Self-Compression QAT and Linear #1342
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/1342
Note: Links to docs will display an error until the docs builds have been completed. ❌ 10 New FailuresAs of commit 64abffd with merge base b2e42ff ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
|
Hi @lliangthomas! Thank you for your pull request and welcome to our community. Action RequiredIn order to merge any pull request (code, docs, etc.), we require contributors to sign our Contributor License Agreement, and we don't seem to have one on file for you. ProcessIn order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA. Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with If you have received this in error or have any questions, please contact us at [email protected]. Thanks! |
|
Thank you for signing our Contributor License Agreement. We can now accept your code for this (and any) Meta Open Source project. Thanks! |
facebook-github-bot
added
the
CLA Signed
|
Hi @lliangthomas, thanks for your contribution. Have you had a chance to do any full training runs with evaluation results? I'm inclined to move this under |
| Forward pass with weight quantization | ||
| """ | ||
| quant_max = torch.maximum(2. ** -self.float_exponents * self.weight, -2. ** (self.bit_depth.relu() - 1)) | ||
| quant_weight = torch.minimum(quant_max, 2. ** (self.bit_depth.relu() - 1) - 1) |
There was a problem hiding this comment.
can we rewrite this as follows so it's easier to read?
qmax = 2. ** (self.bit_depth.relu() - 1) - 1
qmin = -2. ** (self.bit_depth.relu() - 1)
qweight = torch.clamp(2. ** -self.float_exponents * self.weight, qmin, qmax)
| quant_max = torch.maximum(2. ** -self.float_exponents * self.weight, -2. ** (self.bit_depth.relu() - 1)) | ||
| quant_weight = torch.minimum(quant_max, 2. ** (self.bit_depth.relu() - 1) - 1) | ||
| rounded_weight = (quant_weight.round() - quant_weight).detach() + quant_weight | ||
| return F.linear(x, 2. ** self.float_exponents * rounded_weight) |
There was a problem hiding this comment.
What does the backward pass look like? I think we need to wrap all of this in an autograd.Function and define a backward pass since this has non-differentiable ops like round()
| @@ -30,6 +30,82 @@ | |||
| _get_qmin_qmax, | |||
| ) | |||
|
|
|||
| class SelfCompressionQATQuantizer(torch.nn.Module): | |||
There was a problem hiding this comment.
This feels like a pretty experimental feature. I think this belongs better to a separate folder under torchao/prototype
Implemented self-compression QAT and linear layer from this paper as a solution to #658
If there are more features or more integration to be done with the current QAT schemes in TorchAO (self-compression seems orthogonal right now), let me know.
@msaroufim @HDCharles