[quant][gpu][core] Added quantized linear operator in cudnn #73959
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
Summary: This PR is similar to #70622, but for the linear operator. Unlke PR 70622, this implementations directly uses packed parameters, rather than a refactorization, as was done for the conv operator, and also directly implements bias & relu. [ghstack-poisoned]
CI Flow Status⚛️ CI FlowRuleset - Version:
|
Contributor
🔗 Helpful links
💊 CI failures summary and remediationsAs of commit 91511fd (more details on the Dr. CI page): 💚 💚 Looks good so far! There are no failures yet. 💚 💚 This comment was automatically generated by Dr. CI (expand for details).Please report bugs/suggestions to the (internal) Dr. CI Users group. |
This was referenced Mar 9, 2022
Closed
dzdang
added a commit
that referenced
this pull request
Mar 9, 2022
Summary: This PR is similar to #70622, but for the linear operator. Unlke PR 70622, this implementations directly uses packed parameters, rather than a refactorization, as was done for the conv operator, and also directly implements bias & relu. ghstack-source-id: edd733d Pull Request resolved: #73959
Summary: This PR is similar to #70622, but for the linear operator. Unlke PR 70622, this implementations directly uses packed parameters, rather than a refactorization, as was done for the conv operator, and also directly implements bias & relu. [ghstack-poisoned]
Summary: This PR is similar to #70622, but for the linear operator. Unlke PR 70622, this implementations directly uses packed parameters, rather than a refactorization, as was done for the conv operator, and also directly implements bias & relu. [ghstack-poisoned]
dzdang
added a commit
that referenced
this pull request
Mar 10, 2022
Summary: This PR is similar to #70622, but for the linear operator. Unlke PR 70622, this implementations directly uses packed parameters, rather than a refactorization, as was done for the conv operator, and also directly implements bias & relu. ghstack-source-id: b8f25a2 Pull Request resolved: #73959
Summary: This PR is similar to #70622, but for the linear operator. Unlke PR 70622, this implementations directly uses packed parameters, rather than a refactorization, as was done for the conv operator, and also directly implements bias & relu. [ghstack-poisoned]
dzdang
added a commit
that referenced
this pull request
Mar 11, 2022
Summary: This PR is similar to #70622, but for the linear operator. Unlke PR 70622, this implementations directly uses packed parameters, rather than a refactorization, as was done for the conv operator, and also directly implements bias & relu. ghstack-source-id: 4423fcf Pull Request resolved: #73959
Summary: This PR is similar to #70622, but for the linear operator. Unlke PR 70622, this implementations directly uses packed parameters, rather than a refactorization, as was done for the conv operator, and also directly implements bias & relu. [ghstack-poisoned]
dzdang
added a commit
that referenced
this pull request
Mar 11, 2022
Summary: This PR is similar to #70622, but for the linear operator. Unlke PR 70622, this implementations directly uses packed parameters, rather than a refactorization, as was done for the conv operator, and also directly implements bias & relu. ghstack-source-id: 5655128 Pull Request resolved: #73959
Contributor
|
Please add a |
Summary: This PR is similar to #70622, but for the linear operator. Unlke PR 70622, this implementations directly uses packed parameters, rather than a refactorization, as was done for the conv operator, and also directly implements bias & relu. [ghstack-poisoned]
Summary: This PR is similar to #70622, but for the linear operator. Unlke PR 70622, this implementations directly uses packed parameters, rather than a refactorization, as was done for the conv operator, and also directly implements bias & relu. [ghstack-poisoned]
dzdang
added a commit
that referenced
this pull request
Mar 11, 2022
Summary: This PR is similar to #70622, but for the linear operator. Unlke PR 70622, this implementations directly uses packed parameters, rather than a refactorization, as was done for the conv operator, and also directly implements bias & relu. ghstack-source-id: 3bba699 Pull Request resolved: #73959
Contributor
Author
|
@dzdang has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
…n [PR currently incomplete]" Summary: This PR is similar to #70622, but for the linear operator. Unlke PR 70622, this implementations directly uses packed parameters, rather than a refactorization, as was done for the conv operator, and also directly implements bias & relu. Test plan: Differential Revision: [D34824251](https://our.internmc.facebook.com/intern/diff/D34824251) [ghstack-poisoned]
Contributor
Author
|
@dzdang has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
Summary: This PR is similar to #70622, but for the linear operator. Unlke PR 70622, this implementations directly uses packed parameters, rather than a refactorization, as was done for the conv operator, and also directly implements bias & relu. Currently, int8 matrix multiplication is not supported in cudnn. The ETA for this support is in the first half of April 2022. As a temporary workaround, we cast our int8 tensors to fp32 prior to matmul. Test plan: ``` python test/test_quantization.py TestQuantizedLinear.test_qlinear_cudnn ``` Differential Revision: [D34824251](https://our.internmc.facebook.com/intern/diff/D34824251) [ghstack-poisoned]
dzdang
added a commit
that referenced
this pull request
Mar 29, 2022
Summary: This PR is similar to #70622, but for the linear operator. Unlke PR 70622, this implementations directly uses packed parameters, rather than a refactorization, as was done for the conv operator, and also directly implements bias & relu. Currently, int8 matrix multiplication is not supported in cudnn. The ETA for this support is in the first half of April 2022. As a temporary workaround, we cast our int8 tensors to fp32 prior to matmul. Test plan: ``` python test/test_quantization.py TestQuantizedLinear.test_qlinear_cudnn ``` ghstack-source-id: 2e4a852 Pull Request resolved: #73959
Contributor
Author
|
@dzdang has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
jerryzh168
reviewed
Mar 30, 2022
jerryzh168
reviewed
Mar 30, 2022
jerryzh168
reviewed
Mar 30, 2022
| // we need to add trailing dimensions in order to properly broadcast bias, otherwise broadcast_to will fail. | ||
| // the number of trailling dimensions is quantized_output.dim() - 2, so the new size of the broadcast_bias | ||
| // becomes quantized_output.dim() - 2 + 1. nothing needs to be done for the leading dimensions | ||
| std::vector<int64_t> new_size(quantized_output.dim() - 1, 1); |
Contributor
There was a problem hiding this comment.
- is the call
ndim? - I feel maybe just match the dimension is cleaner, i.e. create a new_size to have the same dimension as quantized_output, and set new_size[1] = the expected dimension
jerryzh168
reviewed
Mar 30, 2022
jerryzh168
reviewed
Mar 30, 2022
| auto weight_fp = weight_transposed.int_repr().to(at::kFloat); | ||
|
|
||
| auto run = [&](cudnn_frontend::ManagedOpaqueDescriptor plan_desc) { | ||
| auto workspace_size = 0; |
Contributor
There was a problem hiding this comment.
I feel in general we have a lot of bolierplate code, maybe we can think about creating some helper functions or easier abstractions to make this simpler, this will be helpful when we have more ops in cudnn
jerryzh168
reviewed
Mar 30, 2022
| // .setbMatDesc(cudnn_utils::getTensorDescriptor(orig_weight.sizes(), orig_weight.strides(), CUDNN_DATA_FLOAT, 'w', key.weight_alignment)) | ||
| .setbMatDesc(cudnn_utils::getTensorDescriptor(weight_fp.sizes(), weight_fp.strides(), CUDNN_DATA_FLOAT, 'w', key.weight_alignment)) | ||
| .setcMatDesc(cudnn_utils::getTensorDescriptor(linear_output, 'y', key.output_alignment)) | ||
| .setmatmulDesc(getLinearDescriptor(CUDNN_DATA_FLOAT)) // is this right? should it be float? |
Contributor
There was a problem hiding this comment.
I remember we have a table for the descriptor data type, maybe we can implement that in a function: getting descriptor data type from input data type
jerryzh168
reviewed
Mar 30, 2022
jerryzh168
approved these changes
Mar 30, 2022
Contributor
jerryzh168
left a comment
jerryzh168
left a comment
There was a problem hiding this comment.
Looks good, had some nit comments inline
Summary: This PR is similar to #70622, but for the linear operator. Unlke PR 70622, this implementations directly uses packed parameters, rather than a refactorization, as was done for the conv operator, and also directly implements bias & relu. Currently, int8 matrix multiplication is not supported in cudnn. The ETA for this support is in the first half of April 2022. As a temporary workaround, we cast our int8 tensors to fp32 prior to matmul. Test plan: ``` python test/test_quantization.py TestQuantizedLinear.test_qlinear_cudnn ``` Differential Revision: [D34824251](https://our.internmc.facebook.com/intern/diff/D34824251) [ghstack-poisoned]
dzdang
added a commit
that referenced
this pull request
Mar 31, 2022
Summary: This PR is similar to #70622, but for the linear operator. Unlke PR 70622, this implementations directly uses packed parameters, rather than a refactorization, as was done for the conv operator, and also directly implements bias & relu. Currently, int8 matrix multiplication is not supported in cudnn. The ETA for this support is in the first half of April 2022. As a temporary workaround, we cast our int8 tensors to fp32 prior to matmul. Test plan: ``` python test/test_quantization.py TestQuantizedLinear.test_qlinear_cudnn ``` ghstack-source-id: 586d55e Pull Request resolved: #73959
Contributor
Author
|
@dzdang has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
Summary: This PR is similar to #70622, but for the linear operator. Unlke PR 70622, this implementations directly uses packed parameters, rather than a refactorization, as was done for the conv operator, and also directly implements bias & relu. Currently, int8 matrix multiplication is not supported in cudnn. The ETA for this support is in the first half of April 2022. As a temporary workaround, we cast our int8 tensors to fp32 prior to matmul. Test plan: ``` python test/test_quantization.py TestQuantizedLinear.test_qlinear_cudnn ``` Differential Revision: [D34824251](https://our.internmc.facebook.com/intern/diff/D34824251) [ghstack-poisoned]
Contributor
Author
|
@dzdang has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
Summary: This PR is similar to #70622, but for the linear operator. Unlke PR 70622, this implementations directly uses packed parameters, rather than a refactorization, as was done for the conv operator, and also directly implements bias & relu. Currently, int8 matrix multiplication is not supported in cudnn. The ETA for this support is in the first half of April 2022. As a temporary workaround, we cast our int8 tensors to fp32 prior to matmul. Test plan: ``` python test/test_quantization.py TestQuantizedLinear.test_qlinear_cudnn ``` Differential Revision: [D34824251](https://our.internmc.facebook.com/intern/diff/D34824251) [ghstack-poisoned]
Contributor
Author
|
@dzdang has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
Summary: This PR is similar to #70622, but for the linear operator. Unlke PR 70622, this implementations directly uses packed parameters, rather than a refactorization, as was done for the conv operator, and also directly implements bias & relu. Currently, int8 matrix multiplication is not supported in cudnn. The ETA for this support is in the first half of April 2022. As a temporary workaround, we cast our int8 tensors to fp32 prior to matmul. Test plan: ``` python test/test_quantization.py TestQuantizedLinear.test_qlinear_cudnn ``` Differential Revision: [D34824251](https://our.internmc.facebook.com/intern/diff/D34824251) [ghstack-poisoned]
dzdang
added a commit
that referenced
this pull request
Mar 31, 2022
Summary: This PR is similar to #70622, but for the linear operator. Unlke PR 70622, this implementations directly uses packed parameters, rather than a refactorization, as was done for the conv operator, and also directly implements bias & relu. Currently, int8 matrix multiplication is not supported in cudnn. The ETA for this support is in the first half of April 2022. As a temporary workaround, we cast our int8 tensors to fp32 prior to matmul. Test plan: ``` python test/test_quantization.py TestQuantizedLinear.test_qlinear_cudnn ``` ghstack-source-id: 0a2c6c1 Pull Request resolved: #73959
Contributor
Author
|
@dzdang has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
facebook-github-bot
pushed a commit
that referenced
this pull request
Apr 1, 2022
Summary: Pull Request resolved: #73959 This PR is similar to #70622, but for the linear operator. Unlke PR 70622, this implementations directly uses packed parameters, rather than a refactorization, as was done for the conv operator, and also directly implements bias & relu. Currently, int8 matrix multiplication is not supported in cudnn. The ETA for this support is in the first half of April 2022. As a temporary workaround, we cast our int8 tensors to fp32 prior to matmul. Test Plan: ``` python test/test_quantization.py TestQuantizedLinear.test_qlinear_cudnn ``` Imported from OSS Differential Revision: D34824251 D34824251 Reviewed By: jerryzh168 Pulled By: dzdang fbshipit-source-id: 47139796782ade8d030ba2f9968a9abdd3a91d2f
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):
Summary:
This PR is similar to #70622, but for the linear operator.
Unlke PR 70622, this implementations directly uses packed parameters, rather than a refactorization, as was done for the conv operator,
and also directly implements bias & relu.
Currently, int8 matrix multiplication is not supported in cudnn. The ETA for this support is in the first half of April 2022. As
a temporary workaround, we cast our int8 tensors to fp32 prior to matmul.
Test plan:
Differential Revision: D34824251