Matthias Reso

Some ops use non-deterministic algorithms so some fluctuation is expected. See https://pytorch.org/docs/stable/notes/randomness.html if you can disable non-deterministic behavior but beware that this will have an impact on your training performance.

Hi @cyberyu @Mugheeera @wang-sj16 I looked into llama_adapter and it turns out that the way llama_adapter and FSDP are written they are currently incompatible. Llama_adapter [adds nn.Parameters to the model...

Turns out that in the case of prefix tuning its actually an incompatibility between the current peft and transformers/llama implementation. Previously, past_key_values used tuples as a data structure but "recently"...

Hi @aengusl which techniques are you trying to run? Are you running the latest version of llama-recipes? Please note that prefix fine tuning is currently disabled completely as the peft...

follow up to https://github.com/pytorch/serve/pull/2913/

Thanks for reporting this! Can reproduce this on my end and will have a look into it.

Hi @fredrik-jansson-se seems like for some reason the operator is not implemented for sparse tensors anymore (It was in the past). Will try to dig deeper into this later. To...

> Currently regression test run [test/pytest](https://github.com/pytorch/serve/blob/master/ts_scripts/regression_utils.py#L53). > > With these changes, tt seems that regression test will also run test_sanity. Do we need this duplicate work? @lxning good point! Have...

@lxning @msaroufim I've created a PR with the fix and a unit test to confirm it #2746 Thanks @pavel-sakun for the fix

Hi @LHQUer @ZHANGJINKUI what version of lm_eval are you using?