Add torchao_convert to PARQ's QuantOptimizer

[lisjin]

The original plan was to manually replace each weight tensor with an instance of IntxUnpackedToInt8Tensor. However, I found that this already happens in IntxWeightOnlyConfig and Int8DynamicActivationIntxWeightConfig when they are initialized with version=2.

• I leverage this code and call quantize_ once per regularized param_group in QuantOptimizer.torchao_convert
• For params quantized with StretchedUnifTorchaoQuantizer, I fetch the qparams, scale, zero_point to initialize IntxUnpackedToInt8Tensor (as suggested by Scott)

The PR also adds the num_steps property to QuantOptimizer to mirror D81526700.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/2947

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit 4cfa628 with merge base b99904b ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[metascroy]

Thanks! Overall looks great! Left some comments.

[andrewor14]

Looks good from my side, will let @metascroy do a final pass/stamp

[metascroy]

Overall, I think it looks great! Approving PR, but address the comments above before landing.

[lisjin]

Thanks for the comments @andrewor14 @metascroy :) I think I've addressed them all now, but let me know if I missed anything!

[jerryzh168]

for this it's probably fine to break BC soon since it's a prototype feature?

[lisjin]

Got it, thanks. I kept the version=1 convention for this config to preserve the old functionality (StretchedAffineQuantizedTensor). The new version=2 will convert tensors to IntxUnpackedToInt8Tensor, which is consistent with other config classes like IntxWeightOnlyConfig

@metascroy Should we remove StretchedAffineQuantizedTensor entirely?

[metascroy]

I have no objections to removing it. I assumed you wanted it for Int4CPULayout, but if you don't need that, I think it's fine to remove.

On ExecuTorch side, IntxUnpackedToInt8 tensor will be sufficient

[lisjin]

I think you're right about CPU support. I'll leave it for now but consider removing it in the future

[jerryzh168]

Int4CPULayout is also moved to v2 already: #2845 you can set int4_packing_format to plain_int32 to get it