Stuck on an issue?

Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.

And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.

AOT input mutation warnings for train.step() and hf accelerate

See original GitHub issue

Repro

https://github.com/sgugger/torchdynamo-tests/

Logs

[2022-11-03 23:01:33,934] torch._dynamo.optimizations.training: [WARNING] Unable to use Aot Autograd because of presence of mutation
[2022-11-03 23:01:33,935] torch._inductor.compile_fx: [WARNING] Aot Autograd is not safe to run, so falling back to eager
[2022-11-03 23:01:33,975] torch._dynamo.optimizations.training: [WARNING] Unable to use Aot Autograd because of presence of mutation
[2022-11-03 23:01:33,976] torch._inductor.compile_fx: [WARNING] Aot Autograd is not safe to run, so falling back to eager
[2022-11-03 23:01:34,012] torch._dynamo.optimizations.training: [WARNING] Unable to use Aot Autograd because of presence of mutation
[2022-11-03 23:01:34,012] torch._inductor.compile_fx: [WARNING] Aot Autograd is not safe to run, so falling back to eager
[2022-11-03 23:01:34,036] torch._dynamo.optimizations.training: [WARNING] Unable to use Aot Autograd because of presence of mutation
[2022-11-03 23:01:34,036] torch._inductor.compile_fx: [WARNING] Aot Autograd is not safe to run, so falling back to eager

Issue Analytics

State:
Created a year ago
Comments:5 (5 by maintainers)

Top GitHub Comments

1reaction

bdhirshcommented, Nov 10, 2022

WIP: https://github.com/pytorch/pytorch/pull/88817

This PR is very much a draft, but I tried running the above repro locally and it:

successfully sent everything to aot autograd
ran without error

It doesn’t look like the repro asserts correctness of the output though, so that would be good to check. I’ll continue hardening the PR in the meantime.

1reaction

bdhirshcommented, Nov 4, 2022

@yanboliang i’m planning to push on the input mutations stuff next week, yep. If there’s a quick repro command handy, I can try that out when I’m testing it.

Top Results From Across the Web

See raw diff - Hugging Face

+ st.markdown("Provide a piano solo input and click on the **Taste it!** ... + + if train: + global_loss.backward() + opt.step() + opt.zero_grad()...

accelerated stokesian dynamics: Topics by Science.gov

As a practical example, we provide a detailed, step-by-step explanation of how to perform an accelerated molecular dynamics simulation using a model ...

computational chemistry

We'll report some speed-up timings and describe difficulties porting from Intel to the Apple M1 CPU. A native Apple M1 version of ChimeraX ......

Available CRAN Packages By Name

ALSCPC, Accelerated line search algorithm for simultaneous orthogonal transformation of several positive definite symmetric matrices to nearly diagonal form.

Principal component analysis: a review and recent ... - Journals

Principal component analysis (PCA) is a technique for reducing the dimensionality of such datasets, increasing interpretability but at the same ...