[ET-VK] Fix implementation of int4 quantized linear #6200
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
## Context Fix the existing implementation of int4 weight quantized linear to conform with how the `_weight_int4packed_mm` op works in the ATen library. For some additional context, the current op implementation does not actually match the behaviour of `_weight_int4packed_mm`. The ATen op expects that the weights have already been packed into a specific format, with `inner_k_tiles` as a packing parameter. The packing is accomplished via calling the `_convert_weight_to_int4pack` operator. Thus the current implementation in vulkan is equivalent to calling `_convert_weight_to_int4pack` + `_weight_int4packed_mm`. To address this discrepancy, the operator implementation is registered under the `linear_weight_int4` custom op as of this diff. The problems with the existing implementation were as follows: * The expected sizes of the scales and zeros tensor was incorrect. Previously, the sizes were assumed to be `(2, N, num_groups)` but the correct size is `(num_groups, N, 2)` * Previously, when unpacking a uint8_t into 2 unpacked int4 values, it was assumed that the LSB was the first value and the MSB was the second value. However, this ordering should be flipped * The original implementation expected the output tensor to be channels packed, but in practice we want the output tensor to be width packed This diff addresses the above issues, and introduces a dedicated test binary to test against an equivalent reference implementation expressed with ATen functions. Differential Revision: [D64354773](https://our.internmc.facebook.com/intern/diff/D64354773/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/6200
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 470ff32 with merge base 8673567 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
SS-JIA
added a commit
that referenced
this pull request
Oct 14, 2024
## Context Fix the existing implementation of int4 weight quantized linear to conform with how the `_weight_int4packed_mm` op works in the ATen library. For some additional context, the current op implementation does not actually match the behaviour of `_weight_int4packed_mm`. The ATen op expects that the weights have already been packed into a specific format, with `inner_k_tiles` as a packing parameter. The packing is accomplished via calling the `_convert_weight_to_int4pack` operator. Thus the current implementation in vulkan is equivalent to calling `_convert_weight_to_int4pack` + `_weight_int4packed_mm`. To address this discrepancy, the operator implementation is registered under the `linear_weight_int4` custom op as of this diff. The problems with the existing implementation were as follows: * The expected sizes of the scales and zeros tensor was incorrect. Previously, the sizes were assumed to be `(2, N, num_groups)` but the correct size is `(num_groups, N, 2)` * Previously, when unpacking a uint8_t into 2 unpacked int4 values, it was assumed that the LSB was the first value and the MSB was the second value. However, this ordering should be flipped * The original implementation expected the output tensor to be channels packed, but in practice we want the output tensor to be width packed This diff addresses the above issues, and introduces a dedicated test binary to test against an equivalent reference implementation expressed with ATen functions. Differential Revision: [D64354773](https://our.internmc.facebook.com/intern/diff/D64354773/) ghstack-source-id: 247944123 Pull Request resolved: #6200
|
This pull request was exported from Phabricator. Differential Revision: D64354773 |
## Context Fix the existing implementation of int4 weight quantized linear to conform with how the `_weight_int4packed_mm` op works in the ATen library. For some additional context, the current op implementation does not actually match the behaviour of `_weight_int4packed_mm`. The ATen op expects that the weights have already been packed into a specific format, with `inner_k_tiles` as a packing parameter. The packing is accomplished via calling the `_convert_weight_to_int4pack` operator. Thus the current implementation in vulkan is equivalent to calling `_convert_weight_to_int4pack` + `_weight_int4packed_mm`. To address this discrepancy, the operator implementation is registered under the `linear_weight_int4` custom op as of this diff. The problems with the existing implementation were as follows: * The expected sizes of the scales and zeros tensor was incorrect. Previously, the sizes were assumed to be `(2, N, num_groups)` but the correct size is `(num_groups, N, 2)` * Previously, when unpacking a uint8_t into 2 unpacked int4 values, it was assumed that the LSB was the first value and the MSB was the second value. However, this ordering should be flipped * The original implementation expected the output tensor to be channels packed, but in practice we want the output tensor to be width packed This diff addresses the above issues, and introduces a dedicated test binary to test against an equivalent reference implementation expressed with ATen functions. Differential Revision: [D64354773](https://our.internmc.facebook.com/intern/diff/D64354773/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D64354773 |
SS-JIA
added a commit
that referenced
this pull request
Oct 14, 2024
Pull Request resolved: #6200 ## Context Fix the existing implementation of int4 weight quantized linear to conform with how the `_weight_int4packed_mm` op works in the ATen library. For some additional context, the current op implementation does not actually match the behaviour of `_weight_int4packed_mm`. The ATen op expects that the weights have already been packed into a specific format, with `inner_k_tiles` as a packing parameter. The packing is accomplished via calling the `_convert_weight_to_int4pack` operator. Thus the current implementation in vulkan is equivalent to calling `_convert_weight_to_int4pack` + `_weight_int4packed_mm`. To address this discrepancy, the operator implementation is registered under the `linear_weight_int4` custom op as of this diff. The problems with the existing implementation were as follows: * The expected sizes of the scales and zeros tensor was incorrect. Previously, the sizes were assumed to be `(2, N, num_groups)` but the correct size is `(num_groups, N, 2)` * Previously, when unpacking a uint8_t into 2 unpacked int4 values, it was assumed that the LSB was the first value and the MSB was the second value. However, this ordering should be flipped * The original implementation expected the output tensor to be channels packed, but in practice we want the output tensor to be width packed This diff addresses the above issues, and introduces a dedicated test binary to test against an equivalent reference implementation expressed with ATen functions. ghstack-source-id: 247975468 @exported-using-ghexport Differential Revision: [D64354773](https://our.internmc.facebook.com/intern/diff/D64354773/)
## Context Fix the existing implementation of int4 weight quantized linear to conform with how the `_weight_int4packed_mm` op works in the ATen library. For some additional context, the current op implementation does not actually match the behaviour of `_weight_int4packed_mm`. The ATen op expects that the weights have already been packed into a specific format, with `inner_k_tiles` as a packing parameter. The packing is accomplished via calling the `_convert_weight_to_int4pack` operator. Thus the current implementation in vulkan is equivalent to calling `_convert_weight_to_int4pack` + `_weight_int4packed_mm`. To address this discrepancy, the operator implementation is registered under the `linear_weight_int4` custom op as of this diff. The problems with the existing implementation were as follows: * The expected sizes of the scales and zeros tensor was incorrect. Previously, the sizes were assumed to be `(2, N, num_groups)` but the correct size is `(num_groups, N, 2)` * Previously, when unpacking a uint8_t into 2 unpacked int4 values, it was assumed that the LSB was the first value and the MSB was the second value. However, this ordering should be flipped * The original implementation expected the output tensor to be channels packed, but in practice we want the output tensor to be width packed This diff addresses the above issues, and introduces a dedicated test binary to test against an equivalent reference implementation expressed with ATen functions. Differential Revision: [D64354773](https://our.internmc.facebook.com/intern/diff/D64354773/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D64354773 |
This was referenced Oct 15, 2024
## Context Fix the existing implementation of int4 weight quantized linear to conform with how the `_weight_int4packed_mm` op works in the ATen library. For some additional context, the current op implementation does not actually match the behaviour of `_weight_int4packed_mm`. The ATen op expects that the weights have already been packed into a specific format, with `inner_k_tiles` as a packing parameter. The packing is accomplished via calling the `_convert_weight_to_int4pack` operator. Thus the current implementation in vulkan is equivalent to calling `_convert_weight_to_int4pack` + `_weight_int4packed_mm`. To address this discrepancy, the operator implementation is registered under the `linear_weight_int4` custom op as of this diff. The problems with the existing implementation were as follows: * The expected sizes of the scales and zeros tensor was incorrect. Previously, the sizes were assumed to be `(2, N, num_groups)` but the correct size is `(num_groups, N, 2)` * Previously, when unpacking a uint8_t into 2 unpacked int4 values, it was assumed that the LSB was the first value and the MSB was the second value. However, this ordering should be flipped * The original implementation expected the output tensor to be channels packed, but in practice we want the output tensor to be width packed This diff addresses the above issues, and introduces a dedicated test binary to test against an equivalent reference implementation expressed with ATen functions. Differential Revision: [D64354773](https://our.internmc.facebook.com/intern/diff/D64354773/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D64354773 |
## Context Fix the existing implementation of int4 weight quantized linear to conform with how the `_weight_int4packed_mm` op works in the ATen library. For some additional context, the current op implementation does not actually match the behaviour of `_weight_int4packed_mm`. The ATen op expects that the weights have already been packed into a specific format, with `inner_k_tiles` as a packing parameter. The packing is accomplished via calling the `_convert_weight_to_int4pack` operator. Thus the current implementation in vulkan is equivalent to calling `_convert_weight_to_int4pack` + `_weight_int4packed_mm`. To address this discrepancy, the operator implementation is registered under the `linear_weight_int4` custom op as of this diff. The problems with the existing implementation were as follows: * The expected sizes of the scales and zeros tensor was incorrect. Previously, the sizes were assumed to be `(2, N, num_groups)` but the correct size is `(num_groups, N, 2)` * Previously, when unpacking a uint8_t into 2 unpacked int4 values, it was assumed that the LSB was the first value and the MSB was the second value. However, this ordering should be flipped * The original implementation expected the output tensor to be channels packed, but in practice we want the output tensor to be width packed This diff addresses the above issues, and introduces a dedicated test binary to test against an equivalent reference implementation expressed with ATen functions. Differential Revision: [D64354773](https://our.internmc.facebook.com/intern/diff/D64354773/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D64354773 |
## Context Fix the existing implementation of int4 weight quantized linear to conform with how the `_weight_int4packed_mm` op works in the ATen library. For some additional context, the current op implementation does not actually match the behaviour of `_weight_int4packed_mm`. The ATen op expects that the weights have already been packed into a specific format, with `inner_k_tiles` as a packing parameter. The packing is accomplished via calling the `_convert_weight_to_int4pack` operator. Thus the current implementation in vulkan is equivalent to calling `_convert_weight_to_int4pack` + `_weight_int4packed_mm`. To address this discrepancy, the operator implementation is registered under the `linear_weight_int4` custom op as of this diff. The problems with the existing implementation were as follows: * The expected sizes of the scales and zeros tensor was incorrect. Previously, the sizes were assumed to be `(2, N, num_groups)` but the correct size is `(num_groups, N, 2)` * Previously, when unpacking a uint8_t into 2 unpacked int4 values, it was assumed that the LSB was the first value and the MSB was the second value. However, this ordering should be flipped * The original implementation expected the output tensor to be channels packed, but in practice we want the output tensor to be width packed This diff addresses the above issues, and introduces a dedicated test binary to test against an equivalent reference implementation expressed with ATen functions. Differential Revision: [D64354773](https://our.internmc.facebook.com/intern/diff/D64354773/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D64354773 |
junpi3
approved these changes
Oct 16, 2024
|
This pull request has been merged in 841277a. |
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.
Stack from ghstack (oldest at bottom):
export_llama#6235VkInt4WeightOnlyQuantizerfor vulkan #6234Context
Fix the existing implementation of int4 weight quantized linear to conform with how the
_weight_int4packed_mmop works in the ATen library.For some additional context, the current op implementation does not actually match the behaviour of
_weight_int4packed_mm. The ATen op expects that the weights have already been packed into a specific format, withinner_k_tilesas a packing parameter. The packing is accomplished via calling the_convert_weight_to_int4packoperator. Thus the current implementation in vulkan is equivalent to calling_convert_weight_to_int4pack+_weight_int4packed_mm. To address this discrepancy, the operator implementation is registered under thelinear_weight_int4custom op as of this diff.The problems with the existing implementation were as follows:
(2, N, num_groups)but the correct size is(num_groups, N, 2)This diff addresses the above issues, and introduces a dedicated test binary to test against an equivalent reference implementation expressed with ATen functions.
Differential Revision: D64354773