fix: ping peft >= 0.18.0, < 0.19.0 for torchao compatability issues

[davidberenstein1957]

Summary

• peft>=0.18.1 added a strict is_torchao_available() check that raises ImportError (not returns False) when torchao<0.16.0 is installed
• This broke any LoRA loading through peft (e.g. the Hyper algorithm on stable_diffusion_v1_4) — failing CI job: feat(evaluation): add VLMMetrics #545
• Removes the <0.16.0 upper cap and raises the floor to >=0.16.0

Root cause

Call chain: Hyper._apply → model.load_lora_weights → peft dispatch_torchao → is_torchao_available() → ImportError: Found version 0.15.0, but only versions above 0.16.0 are supported

The original cap was added for a torchao+torch_compile+diffusers interaction. That issue is CPU/torch_compile-specific and does not affect the base test suite. torchao 0.17.0 Python API is compatible with torch>=2.8.0 (our CI uses torch 2.9.1).

Test Plan

• CI test (3.11, base) passes — specifically TestHyper_stable_diffusion_v1_4-cpu
• TestTorchao_sd_tiny_random-cpu and TestTorchao_flux_tiny_random-cpu still pass

[codacy-production]

Up to standards ✅

🟢 Issues 0 issues

Results:
0 new issues

View in Codacy

_{TIP This summary will be updated as you push new changes. Give us feedback}

[begumcig]

Thanks for the PR @davidberenstein1957

I think we currently don’t support torchao > 0.16. When we tested newer versions, our nightlies started failing (@gsprochette would know better about this than me), so we still need to properly investigate compatibility. Also even if we supported the newer torchao, pinning it >=0.16.0 would break the torch versions we support, because they are also tightly coupled.

From my perspective, the next steps would be:

• Investigate whether we can support torchao > 0.16.
• Introduce a new extra (aligned with our dependency separation strategy in Pruna), e.g. recovery, that includes the required versions for peft, trl, termcolor, and torchao. We should also guard all module-level imports of these packages across the codebase (tests and recovery).

Since the nightlies are currently failing and we need a quick fix, I think we can temporarily pin peft to:

peft >= 0.18.0, < 0.19.0

In the PR description, you mentioned that the issue comes from peft>=0.18.1, but after checking the source code, it looks like the breaking change was actually introduced in 0.19.0:
https://github.com/huggingface/peft/blob/6d5a6f4f2f902dbf13d21d2661d57c3c05df1dae/src/peft/import_utils.py#L131
Please correct me if I’m mistaken here.

But anyway, I think overall, pinning peft to a version range above 0.18.0 and below the first version that introduces the torchao incompatibility should be a reasonable short-term fix, until we can properly support newer torchao versions. What do you think?

[davidberenstein1957]

Hi @begumcig,

I have not investigated the nightlies. @gsprochette is still on holiday, right? Let's, for now, go for the easy fix and pin peft to peft >= 0.18.0, < 0.19.0.

[begumcig]

Looks super good to me !! Thank you so much David