[quant] Quantizable MultiheadAttention #49866
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
This is still work in progress, and the tests will be amended into this PR. [ghstack-poisoned]
💊 CI failures summary and remediationsAs of commit a689cc1 (more details on the Dr. CI page):
This comment was automatically generated by Dr. CI (expand for details).Follow this link to opt-out of these comments for your Pull Requests.Please report bugs/suggestions to the (internal) Dr. CI Users group. |
This is still work in progress, and the tests will be amended into this PR. [ghstack-poisoned]
This is still work in progress, and the tests will be amended into this PR. [ghstack-poisoned]
This is still work in progress, and the tests will be amended into this PR. [ghstack-poisoned]
added 2 commits
December 26, 2020 18:48
This is still work in progress, and the tests will be amended into this PR. Differential Revision: [D25706179](https://our.internmc.facebook.com/intern/diff/D25706179) [ghstack-poisoned]
This is still work in progress, and the tests will be amended into this PR. Differential Revision: [D25706179](https://our.internmc.facebook.com/intern/diff/D25706179) [ghstack-poisoned]
This is still work in progress, and the tests will be amended into this PR. Differential Revision: [D25706179](https://our.internmc.facebook.com/intern/diff/D25706179) [ghstack-poisoned]
This is still work in progress, and the tests will be amended into this PR. Differential Revision: [D25706179](https://our.internmc.facebook.com/intern/diff/D25706179) [ghstack-poisoned]
This is still work in progress, and the tests will be amended into this PR. Differential Revision: [D25706179](https://our.internmc.facebook.com/intern/diff/D25706179) [ghstack-poisoned]
This is still work in progress, and the tests will be amended into this PR. Differential Revision: [D25706179](https://our.internmc.facebook.com/intern/diff/D25706179) [ghstack-poisoned]
This is still work in progress, and the tests will be amended into this PR. Differential Revision: [D25706179](https://our.internmc.facebook.com/intern/diff/D25706179) [ghstack-poisoned]
This is still work in progress, and the tests will be amended into this PR. Differential Revision: [D25706179](https://our.internmc.facebook.com/intern/diff/D25706179) [ghstack-poisoned]
- Adds the `torch.nn.quantizable.MultiheadAttention` The quantizable version can serve as a fully equivalent to `torch.nn.MultiheadAttention` module. The main difference is that it allows for linear units observation after the `prepare` step in the quantization flow. Note: The `from_observed` method (called during the `convert`) removes the `bias_k` and `bias_v` parameters, and resets them as attributes. This is done to avoid an error of assigning a quantized tensor to the `torch.nn.Parameter`. Test Plan: ``` python test/test_quantization.py TestQuantizedOps.test_custom_module_multi_head_attention ``` Differential Revision: [D25706179](https://our.internmc.facebook.com/intern/diff/D25706179) [ghstack-poisoned]
vkuzo
requested changes
Jan 13, 2021
- Adds the `torch.nn.quantizable.MultiheadAttention` The quantizable version can serve as a fully equivalent to `torch.nn.MultiheadAttention` module. The main difference is that it allows for linear units observation after the `prepare` step in the quantization flow. Note: The `from_observed` method (called during the `convert`) removes the `bias_k` and `bias_v` parameters, and resets them as attributes. This is done to avoid an error of assigning a quantized tensor to the `torch.nn.Parameter`. Test Plan: ``` python test/test_quantization.py TestQuantizedOps.test_custom_module_multi_head_attention ``` Differential Revision: [D25706179](https://our.internmc.facebook.com/intern/diff/D25706179) [ghstack-poisoned]
added 2 commits
January 26, 2021 13:34
- Adds the `torch.nn.quantizable.MultiheadAttention` The quantizable version can serve as a fully equivalent to `torch.nn.MultiheadAttention` module. The main difference is that it allows for linear units observation after the `prepare` step in the quantization flow. Note: The `from_observed` method (called during the `convert`) removes the `bias_k` and `bias_v` parameters, and resets them as attributes. This is done to avoid an error of assigning a quantized tensor to the `torch.nn.Parameter`. Test Plan: ``` python test/test_quantization.py TestQuantizedOps.test_custom_module_multi_head_attention ``` Differential Revision: [D25706179](https://our.internmc.facebook.com/intern/diff/D25706179) [ghstack-poisoned]
- Adds the `torch.nn.quantizable.MultiheadAttention` The quantizable version can serve as a fully equivalent to `torch.nn.MultiheadAttention` module. The main difference is that it allows for linear units observation after the `prepare` step in the quantization flow. Note: The `from_observed` method (called during the `convert`) removes the `bias_k` and `bias_v` parameters, and resets them as attributes. This is done to avoid an error of assigning a quantized tensor to the `torch.nn.Parameter`. Test Plan: ``` python test/test_quantization.py TestQuantizedOps.test_custom_module_multi_head_attention ``` Differential Revision: [D25706179](https://our.internmc.facebook.com/intern/diff/D25706179) [ghstack-poisoned]
- Adds the `torch.nn.quantizable.MultiheadAttention` The quantizable version can serve as a fully equivalent to `torch.nn.MultiheadAttention` module. The main difference is that it allows for linear units observation after the `prepare` step in the quantization flow. Note: The `from_observed` method (called during the `convert`) removes the `bias_k` and `bias_v` parameters, and resets them as attributes. This is done to avoid an error of assigning a quantized tensor to the `torch.nn.Parameter`. Test Plan: ``` python test/test_quantization.py TestQuantizedOps.test_custom_module_multi_head_attention ``` Differential Revision: [D25706179](https://our.internmc.facebook.com/intern/diff/D25706179) [ghstack-poisoned]
vkuzo
approved these changes
Jan 26, 2021
| @@ -0,0 +1,439 @@ | |||
| import torch | |||
There was a problem hiding this comment.
nit on the filename, is activation.py the best filename for it? I would have expected something like multiheadattention.py, etc
There was a problem hiding this comment.
MHA is a type of activation by convention -- it's more of a head (imho), but pytorch does not have a "head" subgroup for the layers, so it is put in the activation group
| self.dequant_v = torch.quantization.DeQuantStub() | ||
|
|
||
| def _get_name(self): | ||
| return 'QuantizableMultiheadAttention' |
There was a problem hiding this comment.
is this going to return QuantizableMultiheadAttention in both fp32 with observers and quantized state? If yes, would it be possible to refactor the code so that the final quantized module prints out QuantizedMultiheadAttention, to help with debugging?
There was a problem hiding this comment.
It is not as straight-forward. This will require reflection design pattern with the change of function when the from_observed is called. I would not recommend doing it, as this is a power-feature. Alternatively, one can check if the layer weights are quantized, and change the name accordingly. But I think it might be error prone -- I'll think about it...
There was a problem hiding this comment.
it would be good to fix this in the long term, the current state will be confusing to people who use this without reading the code
|
|
||
| q = self.q_scaling_product.mul_scalar(q, scaling) | ||
|
|
||
| if attn_mask is not None: |
| for bias, add_bias_kv, add_zero_attn in itertools.product( | ||
| Bias, Add_bias_kv, Add_zero_attn): | ||
| min_power = 20 | ||
| max_mse = 4 |
There was a problem hiding this comment.
Can we have a tighter check on the quantized module vs expected? Perhaps having a reference op that does exactly the same numerics with fp32 operations can be done.
There was a problem hiding this comment.
I already implemented the dequantize method inside the module that creates a dequantized reference -- the numerics don't exceed 20dB
There was a problem hiding this comment.
hi z-a-f
now,how can i dynamic_quantize wav2vec model?
- Adds the `torch.nn.quantizable.MultiheadAttention` The quantizable version can serve as a fully equivalent to `torch.nn.MultiheadAttention` module. The main difference is that it allows for linear units observation after the `prepare` step in the quantization flow. Note: The `from_observed` method (called during the `convert`) removes the `bias_k` and `bias_v` parameters, and resets them as attributes. This is done to avoid an error of assigning a quantized tensor to the `torch.nn.Parameter`. Test Plan: ``` python test/test_quantization.py TestQuantizedOps.test_custom_module_multi_head_attention ``` Differential Revision: [D25706179](https://our.internmc.facebook.com/intern/diff/D25706179) [ghstack-poisoned]
This was referenced Jan 30, 2021
added 7 commits
February 3, 2021 11:49
- Adds the `torch.nn.quantizable.MultiheadAttention` The quantizable version can serve as a fully equivalent to `torch.nn.MultiheadAttention` module. The main difference is that it allows for linear units observation after the `prepare` step in the quantization flow. Note: The `from_observed` method (called during the `convert`) removes the `bias_k` and `bias_v` parameters, and resets them as attributes. This is done to avoid an error of assigning a quantized tensor to the `torch.nn.Parameter`. Test Plan: ``` python test/test_quantization.py TestQuantizedOps.test_custom_module_multi_head_attention ``` Differential Revision: [D25706179](https://our.internmc.facebook.com/intern/diff/D25706179) [ghstack-poisoned]
- Adds the `torch.nn.quantizable.MultiheadAttention` The quantizable version can serve as a fully equivalent to `torch.nn.MultiheadAttention` module. The main difference is that it allows for linear units observation after the `prepare` step in the quantization flow. Note: The `from_observed` method (called during the `convert`) removes the `bias_k` and `bias_v` parameters, and resets them as attributes. This is done to avoid an error of assigning a quantized tensor to the `torch.nn.Parameter`. Test Plan: ``` python test/test_quantization.py TestQuantizedOps.test_custom_module_multi_head_attention ``` Differential Revision: [D25706179](https://our.internmc.facebook.com/intern/diff/D25706179) [ghstack-poisoned]
- Adds the `torch.nn.quantizable.MultiheadAttention` The quantizable version can serve as a fully equivalent to `torch.nn.MultiheadAttention` module. The main difference is that it allows for linear units observation after the `prepare` step in the quantization flow. Note: The `from_observed` method (called during the `convert`) removes the `bias_k` and `bias_v` parameters, and resets them as attributes. This is done to avoid an error of assigning a quantized tensor to the `torch.nn.Parameter`. Test Plan: ``` python test/test_quantization.py TestQuantizedOps.test_custom_module_multi_head_attention ``` Differential Revision: [D25706179](https://our.internmc.facebook.com/intern/diff/D25706179) [ghstack-poisoned]
- Adds the `torch.nn.quantizable.MultiheadAttention` The quantizable version can serve as a fully equivalent to `torch.nn.MultiheadAttention` module. The main difference is that it allows for linear units observation after the `prepare` step in the quantization flow. Note: The `from_observed` method (called during the `convert`) removes the `bias_k` and `bias_v` parameters, and resets them as attributes. This is done to avoid an error of assigning a quantized tensor to the `torch.nn.Parameter`. Test Plan: ``` python test/test_quantization.py TestQuantizedOps.test_custom_module_multi_head_attention ``` Differential Revision: [D25706179](https://our.internmc.facebook.com/intern/diff/D25706179) [ghstack-poisoned]
- Adds the `torch.nn.quantizable.MultiheadAttention` The quantizable version can serve as a fully equivalent to `torch.nn.MultiheadAttention` module. The main difference is that it allows for linear units observation after the `prepare` step in the quantization flow. Note: The `from_observed` method (called during the `convert`) removes the `bias_k` and `bias_v` parameters, and resets them as attributes. This is done to avoid an error of assigning a quantized tensor to the `torch.nn.Parameter`. Test Plan: ``` python test/test_quantization.py TestQuantizedOps.test_custom_module_multi_head_attention ``` Differential Revision: [D25706179](https://our.internmc.facebook.com/intern/diff/D25706179) [ghstack-poisoned]
- Adds the `torch.nn.quantizable.MultiheadAttention` The quantizable version can serve as a fully equivalent to `torch.nn.MultiheadAttention` module. The main difference is that it allows for linear units observation after the `prepare` step in the quantization flow. Note: The `from_observed` method (called during the `convert`) removes the `bias_k` and `bias_v` parameters, and resets them as attributes. This is done to avoid an error of assigning a quantized tensor to the `torch.nn.Parameter`. Test Plan: ``` python test/test_quantization.py TestQuantizedOps.test_custom_module_multi_head_attention ``` Differential Revision: [D25706179](https://our.internmc.facebook.com/intern/diff/D25706179) [ghstack-poisoned]
- Adds the `torch.nn.quantizable.MultiheadAttention` The quantizable version can serve as a fully equivalent to `torch.nn.MultiheadAttention` module. The main difference is that it allows for linear units observation after the `prepare` step in the quantization flow. Note: The `from_observed` method (called during the `convert`) removes the `bias_k` and `bias_v` parameters, and resets them as attributes. This is done to avoid an error of assigning a quantized tensor to the `torch.nn.Parameter`. Test Plan: ``` python test/test_quantization.py TestQuantizedOps.test_custom_module_multi_head_attention ``` Differential Revision: [D25706179](https://our.internmc.facebook.com/intern/diff/D25706179) [ghstack-poisoned]
|
This pull request has been merged in b8584b8. |
xsacha
pushed a commit
to xsacha/pytorch
that referenced
this pull request
Mar 31, 2021
Summary: Pull Request resolved: pytorch#49866 - Adds the `torch.nn.quantizable.MultiheadAttention` The quantizable version can serve as a fully equivalent to `torch.nn.MultiheadAttention` module. The main difference is that it allows for linear units observation after the `prepare` step in the quantization flow. Note: The `from_observed` method (called during the `convert`) removes the `bias_k` and `bias_v` parameters, and resets them as attributes. This is done to avoid an error of assigning a quantized tensor to the `torch.nn.Parameter`. (Note: this ignores all push blocking failures!) Test Plan: ``` python test/test_quantization.py TestQuantizedOps.test_custom_module_multi_head_attention ``` Imported from OSS Reviewed By: vkuzo Differential Revision: D25706179 fbshipit-source-id: e27ab641d8d1eccc64cf9e44343459331f89eea4
jasperzhong
pushed a commit
to jasperzhong/swift
that referenced
this pull request
Nov 25, 2021
- Adds the `torch.nn.quantizable.MultiheadAttention` The quantizable version can serve as a fully equivalent to `torch.nn.MultiheadAttention` module. The main difference is that it allows for linear units observation after the `prepare` step in the quantization flow. Note: The `from_observed` method (called during the `convert`) removes the `bias_k` and `bias_v` parameters, and resets them as attributes. This is done to avoid an error of assigning a quantized tensor to the `torch.nn.Parameter`. Test Plan: ``` python test/test_quantization.py TestQuantizedOps.test_custom_module_multi_head_attention ``` ghstack-source-id: db53795 Pull Request resolved: pytorch/pytorch#49866
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:
torch.nn.quantizable.MultiheadAttentionThe quantizable version can serve as a fully equivalent to
torch.nn.MultiheadAttentionmodule.The main difference is that it allows for linear units observation after the
preparestep in the quantization flow.Note: The
from_observedmethod (called during theconvert) removes thebias_kandbias_vparameters, and resets them as attributes.This is done to avoid an error of assigning a quantized tensor to the
torch.nn.Parameter.Test Plan:
Differential Revision: D25706179