optimization of perKVhead quantization #4161
Closed
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
✅ Deploy Preview for pytorch-fbgemm-docs ready!
To edit notification comments on pull requests, go to your Netlify project configuration. |
|
This pull request was exported from Phabricator. Differential Revision: D74924275 |
Summary: as title. This is needed to handle this case: https://www.internalfb.com/diff/D73833204?dst_version_fbid=9500286030082255&transaction_fbid=676020828512263 This will help avoid amax calc in rope for decode and partial prefill batch lanes. Also, we can rely on it in Kernel2, to return back and avoid unneccessary quantization. Differential Revision: D73478483 Reviewed By: y-sq
Aya-ZIbra
force-pushed
the
export-D74924275
branch
from
5182026 to
8e14af4
Compare
Aya-ZIbra
added a commit
to Aya-ZIbra/FBGEMM
that referenced
this pull request
May 21, 2025
Summary: X-link: facebookresearch/FBGEMM#1241 y-sq noticed that for prefill chunk of 64k, the improvement in attention kernel runtime for local layers is cancelled out by around 0.4 ms overhead from the quantization kernel. https://docs.google.com/document/d/193GL7o5GMlpVlwEDVxqoDB6O85zDuS8A5-PUOSsZc1s/edit?tab=t.0#bookmark=id.zh92spta1uxw Before: BS =1 , Seqlen = 64k Elapsed Cycles cycle 530,268 Memory Throughput % 17.24 Duration us 392.93 After: ----------------------- ----------- ------------ DRAM Frequency Ghz 1.59 SM Frequency Ghz 1.34 Elapsed Cycles cycle 192,884 Memory Throughput % 46.01 DRAM Throughput % 46.01 Duration us 143.23 L1/TEX Cache Throughput % 15.15 L2 Cache Throughput % 39.31 SM Active Cycles cycle 181,953.16 Compute (SM) Throughput % 71.92 ----------------------- ----------- ------------ Reviewed By: y-sq Differential Revision: D74924275
Aya-ZIbra
added a commit
to Aya-ZIbra/FBGEMM
that referenced
this pull request
May 21, 2025
Summary: X-link: facebookresearch/FBGEMM#1241 y-sq noticed that for prefill chunk of 64k, the improvement in attention kernel runtime for local layers is cancelled out by around 0.4 ms overhead from the quantization kernel. https://docs.google.com/document/d/193GL7o5GMlpVlwEDVxqoDB6O85zDuS8A5-PUOSsZc1s/edit?tab=t.0#bookmark=id.zh92spta1uxw Before: BS =1 , Seqlen = 64k Elapsed Cycles cycle 530,268 Memory Throughput % 17.24 Duration us 392.93 After: ----------------------- ----------- ------------ DRAM Frequency Ghz 1.59 SM Frequency Ghz 1.34 Elapsed Cycles cycle 192,884 Memory Throughput % 46.01 DRAM Throughput % 46.01 Duration us 143.23 L1/TEX Cache Throughput % 15.15 L2 Cache Throughput % 39.31 SM Active Cycles cycle 181,953.16 Compute (SM) Throughput % 71.92 ----------------------- ----------- ------------ Reviewed By: y-sq Differential Revision: D74924275
Aya-ZIbra
force-pushed
the
export-D74924275
branch
from
8e14af4 to
2779dd7
Compare
|
This pull request was exported from Phabricator. Differential Revision: D74924275 |
Aya-ZIbra
added a commit
to Aya-ZIbra/FBGEMM
that referenced
this pull request
May 21, 2025
Summary: Pull Request resolved: pytorch#4161 X-link: facebookresearch/FBGEMM#1241 y-sq noticed that for prefill chunk of 64k, the improvement in attention kernel runtime for local layers is cancelled out by around 0.4 ms overhead from the quantization kernel. https://docs.google.com/document/d/193GL7o5GMlpVlwEDVxqoDB6O85zDuS8A5-PUOSsZc1s/edit?tab=t.0#bookmark=id.zh92spta1uxw Before: BS =1 , Seqlen = 64k Elapsed Cycles cycle 530,268 Memory Throughput % 17.24 Duration us 392.93 After: ----------------------- ----------- ------------ DRAM Frequency Ghz 1.59 SM Frequency Ghz 1.34 Elapsed Cycles cycle 192,884 Memory Throughput % 46.01 DRAM Throughput % 46.01 Duration us 143.23 L1/TEX Cache Throughput % 15.15 L2 Cache Throughput % 39.31 SM Active Cycles cycle 181,953.16 Compute (SM) Throughput % 71.92 ----------------------- ----------- ------------ Reviewed By: y-sq Differential Revision: D74924275
Aya-ZIbra
force-pushed
the
export-D74924275
branch
from
2779dd7 to
f2a130f
Compare
Summary: Pull Request resolved: pytorch#4161 X-link: facebookresearch/FBGEMM#1241 y-sq noticed that for prefill chunk of 64k, the improvement in attention kernel runtime for local layers is cancelled out by around 0.4 ms overhead from the quantization kernel. https://docs.google.com/document/d/193GL7o5GMlpVlwEDVxqoDB6O85zDuS8A5-PUOSsZc1s/edit?tab=t.0#bookmark=id.zh92spta1uxw Before: BS =1 , Seqlen = 64k Elapsed Cycles cycle 530,268 Memory Throughput % 17.24 Duration us 392.93 After: ----------------------- ----------- ------------ DRAM Frequency Ghz 1.59 SM Frequency Ghz 1.34 Elapsed Cycles cycle 192,884 Memory Throughput % 46.01 DRAM Throughput % 46.01 Duration us 143.23 L1/TEX Cache Throughput % 15.15 L2 Cache Throughput % 39.31 SM Active Cycles cycle 181,953.16 Compute (SM) Throughput % 71.92 ----------------------- ----------- ------------ Reviewed By: y-sq Differential Revision: D74924275
|
This pull request was exported from Phabricator. Differential Revision: D74924275 |
Aya-ZIbra
force-pushed
the
export-D74924275
branch
from
f2a130f to
6005e07
Compare
|
This pull request has been merged in aa2fe3d. |
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.
Summary:
y-sq noticed that for prefill chunk of 64k, the improvement in attention kernel runtime for local layers is cancelled out by around 0.4 ms overhead from the quantization kernel.
https://docs.google.com/document/d/193GL7o5GMlpVlwEDVxqoDB6O85zDuS8A5-PUOSsZc1s/edit?tab=t.0#bookmark=id.zh92spta1uxw
Before:
BS =1 , Seqlen = 64k
Elapsed Cycles cycle 530,268
Memory Throughput % 17.24
Duration us 392.93
After:
----------------------- ----------- ------------
DRAM Frequency Ghz 1.59
SM Frequency Ghz 1.34
Elapsed Cycles cycle 192,884
Memory Throughput % 46.01
DRAM Throughput % 46.01
Duration us 143.23
L1/TEX Cache Throughput % 15.15
L2 Cache Throughput % 39.31
SM Active Cycles cycle 181,953.16
Compute (SM) Throughput % 71.92
----------------------- ----------- ------------
Reviewed By: y-sq
Differential Revision: D74924275