[v2] Branching simulation modes & execution-time entry points

[vaseline555]

Describe the Request

APPFL currently supports run_* under /src/appfl. (MPI/gPRC/Globus)
While run_serial.py exists, it will be deprecated in future, which seems useful for laptop-scale simulation experiments.

1. branching simulation modes

• I suggest to revive the run_serial.py, by having four modes internally:
  • run_serial: running on CPU/GPU in serial manner (for-loop style)
  • run_gloo: running in parallel on CPUs (torch.distributed.init_process_group(backend='gloo'))
  • run_nccl: running in parallel on multiple GPUs (torch.distributed.init_process_group(backend='nccl'))
  • run_mpi: running in parallel on multiple GPUs for intel GPUs (i.e., XPU)
• While gloo can be regarded as an alternative of mpi, but it is thread-safe, pytorch-native backend, and thus no dependency on mpiexec.
• (draft example):

  APPFL/src/appfl/sim/runner.py

  Line 580 in 4a26e03

  def run_distributed(config, backend: str) -> None:

2. branching entry points

• appfl
  • run
    • mode=real
      • comm=grpc
      • comm=globus
      • comm=ray
    • mode=sim
      • backend=serial
      • backend=gloo
      • backend=nccl
      • backend=mpi
  • commit
    • interface=chat // vibe coding interface to auto generate essential files for algorithm implementation.
    • interface=manual // sanity check interface to confirm manually-coded essential modules by users (e.g., aggregator,scheduler,trainer)
• Currently a simple suggestion -- needs further discussion on this if acceptable

Sample Code

// simulation modes
appfl run mode=sim backend=serial config=...
appfl run mode=sim backend=gloo config=...
appfl run mode=sim backend=ncll config=...

// deployment modes
appfl run mode=real comm=grpc config=...
appfl run mode=real comm=globus config=...
...

// commit modes
appfl commit interface=chat config=...
appfl commit interface=manual config=...

Additional Code or Information

To-Do

• Check and test torch.distributed compatibility with Intel XPU
• Determine run_mpi design
• Devise and confirm entry points design pattern
• Research existing Generative UI design for the chat commit modes
• Envision specific paths and plausible scenarios for the sustainable maintenance of the commit mode (for both advanced / new users)