Logan Adams comments

Results 294 comments of


                                            Logan Adams

Rename config parameters for pydantic v2 changes

> Looks like v2 will not be compatible with v1 - we will need to add `pydantic>=2.0.0` to requirements once we make the switch. It might be worth adding `pydantic

when I execute the command "python train.py --actor-model facebook/opt-1.3b --reward-model facebook/opt-350m --deployment-type single_gpu" I got the follow result, and I try many solutions but all failure.

Closing this issue for now - if there are new installation issues, please feel free to open new tickets.

Switch to adding AMD/ROCm pre-compile too

For some reason the build can show as passing but only the cuda job passes, the rocm/rocr one fails.

[BUG] Triton Error [CUDA]: invalid argument

Can you test again with the latest DeepSpeed with the triton versions updated? If you are still seeing this, can you re-open this issue?

[BUG] Triton Error [CUDA]: invalid argument

@hayday100 - for now, can you use the triton version listed in requirements-sd.txt? That version specifically works when running our unit tests.

If you want to add arguments to the training such as the ones you list above (e.g., --gradient_checkpointing) you'll need to add them after `main.py` in the script for example:

Hi @TinyQi - did you mean to make this a new issue? This looks like it is still being covered in the original linked issue. @khai0617 - what parameters you...

If you want to add arguments to the training such as the ones you list above (e.g., --gradient_checkpointing) you'll need to add them after `main.py` in the script for example:

Closing this issue for preference to track in the main issue. If you need this one re-opened with a different question, please re-open.

Logan Adams

Rename config parameters for pydantic v2 changes

when I execute the command "python train.py --actor-model facebook/opt-1.3b --reward-model facebook/opt-350m --deployment-type single_gpu" I got the follow result, and I try many solutions but all failure.

Switch to adding AMD/ROCm pre-compile too

[BUG] Triton Error [CUDA]: invalid argument

[BUG] Triton Error [CUDA]: invalid argument

If you want to add arguments to the training such as the ones you list above (e.g., --gradient_checkpointing) you'll need to add them after `main.py` in the script for example:

If you want to add arguments to the training such as the ones you list above (e.g., --gradient_checkpointing) you'll need to add them after `main.py` in the script for example:

Fix a bug in the implementation of dequantization for inference

About DEFAULT_MIN_MEM_CONFIG in autoning

About DEFAULT_MIN_MEM_CONFIG in autoning