openmpi launchers for deepseed should not set unsupported --num_gpus

[chiragjn]

To reproduce:

accelerate launch --mixed_precision bf16 --use_deepspeed --deepspeed_multinode_launcher openmpi --deepspeed_hostfile /etc/mpi/hostfile --same_network --monitor_interval "30" --num_machines "2" --num_processes "2" --main_process_port 29500 --machine_rank "-1" train.py

Which translates to

deepspeed --no_local_rank --hostfile /etc/mpi/hostfile --launcher openmpi --master_port 29500 --num_gpus 1 train.py --deepspeed ./deepspeed_configs/5_ds_z3_config.json --model_name_or_path Qwen/Qwen2.5-0.5B --dataset_name trl-lib/Capybara --learning_rate 2.0e-5 --num_train_epochs 1 --packing --per_device_train_batch_size 2 --gradient_accumulation_steps 8 --gradient_checkpointing --logging_steps 25 --eval_strategy steps --eval_steps 100 --output_dir /mnt/shared/Qwen2-0.5B-SFT --use_liger

Which will lead to

openmpi backend does not support limiting num nodes/gpus

Deepspeed error reference:

https://github.com/microsoft/DeepSpeed/blob/3573858e7ce2c723b8c43231c6c6b0cf97dca2fc/deepspeed/launcher/multinode_runner.py#L141C44-L141C92

---

Accelerate should skip setting num gpus for openmpi based launchers in the else clause here:

accelerate/src/accelerate/utils/launch.py

Lines 323 to 341 in d6d3e03

	if num_machines > 1 and args.deepspeed_multinode_launcher != DEEPSPEED_MULTINODE_LAUNCHERS[1]:
	cmd = ["deepspeed", "--no_local_rank"]
	cmd.extend(["--hostfile", str(args.deepspeed_hostfile), "--launcher", str(args.deepspeed_multinode_launcher)])
	if args.deepspeed_exclusion_filter is not None:
	cmd.extend(
	[
	"--exclude",
	str(args.deepspeed_exclusion_filter),
	]
	)
	elif args.deepspeed_inclusion_filter is not None:
	cmd.extend(
	[
	"--include",
	str(args.deepspeed_inclusion_filter),
	]
	)
	else:
	cmd.extend(["--num_gpus", str(args.num_processes // args.num_machines)])