 416 | 2048 | torch.bfloat16 | 128 | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ ``` [ghstack-poisoned]
cpuhrsch
reviewed
Dec 2, 2023
cpuhrsch
reviewed
Dec 2, 2023
cpuhrsch
reviewed
Dec 2, 2023
cpuhrsch
reviewed
Dec 2, 2023
Results from benchmarking, I improved the meff_attention perf hence the slightly decreased max perf. ```Shell +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Type | Speedup | batch_size | num_heads | q_seq_len | k_seq_len | embed_dim | dtype | head_dim | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Average | 1.2388050062214226 | | | | | | | | | Max | 1.831672915579016 | 128 | 32 | 1024 | 2048 | 2048 | torch.bfloat16 | 64 | | Min | 0.9430534166730135 | 1 | 16 | 256 | 416 | 2048 | torch.bfloat16 | 128 | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ ``` [ghstack-poisoned]
Results from benchmarking, I improved the meff_attention perf hence the slightly decreased max perf. ```Shell +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Type | Speedup | batch_size | num_heads | q_seq_len | k_seq_len | embed_dim | dtype | head_dim | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Average | 1.2388050062214226 | | | | | | | | | Max | 1.831672915579016 | 128 | 32 | 1024 | 2048 | 2048 | torch.bfloat16 | 64 | | Min | 0.9430534166730135 | 1 | 16 | 256 | 416 | 2048 | torch.bfloat16 | 128 | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ ``` [ghstack-poisoned]
…sking" # Summary This PR introduces a new Tensor subclass that is designed to be used with torch.nn.functional.scaled_dot_product_attention. Currently we have a boolean `is_causal` flag that allows users to do do causal masking without the need to actually create the "realized" attention bias and pass into sdpa. We originally added this flag since there is native support in both fused kernels we support. This provides a big performance gain ( the kernels only need to iterate over ~0.5x the sequence, and for very large sequence lengths this can provide vary large memory improvements. The flag was introduced when the early on in the kernel development and at the time it was implicitly meant to "upper_left" causal attention. This distinction only matters when the attention_bias is not square. For a more detailed break down see: #108108. The kernels default behavior has since changed, largely due to the rise of autogressive text generation. And unfortunately this would lead to a BC break. In the long term it may actually be beneficial to change the default meaning of `is_causal` to represent lower_right causal masking. The larger theme though is laid here: #110681. The thesis being that there is alot of innovation in SDPA revolving around the attention_bias being used. This is the first in hopefully a few more attention_biases that we would like to add. The next interesting one would be `sliding_window` which is used by the popular mistral model family. Results from benchmarking, I improved the meff_attention perf hence the slightly decreased max perf. ```Shell +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Type | Speedup | batch_size | num_heads | q_seq_len | k_seq_len | embed_dim | dtype | head_dim | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Average | 1.2388050062214226 | | | | | | | | | Max | 1.831672915579016 | 128 | 32 | 1024 | 2048 | 2048 | torch.bfloat16 | 64 | | Min | 0.9430534166730135 | 1 | 16 | 256 | 416 | 2048 | torch.bfloat16 | 128 | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ ``` [ghstack-poisoned]
…sking" # Summary This PR introduces a new Tensor subclass that is designed to be used with torch.nn.functional.scaled_dot_product_attention. Currently we have a boolean `is_causal` flag that allows users to do do causal masking without the need to actually create the "realized" attention bias and pass into sdpa. We originally added this flag since there is native support in both fused kernels we support. This provides a big performance gain ( the kernels only need to iterate over ~0.5x the sequence, and for very large sequence lengths this can provide vary large memory improvements. The flag was introduced when the early on in the kernel development and at the time it was implicitly meant to "upper_left" causal attention. This distinction only matters when the attention_bias is not square. For a more detailed break down see: #108108. The kernels default behavior has since changed, largely due to the rise of autogressive text generation. And unfortunately this would lead to a BC break. In the long term it may actually be beneficial to change the default meaning of `is_causal` to represent lower_right causal masking. The larger theme though is laid here: #110681. The thesis being that there is alot of innovation in SDPA revolving around the attention_bias being used. This is the first in hopefully a few more attention_biases that we would like to add. The next interesting one would be `sliding_window` which is used by the popular mistral model family. Results from benchmarking, I improved the meff_attention perf hence the slightly decreased max perf. ```Shell +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Type | Speedup | batch_size | num_heads | q_seq_len | k_seq_len | embed_dim | dtype | head_dim | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Average | 1.2388050062214226 | | | | | | | | | Max | 1.831672915579016 | 128 | 32 | 1024 | 2048 | 2048 | torch.bfloat16 | 64 | | Min | 0.9430534166730135 | 1 | 16 | 256 | 416 | 2048 | torch.bfloat16 | 128 | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ ``` [ghstack-poisoned]
…sking" # Summary This PR introduces a new Tensor subclass that is designed to be used with torch.nn.functional.scaled_dot_product_attention. Currently we have a boolean `is_causal` flag that allows users to do do causal masking without the need to actually create the "realized" attention bias and pass into sdpa. We originally added this flag since there is native support in both fused kernels we support. This provides a big performance gain ( the kernels only need to iterate over ~0.5x the sequence, and for very large sequence lengths this can provide vary large memory improvements. The flag was introduced when the early on in the kernel development and at the time it was implicitly meant to "upper_left" causal attention. This distinction only matters when the attention_bias is not square. For a more detailed break down see: #108108. The kernels default behavior has since changed, largely due to the rise of autogressive text generation. And unfortunately this would lead to a BC break. In the long term it may actually be beneficial to change the default meaning of `is_causal` to represent lower_right causal masking. The larger theme though is laid here: #110681. The thesis being that there is alot of innovation in SDPA revolving around the attention_bias being used. This is the first in hopefully a few more attention_biases that we would like to add. The next interesting one would be `sliding_window` which is used by the popular mistral model family. Results from benchmarking, I improved the meff_attention perf hence the slightly decreased max perf. ```Shell +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Type | Speedup | batch_size | num_heads | q_seq_len | k_seq_len | embed_dim | dtype | head_dim | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Average | 1.2388050062214226 | | | | | | | | | Max | 1.831672915579016 | 128 | 32 | 1024 | 2048 | 2048 | torch.bfloat16 | 64 | | Min | 0.9430534166730135 | 1 | 16 | 256 | 416 | 2048 | torch.bfloat16 | 128 | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ ``` [ghstack-poisoned]
…sking" # Summary This PR introduces a new Tensor subclass that is designed to be used with torch.nn.functional.scaled_dot_product_attention. Currently we have a boolean `is_causal` flag that allows users to do do causal masking without the need to actually create the "realized" attention bias and pass into sdpa. We originally added this flag since there is native support in both fused kernels we support. This provides a big performance gain ( the kernels only need to iterate over ~0.5x the sequence, and for very large sequence lengths this can provide vary large memory improvements. The flag was introduced when the early on in the kernel development and at the time it was implicitly meant to "upper_left" causal attention. This distinction only matters when the attention_bias is not square. For a more detailed break down see: #108108. The kernels default behavior has since changed, largely due to the rise of autogressive text generation. And unfortunately this would lead to a BC break. In the long term it may actually be beneficial to change the default meaning of `is_causal` to represent lower_right causal masking. The larger theme though is laid here: #110681. The thesis being that there is alot of innovation in SDPA revolving around the attention_bias being used. This is the first in hopefully a few more attention_biases that we would like to add. The next interesting one would be `sliding_window` which is used by the popular mistral model family. Results from benchmarking, I improved the meff_attention perf hence the slightly decreased max perf. ```Shell +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Type | Speedup | batch_size | num_heads | q_seq_len | k_seq_len | embed_dim | dtype | head_dim | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Average | 1.2388050062214226 | | | | | | | | | Max | 1.831672915579016 | 128 | 32 | 1024 | 2048 | 2048 | torch.bfloat16 | 64 | | Min | 0.9430534166730135 | 1 | 16 | 256 | 416 | 2048 | torch.bfloat16 | 128 | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ ``` [ghstack-poisoned]
Contributor
Author
|
@pytorchbot merge |
Collaborator
Merge failedReason: This PR needs a If not, please add the To add a label, you can comment to pytorchbot, for example For more information, see Details for Dev Infra teamRaised by workflow job |
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 |
dmenig
pushed a commit
to dmenig/pytorch
that referenced
this pull request
Dec 21, 2023
…rch#114823) # Summary This PR introduces a new Tensor subclass that is designed to be used with torch.nn.functional.scaled_dot_product_attention. Currently we have a boolean `is_causal` flag that allows users to do do causal masking without the need to actually create the "realized" attention bias and pass into sdpa. We originally added this flag since there is native support in both fused kernels we support. This provides a big performance gain ( the kernels only need to iterate over ~0.5x the sequence, and for very large sequence lengths this can provide vary large memory improvements. The flag was introduced when the early on in the kernel development and at the time it was implicitly meant to "upper_left" causal attention. This distinction only matters when the attention_bias is not square. For a more detailed break down see: pytorch#108108. The kernels default behavior has since changed, largely due to the rise of autogressive text generation. And unfortunately this would lead to a BC break. In the long term it may actually be beneficial to change the default meaning of `is_causal` to represent lower_right causal masking. The larger theme though is laid here: pytorch#110681. The thesis being that there is alot of innovation in SDPA revolving around the attention_bias being used. This is the first in hopefully a few more attention_biases that we would like to add. The next interesting one would be `sliding_window` which is used by the popular mistral model family. Results from benchmarking, I improved the meff_attention perf hence the slightly decreased max perf. ```Shell +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Type | Speedup | batch_size | num_heads | q_seq_len | k_seq_len | embed_dim | dtype | head_dim | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ | Average | 1.2388050062214226 | | | | | | | | | Max | 1.831672915579016 | 128 | 32 | 1024 | 2048 | 2048 | torch.bfloat16 | 64 | | Min | 0.9430534166730135 | 1 | 16 | 256 | 416 | 2048 | torch.bfloat16 | 128 | +---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+ ``` Pull Request resolved: pytorch#114823 Approved by: https://github.com/cpuhrsch
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
3 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
This PR introduces a new Tensor subclass that is designed to be used with torch.nn.functional.scaled_dot_product_attention. Currently we have a boolean
is_causalflag that allows users to do do causal masking without the need to actually create the "realized" attention bias and pass into sdpa. We originally added this flag since there is native support in both fused kernels we support. This provides a big performance gain ( the kernels only need to iterate over ~0.5x the sequence, and for very large sequence lengths this can provide vary large memory improvements.The flag was introduced when the early on in the kernel development and at the time it was implicitly meant to "upper_left" causal attention. This distinction only matters when the attention_bias is not square. For a more detailed break down see: #108108. The kernels default behavior has since changed, largely due to the rise of autogressive text generation. And unfortunately this would lead to a BC break. In the long term it may actually be beneficial to change the default meaning of
is_causalto represent lower_right causal masking.The larger theme though is laid here: #110681. The thesis being that there is alot of innovation in SDPA revolving around the attention_bias being used. This is the first in hopefully a few more attention_biases that we would like to add. The next interesting one would be
sliding_windowwhich is used by the popular mistral model family.Stack from ghstack (oldest at bottom):
Results from benchmarking, I improved the meff_attention perf hence the slightly decreased max perf.
cc @albanD @mruberry @jbschlosser @walterddr @mikaylagawarecki