Compatibility with PyTorch 2.7.0 and FBGEMM_GPU 1.2.0?

[xivid]

torch 2.7.0+cu128 and fbgemm_gpu 1.2.0+cu128 have been released for some time, but the current torchrec release is still 1.1.0, and fails to run with the latest torch+fbgemm_gpu versions.

Is there a branch that could work?

[kausv]

Hi @xivid have you checked for stable release? We run our tests against specific FBGEMM nightly versions

[xivid]

hi @kausv Thanks for your reply!

Yes, I have tried the stable release torchrec==1.1.0, however it does not run out of the box in a torch==2.7.0+cu128 and fbgemm_gpu==1.2.0+cu128 environment. I have been using the dlrm example as my test program.

After some debugging, I managed to make it run by modifying the following line:

torchrec/torchrec/distributed/batched_embedding_kernel.py

Line 194 in cac04c8

"lr": emb_module.optimizer_args.learning_rate,

emb_module.optimizer_args.learning_rate needs to be changed to emb_module.get_learning_rate().

I'm not sure if this is the only compatibility issue that exists, but at least I can confirm, with this change, the dlrm example can run successfully now.

[jiyuanq]

same question here, I though torchrec, like other pytorch official domain libraries, has one to one version mapping to torch version. torchrec 1.1.0 is built against torch 2.6.x, so we should have a torchrec 1.2.0 built against torch 2.7.0, which was released almost a month ago.