[WIP] Streaming DiLoCo prototype

[H-Huang]

Creating a small script to quickly hack on the implementation for streaming DiLoCo.

Run with (start lighthouse first by looking at command in README.md):

cd streaming_diloco_prototype
torchx run

Issues found:

1. Quantization only supports 2D tensors (can workaround)

replica_0/0     File "/home/howardhuang/.conda/envs/torchft/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 357, in wrapper
replica_0/0       return f(*args, **kwargs)
replica_0/0     File "/data/users/howardhuang/torchft/streaming_diloco_prototype/train.py", line 275, in streaming_diloco
replica_0/0       fut = allreduce_quantized(params_data, ReduceOp.AVG, pg)
replica_0/0     File "/data/users/howardhuang/torchft/torchft/collectives.py", line 104, in allreduce_quantized
replica_0/0       quantized_tensors = fused_quantize_into_fp8(tensors, world_size)
replica_0/0     File "/data/users/howardhuang/torchft/torchft/quantization.py", line 520, in fused_quantize_into_fp8
replica_0/0       ) = _prepare_quantize_fp8(inputs, all_reduce_group_size)
replica_0/0     File "/data/users/howardhuang/torchft/torchft/quantization.py", line 450, in _prepare_quantize_fp8
replica_0/0       assert len(inputs[i].shape) == 2, "Only 2D tensors are supported"
replica_0/0   AssertionError: Only 2D tensors are supported

2. triton runtime jit issue when calling _fused_kernel_quantize_into_fp8[grid]

replica_0/0     File "/data/users/howardhuang/torchft/streaming_diloco_prototype/train.py", line 275, in streaming_diloco
replica_0/0       fut = allreduce_quantized(params_data, ReduceOp.AVG, pg)
replica_0/0     File "/data/users/howardhuang/torchft/torchft/collectives.py", line 104, in allreduce_quantized
replica_0/0       quantized_tensors = fused_quantize_into_fp8(tensors, world_size)
replica_0/0     File "/data/users/howardhuang/torchft/torchft/quantization.py", line 531, in fused_quantize_into_fp8
replica_0/0       _fused_kernel_quantize_into_fp8[grid](
replica_0/0     File "/home/howardhuang/.conda/envs/torchft/lib/python3.10/site-packages/triton/runtime/jit.py", line 499, in run
replica_0/0       if key not in self.cache[device]:
replica_0/0   TypeError: unhashable type: 'constexpr'

[meta-cla]

Hi @H-Huang!

Thank you for your pull request.

We require contributors to sign our Contributor License Agreement, and yours needs attention.

You currently have a record in our system, but the CLA is no longer valid, and will need to be resubmitted.

Process

In order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA.

Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with CLA signed. The tagging process may take up to 1 hour after signing. Please give it that time before contacting us about it.

If you have received this in error or have any questions, please contact us at cla@meta.com. Thanks!