[ET-VK][Ops] linear_qta8a_qga4w_qta8o test framework #12005
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 This test framework establishes the foundation for validating the `linear_qta8a_qga4w_qta8o` operator implementation as part of enabling dynamic quantization. The motivation stems from advancing beyond weight-only quantization to full activation and weight quantized linear operations, enabling true integer arithmetic throughout the matrix multiplication process for improved performance on GPU hardware. The current weight-only quantized linear implementations in ET-VK dequantize weights to floating point before computation, missing the performance benefits of integer arithmetic. This operator nomenclature breakdown: - **qta8a**: Quantized per-token affine 8-bit activation inputs - **qga4w**: Quantized per-group affine 4-bit weights - **qta8o**: Quantized per-token affine 8-bit outputs # Changes The reference implementation (`linear_qta8a_qga4w_qta8o_4bit_dequant_impl`) provides a baseline for validating the GPU shader implementation through a deliberately simplified computation path. The quantized int8 input tensor is dequantized using the standard affine transformation `(quantized_input.to(at::kFloat) - input_zero_point) * input_scale`. After dequantization, the implementation performs standard floating point linear operation `at::linear(x_float, weights_dequantized)`, then manually quantizes the result using `at::round(linear_result / output_scale) + output_zero_point` with clamping to the int8 range [-128,127]. This two-stage approach of dequantize → compute → quantize provides a clear reference against which the GPU's integer arithmetic implementation can be validated. Differential Revision: [D77173442](https://our.internmc.facebook.com/intern/diff/D77173442/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/12005
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: ✅ No FailuresAs of commit 2e1cbf0 with merge base 85cf6ce ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D77173442 |
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):
Context
This test framework establishes the foundation for validating the
linear_qta8a_qga4w_qta8ooperator implementation as part of enabling dynamic quantization. The motivation stems from advancing beyond weight-only quantization to full activation and weight quantized linear operations, enabling true integer arithmetic throughout the matrix multiplication process for improved performance on GPU hardware.The current weight-only quantized linear implementations in ET-VK dequantize weights to floating point before computation, missing the performance benefits of integer arithmetic.
This operator nomenclature breakdown:
Changes
The reference implementation (
linear_qta8a_qga4w_qta8o_4bit_dequant_impl) provides a baseline for validating the GPU shader implementation through a deliberately simplified computation path. The quantized int8 input tensor is dequantized using the standard affine transformation(quantized_input.to(at::kFloat) - input_zero_point) * input_scale. After dequantization, the implementation performs standard floating point linear operationat::linear(x_float, weights_dequantized), then manually quantizes the result usingat::round(linear_result / output_scale) + output_zero_pointwith clamping to the int8 range [-128,127]. This two-stage approach of dequantize → compute → quantize provides a clear reference against which the GPU's integer arithmetic implementation can be validated.Differential Revision: D77173442