Pin numpy version

[BLOrange-AMD]

This PR is to fix incompatible numpy version of pyt_deepspeed_megatron_gpt2 and pyt_train_deepspeed_megatron_gpt2 with ROCm PyTorch release/2.5 branch.

[loadams]

Hi @BLOrange-AMD - I recall there being a dependency between the torch version and numpy version for us, could you share more info on the error you are seeing?

[BLOrange-AMD]

@loadams With ROCm PyTorch 2.5, on pyt_deepspeed_megatron_gpt2 and pyt_train_deepspeed_megatron_gpt2 models, newer numpy-2.2.1 is downloaded and used instead of using cached numpy-1.26.4, which causes "RuntimeError: Could not infer dtype of numpy.int64". So fixed numpy version from DeepSpeed could be a more stable way to solve the issue.

[loadams]

@loadams With ROCm PyTorch 2.5, on pyt_deepspeed_megatron_gpt2 and pyt_train_deepspeed_megatron_gpt2 models, newer numpy-2.2.1 is downloaded and used instead of using cached numpy-1.26.4, which causes "RuntimeError: Could not infer dtype of numpy.int64". So fixed numpy version from DeepSpeed could be a more stable way to solve the issue.

@BLOrange-AMD - I see, however the issue is that DeepSpeed doesn't strictly require a lower numpy version. Here is a sample workflow that uses numpy>2.0.0. Is there another way to pin the numpy version? Or do you believe this will be fixed soon in torch 2.6?

Or perhaps torch just needs to be built with numpy support too?