[BUG] Training with RoPE is broken: Can't stop FP32 layers from being cast to FP16/BF16 during training

Describe the bug
Many modern transformer components (e.g., RoPE, certain Layer Norm setups) need to be stored and run in FP32. Most of the time, we can accomplish this by disabling autocast or overriding the to() method. When using DeepSpeed's BF16 or FP16 modes with ZeRO it appears neither of these works. I can force cast the types back to what they should be in the forward method of RoPE, but I fear the precision loss in the buffer will have non-obvious negative performance effects.

It seems this was mentioned in discussions a month ago but was never answered: https://github.com/microsoft/DeepSpeed/discussions/5678

To Reproduce
Steps to reproduce the behavior:

1. Make a layer with a FP32 weight or buffer
2. Disable autocast and override to()
3. Enable BF16
4. Print the dtype in forward, it will be bf16. The to() method is not called and yet the dtype is changed.

Expected behavior
A clear and concise description of what you expected to happen.
to() should be called so we can stop param types from being changed.

[2024-07-30 18:59:46,909] [INFO] [real_accelerator.py:203:get_accelerator] Setting ds_accelerator to cuda (auto detect)
 [WARNING]  async_io requires the dev libaio .so object and headers but these were not found.
 [WARNING]  async_io: please install the libaio-dev package with apt
 [WARNING]  If libaio is already installed (perhaps from source), try setting the CFLAGS and LDFLAGS environment variables to where it can be found.
 [WARNING]  Please specify the CUTLASS repo directory as environment variable $CUTLASS_PATH
 [WARNING]  sparse_attn requires a torch version >= 1.5 and < 2.0 but detected 2.3
 [WARNING]  using untested triton version (2.3.1), only 1.0.0 is known to be compatible
--------------------------------------------------
DeepSpeed C++/CUDA extension op report
--------------------------------------------------
NOTE: Ops not installed will be just-in-time (JIT) compiled at
      runtime if needed. Op compatibility means that your system
      meet the required dependencies to JIT install the op.
--------------------------------------------------
JIT compiled ops requires ninja
ninja .................. [OKAY]
--------------------------------------------------
op name ................ installed .. compatible
--------------------------------------------------
 [WARNING]  async_io requires the dev libaio .so object and headers but these were not found.
 [WARNING]  async_io: please install the libaio-dev package with apt
 [WARNING]  If libaio is already installed (perhaps from source), try setting the CFLAGS and LDFLAGS environment variables to where it can be found.
async_io ............... [NO] ....... [NO]
fused_adam ............. [NO] ....... [OKAY]
cpu_adam ............... [NO] ....... [OKAY]
cpu_adagrad ............ [NO] ....... [OKAY]
cpu_lion ............... [NO] ....... [OKAY]
 [WARNING]  Please specify the CUTLASS repo directory as environment variable $CUTLASS_PATH
evoformer_attn ......... [NO] ....... [NO]
fp_quantizer ........... [NO] ....... [OKAY]
fused_lamb ............. [NO] ....... [OKAY]
fused_lion ............. [NO] ....... [OKAY]
inference_core_ops ..... [NO] ....... [OKAY]
cutlass_ops ............ [NO] ....... [OKAY]
transformer_inference .. [NO] ....... [OKAY]
quantizer .............. [NO] ....... [OKAY]
ragged_device_ops ...... [NO] ....... [OKAY]
ragged_ops ............. [NO] ....... [OKAY]
random_ltd ............. [NO] ....... [OKAY]
 [WARNING]  sparse_attn requires a torch version >= 1.5 and < 2.0 but detected 2.3
 [WARNING]  using untested triton version (2.3.1), only 1.0.0 is known to be compatible
sparse_attn ............ [NO] ....... [NO]
spatial_inference ...... [NO] ....... [OKAY]
transformer ............ [NO] ....... [OKAY]
stochastic_transformer . [NO] ....... [OKAY]
--------------------------------------------------
DeepSpeed general environment info:
torch install path ............... ['/home/azureuser/.cache/pypoetry/virtualenvs/the-model-eAEmNTqV-py3.11/lib/python3.11/site-packages/torch']
torch version .................... 2.3.1+cu121
deepspeed install path ........... ['/home/azureuser/.cache/pypoetry/virtualenvs/the-model-eAEmNTqV-py3.11/lib/python3.11/site-packages/deepspeed']
deepspeed info ................... 0.14.4, unknown, unknown
torch cuda version ............... 12.1
torch hip version ................ None
nvcc version ..................... 12.2
deepspeed wheel compiled w. ...... torch 0.0, cuda 0.0
shared memory (/dev/shm) size .... 433.06 GB
(base) azureuser@ray-gpu4modeltest2bf16-7a79-head-3a42-90820:~/sky_workdir$ 

No source file or test is named. Start with the reproduction steps and inspect the DeepSpeed ZeRO path used by BF16/FP16 training, focusing on how FP32 parameters or buffers are cast and whether the layer’s to() override is bypassed. Done means a focused regression test demonstrates that the intended FP32 dtype is preserved.