fix default value for rank size in cpu threads_per_process assignment…

… logic (#3009)

* fix default value for rank size

* fix style

* apply int in case ratio is decimal

* style quality fix

src/accelerate/commands/launch.py

@@ -1126,7 +1126,8 @@ def _validate_launch_command(args):
         args.num_cpu_threads_per_process = get_int_from_env(["OMP_NUM_THREADS"], 1)
         if args.use_cpu and args.num_processes >= 1 and get_int_from_env(["OMP_NUM_THREADS"], 0) == 0:
             local_size = get_int_from_env(
-                ["MPI_LOCALNRANKS", "OMPI_COMM_WORLD_LOCAL_SIZE", "MV2_COMM_WORLD_LOCAL_SIZE"], 1
+                ["MPI_LOCALNRANKS", "OMPI_COMM_WORLD_LOCAL_SIZE", "MV2_COMM_WORLD_LOCAL_SIZE"],
+                max(int(args.num_processes / args.num_machines), 1),
             )
             threads_per_process = int(psutil.cpu_count(logical=False) / local_size)
             if threads_per_process > 1: