Separate output and accumulator type for Flash Attention Prefill #443
Conversation
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
muhammad-tanvir-1211
force-pushed
the
flash_prefill_separate_out_type
branch
from
e948f6d to
f64c7b7
Compare
aacostadiaz
approved these changes
Jun 26, 2025
This was referenced Jun 26, 2025
joeatodd
approved these changes
Jun 27, 2025
|
|
||
| std::vector<ElementAccumulator> vec_acc(block_acc.size()); | ||
| syclcompat::memcpy<ElementAccumulator>(vec_acc.data(), block_acc.get(), vec_acc.size()); | ||
| syclcompat::wait(); |
There was a problem hiding this comment.
syclcompat::memcpy is synchronous/blocking so the waits here are unneeded.
|
|
||
| std::vector<ElementAccumulator> vec_acc(block_acc.size()); | ||
| syclcompat::memcpy<ElementAccumulator>(vec_acc.data(), block_acc.get(), vec_acc.size()); | ||
| syclcompat::wait(); |
| vec_out[i] = static_cast<ElementOutput>(vec_acc[i]); | ||
| } | ||
| syclcompat::memcpy<ElementOutput>(block_ref_O.get() + offset_o, vec_out.data(), vec_out.size()); | ||
| syclcompat::wait(); |
| syclcompat::memcpy<ElementOutput>(host_S.data(), block_S.get(), host_S.size()); | ||
| std::vector<ElementAccumulator> host_S(block_S.size()); | ||
| syclcompat::memcpy<ElementAccumulator>(host_S.data(), block_S.get(), host_S.size()); | ||
| syclcompat::wait(); |
|
|
||
| std::vector<ElementAccumulator> vec_acc(block_acc.size()); | ||
| syclcompat::memcpy<ElementAccumulator>(vec_acc.data(), block_acc.get(), vec_acc.size()); | ||
| syclcompat::wait(); |
| vec_out[i] = static_cast<ElementOutput>(vec_acc[i]); | ||
| } | ||
| syclcompat::memcpy<ElementOutput>(block_ref_O.get() + offset_o, vec_out.data(), vec_out.size()); | ||
| syclcompat::wait(); |
joeatodd
reviewed
Jun 27, 2025
|
|
||
| copy(params.xe_store_o, out_reg, tOgO); | ||
| Tensor final_out_reg = make_fragment_like<ElementOutput>(out_reg); | ||
| if constexpr (cute::is_any_of_v<ElementOutput, cute::float_e5m2_t, cute::float_e4m3_t> || cute::is_same_v<ElementOutput, ElementCompute>) { |
There was a problem hiding this comment.
Could you add a comment here to explain this if condition?
There was a problem hiding this comment.
Added this comment, not sure why it doesn't show up here.
* Add comment on final output type conversion
aacostadiaz
pushed a commit
that referenced
this pull request
Jun 30, 2025
This PR adds tests for all the different data types supported with Flash Attention Prefill. It is a continuation of PR #443
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.
This PR separates the output type and accumulator type for Flash Attention Prefill. Combinations supported are:
Tests added in: #446
Benchmarks added in: #447