WIP: DP4AMatMul fix matmul for subgoup size 64 GPUs #23637
Merged
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
Contributor
Author
|
@qjia7 - FYI, I am not able to add you as a reviewer but wanted to share for awareness. |
qjia7
reviewed
Feb 11, 2025
sushraja-msft
changed the title
sushraja-msft
force-pushed
the
user/sushraja/fix_dp4_quantization
branch
2 times, most recently
from
1ab5a6a to
1decc48
Compare
sushraja-msft
force-pushed
the
user/sushraja/fix_dp4_quantization
branch
from
1decc48 to
4f473cb
Compare
qjia7
reviewed
Feb 13, 2025
guschmue
approved these changes
Feb 13, 2025
guschmue
pushed a commit
that referenced
this pull request
Mar 6, 2025
### Description This change moves away from using subgroup ops for quantization. This is because on AMD GPUs subgroup size is 64 and that is not handled in our quantization function, resulting in garbage output. Implementing subgroup size 64 quantization requires changing the workgroup size and then implementing support for subgroup size 128 becomes a challenge. With the new implementation perf on intel ALD remains about the same 4.36s for 1000K prefill. Tests for this change are present here https://github.com/microsoft/onnxruntime/blob/e66650350b85cb5e3a408f6576fe6a7f4f4ddebc/onnxruntime/test/contrib_ops/matmul_4bits_test.cc However, to trigger the current issue they must be run on a GPU with subgroup size 64.
ashrit-ms
pushed a commit
that referenced
this pull request
Mar 17, 2025
### Description This change moves away from using subgroup ops for quantization. This is because on AMD GPUs subgroup size is 64 and that is not handled in our quantization function, resulting in garbage output. Implementing subgroup size 64 quantization requires changing the workgroup size and then implementing support for subgroup size 128 becomes a challenge. With the new implementation perf on intel ALD remains about the same 4.36s for 1000K prefill. Tests for this change are present here https://github.com/microsoft/onnxruntime/blob/e66650350b85cb5e3a408f6576fe6a7f4f4ddebc/onnxruntime/test/contrib_ops/matmul_4bits_test.cc However, to trigger the current issue they must be run on a GPU with subgroup size 64.
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.
Description
This change moves away from using subgroup ops for quantization. This is because on AMD GPUs subgroup size is 64 and that is not handled in our quantization function, resulting in garbage output. Implementing subgroup size 64 quantization requires changing the workgroup size and then implementing support for subgroup size 128 becomes a challenge.
With the new implementation perf on intel ALD remains about the same 4.36s for 1000K prefill.
Tests for this change are present here
https://github.com/microsoft/onnxruntime/blob/e66650350b85cb5e3a408f6576fe6a7f4f4ddebc/onnxruntime/test/contrib_ops/matmul_4bits_test.cc
However, to trigger the current issue they must be run on a GPU with subgroup size 64.