Compute the mask in-place, with less memory reads, and on CUDA on XLNetLMHeadModel
#23332
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
…NetLMHeadModel` When working on TorchInductor, I realised that there was a part from `XLNetLMHeadModel` that was being compiled to CPU code. This PR should allow to fuse this operation with other CUDA operations in `torch.compile`. It also should be faster on eager mode, as it has a this implementation has a lower foot-print. If in-place operations are not allowed even in non-grad context, I still believe that doing ones + tril rather than a ones + tril + zeros + cat should be faster simply due to the number of memory reads/writes. I tested that this code produces the same results for `0 <= qlen,mlen < 10` and `same_length in (True, False)`.
|
The documentation is not available anymore as the PR was closed or merged. |
younesbelkada
approved these changes
May 12, 2023
amyeroberts
approved these changes
May 12, 2023
sheonhan
pushed a commit
to sheonhan/transformers
that referenced
this pull request
May 15, 2023
…NetLMHeadModel` (huggingface#23332) When working on TorchInductor, I realised that there was a part from `XLNetLMHeadModel` that was being compiled to CPU code. This PR should allow to fuse this operation with other CUDA operations in `torch.compile`. It also should be faster on eager mode, as it has a this implementation has a lower foot-print. If in-place operations are not allowed even in non-grad context, I still believe that doing ones + tril rather than a ones + tril + zeros + cat should be faster simply due to the number of memory reads/writes. I tested that this code produces the same results for `0 <= qlen,mlen < 10` and `same_length in (True, False)`.
gojiteji
pushed a commit
to gojiteji/transformers
that referenced
this pull request
Jun 5, 2023
…NetLMHeadModel` (huggingface#23332) When working on TorchInductor, I realised that there was a part from `XLNetLMHeadModel` that was being compiled to CPU code. This PR should allow to fuse this operation with other CUDA operations in `torch.compile`. It also should be faster on eager mode, as it has a this implementation has a lower foot-print. If in-place operations are not allowed even in non-grad context, I still believe that doing ones + tril rather than a ones + tril + zeros + cat should be faster simply due to the number of memory reads/writes. I tested that this code produces the same results for `0 <= qlen,mlen < 10` and `same_length in (True, False)`.
novice03
pushed a commit
to novice03/transformers
that referenced
this pull request
Jun 23, 2023
…NetLMHeadModel` (huggingface#23332) When working on TorchInductor, I realised that there was a part from `XLNetLMHeadModel` that was being compiled to CPU code. This PR should allow to fuse this operation with other CUDA operations in `torch.compile`. It also should be faster on eager mode, as it has a this implementation has a lower foot-print. If in-place operations are not allowed even in non-grad context, I still believe that doing ones + tril rather than a ones + tril + zeros + cat should be faster simply due to the number of memory reads/writes. I tested that this code produces the same results for `0 <= qlen,mlen < 10` and `same_length in (True, False)`.
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.
When working on TorchInductor, I realised that there was a part from
XLNetLMHeadModelthat was being compiled to CPU code.This PR should allow to fuse this operation with other CUDA operations in
torch.compile. It also should be faster on eager mode, as it has a this implementation has a lower foot-print.If in-place operations are not allowed even in non-grad context, I still believe that doing ones + tril rather than a ones + tril + zeros + cat should be faster simply due to the number of memory reads/writes.
I tested that this code produces the same results for
0 <= qlen,mlen < 10andsame_length in (True, False).@ArthurZucker @younesbelkada