Multi-Node deepspeed calling running instead of launcher

System Info

- `Accelerate` version: 0.10.0
- Platform: Linux-5.10.112-108.499.amzn2.x86_64-x86_64-with-glibc2.2.5
- Python version: 3.7.10
- Numpy version: 1.21.6
- PyTorch version (GPU?): 1.12.0+cu113 (True)
- `Accelerate` default config:
	- compute_environment: LOCAL_MACHINE
	- distributed_type: DEEPSPEED
	- mixed_precision: no
	- use_cpu: False
	- num_processes: 8
	- machine_rank: 0
	- num_machines: 1
	- main_process_ip: None
	- main_process_port: None
	- main_training_function: main
	- deepspeed_config: {'deepspeed_config_file': '/path/to/deepspeed_config.json', 'zero3_init_flag': False}
	- fsdp_config: {}

deepspeed==0.6.5

{
    "train_batch_size": 128,
    "gradient_accumulation_steps": 1,
    "gradient_clipping": 1,
    "fp16": {
        "enabled": true,
        "loss_scale": 0,
        "loss_scale_window": 1000,
        "hysteresis": 2,
        "min_loss_scale": 0.00001
    },
    "zero_optimization": {
        "stage": 2,
        "cpu_offload": true,
        "contiguous_gradients": true,
        "overlap_comm": true
    }
}

Running on a slurm HPC.

Information

• The official example scripts
• My own modified scripts

Tasks

• One of the scripts in the examples/ folder of Accelerate or an officially supported no_trainer script in the examples folder of the transformers repo (such as run_no_trainer_glue.py)
• My own task or dataset (give details below)

Reproduction

Install: pip install accelerate pip install deepspeed

Create the config: Note: deepspeed config can be empty as the crash happens before it is opened. Similarly, main_process_ip and main_process_port can be anything as they are not used before the crash.

{
  "compute_environment": "LOCAL_MACHINE",
  "deepspeed_config": {
    "deepspeed_config_file": "/path/to/deepspeed_config.json",
    "zero3_init_flag": false
  },
  "distributed_type": "DEEPSPEED",
  "fsdp_config": {},
  "machine_rank": 0,
  "main_process_ip": "0.0.0.0",
  "main_process_port": 0000,
  "main_training_function": "main",
  "mixed_precision": "no",
  "num_machines": 2,
  "num_processes": 8,
  "use_cpu": false
}

Launch accelerate Again, since the training script is not launched, it can be empty. accelerate launch --config_file /path/to/accelerate_config.json /path/to/empty_script.py

Expected behavior

When launching a multinode deepspeed training script with this code https://github.com/huggingface/accelerate/blob/86ce737d7fc94f8000dbd5e13021d0411bb4204a/src/accelerate/commands/launch.py#L312-L327 it appears to be running the deepspeed runner where the arguments supplied are for the deepspeed launcher.

This means that I am getting this error when accelerate tries to launch deepspeed.

usage: deepspeed [-h] [-H HOSTFILE] [-i INCLUDE] [-e EXCLUDE]
                 [--num_nodes NUM_NODES] [--num_gpus NUM_GPUS]
                 [--master_port MASTER_PORT] [--master_addr MASTER_ADDR]
                 [--launcher LAUNCHER] [--launcher_args LAUNCHER_ARGS]
                 [--force_multi] [--autotuning {tune,run}]
                 user_script ...
deepspeed: error: unrecognized arguments: --no_local_rank

Since accelerate is performing the same function as the deepspeed runner, I would expect accelerate to call the launcher directly on each of the nodes. Instead, it appears to be calling the runner on each of the nodes.