[Sigmas] Keep sigmas on CPU

src/diffusers/schedulers/scheduling_consistency_models.py

@@ -98,6 +98,7 @@ def __init__(
         self.custom_timesteps = False
         self.is_scale_input_called = False
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
 
     def index_for_timestep(self, timestep, schedule_timesteps=None):
         if schedule_timesteps is None:
@@ -230,6 +231,7 @@ def set_timesteps(
             self.timesteps = torch.from_numpy(timesteps).to(device=device)
 
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
 
     # Modified _convert_to_karras implementation that takes in ramp as argument
     def _convert_to_karras(self, ramp):

src/diffusers/schedulers/scheduling_deis_multistep.py

@@ -187,6 +187,7 @@ def __init__(
         self.model_outputs = [None] * solver_order
         self.lower_order_nums = 0
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
 
     @property
     def step_index(self):
@@ -254,6 +255,7 @@ def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.devic
 
         # add an index counter for schedulers that allow duplicated timesteps
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
 
     # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
     def _threshold_sample(self, sample: torch.FloatTensor) -> torch.FloatTensor:

src/diffusers/schedulers/scheduling_dpmsolver_multistep.py

@@ -214,6 +214,7 @@ def __init__(
         self.model_outputs = [None] * solver_order
         self.lower_order_nums = 0
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
 
     @property
     def step_index(self):
@@ -290,6 +291,7 @@ def set_timesteps(self, num_inference_steps: int = None, device: Union[str, torc
 
         # add an index counter for schedulers that allow duplicated timesteps
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
 
     # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
     def _threshold_sample(self, sample: torch.FloatTensor) -> torch.FloatTensor:

src/diffusers/schedulers/scheduling_dpmsolver_multistep_inverse.py

@@ -209,6 +209,7 @@ def __init__(
         self.model_outputs = [None] * solver_order
         self.lower_order_nums = 0
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
         self.use_karras_sigmas = use_karras_sigmas
 
     @property
@@ -289,6 +290,7 @@ def set_timesteps(self, num_inference_steps: int = None, device: Union[str, torc
 
         # add an index counter for schedulers that allow duplicated timesteps
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
 
     # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
     def _threshold_sample(self, sample: torch.FloatTensor) -> torch.FloatTensor:

src/diffusers/schedulers/scheduling_dpmsolver_sde.py

@@ -198,6 +198,7 @@ def __init__(
         self.noise_sampler = None
         self.noise_sampler_seed = noise_sampler_seed
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
 
     # Copied from diffusers.schedulers.scheduling_heun_discrete.HeunDiscreteScheduler.index_for_timestep
     def index_for_timestep(self, timestep, schedule_timesteps=None):
@@ -347,6 +348,7 @@ def set_timesteps(
         self.mid_point_sigma = None
 
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
         self.noise_sampler = None
 
         # for exp beta schedules, such as the one for `pipeline_shap_e.py`

src/diffusers/schedulers/scheduling_dpmsolver_singlestep.py

@@ -197,6 +197,7 @@ def __init__(
         self.sample = None
         self.order_list = self.get_order_list(num_train_timesteps)
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
 
     def get_order_list(self, num_inference_steps: int) -> List[int]:
         """
@@ -288,6 +289,7 @@ def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.devic
 
         # add an index counter for schedulers that allow duplicated timesteps
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
 
     # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
     def _threshold_sample(self, sample: torch.FloatTensor) -> torch.FloatTensor:

src/diffusers/schedulers/scheduling_euler_ancestral_discrete.py

@@ -166,6 +166,7 @@ def __init__(
         self.is_scale_input_called = False
 
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
 
     @property
     def init_noise_sigma(self):
@@ -249,6 +250,7 @@ def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.devic
 
         self.timesteps = torch.from_numpy(timesteps).to(device=device)
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
 
     # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index
     def _init_step_index(self, timestep):

src/diffusers/schedulers/scheduling_euler_discrete.py

@@ -237,6 +237,7 @@ def __init__(
         self.use_karras_sigmas = use_karras_sigmas
 
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
 
     @property
     def init_noise_sigma(self):
@@ -341,6 +342,7 @@ def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.devic
 
         self.sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)])
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
 
     def _sigma_to_t(self, sigma, log_sigmas):
         # get log sigma

src/diffusers/schedulers/scheduling_heun_discrete.py

@@ -148,6 +148,7 @@ def __init__(
         self.use_karras_sigmas = use_karras_sigmas
 
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
 
     def index_for_timestep(self, timestep, schedule_timesteps=None):
         if schedule_timesteps is None:
@@ -269,6 +270,7 @@ def set_timesteps(
         self.dt = None
 
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
 
         # (YiYi Notes: keep this for now since we are keeping add_noise function which use index_for_timestep)
         # for exp beta schedules, such as the one for `pipeline_shap_e.py`

src/diffusers/schedulers/scheduling_k_dpm_2_ancestral_discrete.py

@@ -140,6 +140,7 @@ def __init__(
         #  set all values
         self.set_timesteps(num_train_timesteps, None, num_train_timesteps)
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
 
     # Copied from diffusers.schedulers.scheduling_heun_discrete.HeunDiscreteScheduler.index_for_timestep
     def index_for_timestep(self, timestep, schedule_timesteps=None):
@@ -295,6 +296,7 @@ def set_timesteps(
         self._index_counter = defaultdict(int)
 
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
 
     # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t
     def _sigma_to_t(self, sigma, log_sigmas):

src/diffusers/schedulers/scheduling_k_dpm_2_discrete.py

@@ -140,6 +140,7 @@ def __init__(
         self.set_timesteps(num_train_timesteps, None, num_train_timesteps)
 
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
 
     # Copied from diffusers.schedulers.scheduling_heun_discrete.HeunDiscreteScheduler.index_for_timestep
     def index_for_timestep(self, timestep, schedule_timesteps=None):
@@ -284,6 +285,7 @@ def set_timesteps(
         self._index_counter = defaultdict(int)
 
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
 
     @property
     def state_in_first_order(self):

src/diffusers/schedulers/scheduling_lms_discrete.py

@@ -168,6 +168,7 @@ def __init__(
         self.is_scale_input_called = False
 
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
 
     @property
     def init_noise_sigma(self):
@@ -279,6 +280,7 @@ def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.devic
         self.sigmas = torch.from_numpy(sigmas).to(device=device)
         self.timesteps = torch.from_numpy(timesteps).to(device=device)
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
 
         self.derivatives = []
 

src/diffusers/schedulers/scheduling_unipc_multistep.py

@@ -198,6 +198,7 @@ def __init__(
         self.solver_p = solver_p
         self.last_sample = None
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
 
     @property
     def step_index(self):
@@ -268,6 +269,7 @@ def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.devic
 
         # add an index counter for schedulers that allow duplicated timesteps
         self._step_index = None
+        self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication
 
     # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
     def _threshold_sample(self, sample: torch.FloatTensor) -> torch.FloatTensor: