Merge branch 'main' into hunyuan-video

Author: a-r-r-o-w

examples/community/pipeline_flux_rf_inversion.py

@@ -648,6 +648,8 @@ def __call__(
         height: Optional[int] = None,
         width: Optional[int] = None,
         eta: float = 1.0,
+        decay_eta: Optional[bool] = False,
+        eta_decay_power: Optional[float] = 1.0,
         strength: float = 1.0,
         start_timestep: float = 0,
         stop_timestep: float = 0.25,
@@ -880,12 +882,9 @@ def __call__(
                     v_t = -noise_pred
                     v_t_cond = (y_0 - latents) / (1 - t_i)
                     eta_t = eta if start_timestep <= i < stop_timestep else 0.0
-                    if start_timestep <= i < stop_timestep:
-                        # controlled vector field
-                        v_hat_t = v_t + eta * (v_t_cond - v_t)
-
-                    else:
-                        v_hat_t = v_t
+                    if decay_eta:
+                        eta_t = eta_t * (1 - i / num_inference_steps) ** eta_decay_power  # Decay eta over the loop
+                    v_hat_t = v_t + eta_t * (v_t_cond - v_t)
 
                     # SDE Eq: 17 from https://arxiv.org/pdf/2410.10792
                     latents = latents + v_hat_t * (sigmas[i] - sigmas[i + 1])

src/diffusers/models/activations.py

@@ -18,7 +18,7 @@
 from torch import nn
 
 from ..utils import deprecate
-from ..utils.import_utils import is_torch_npu_available
+from ..utils.import_utils import is_torch_npu_available, is_torch_version
 
 
 if is_torch_npu_available():
@@ -79,10 +79,10 @@ def __init__(self, dim_in: int, dim_out: int, approximate: str = "none", bias: b
         self.approximate = approximate
 
     def gelu(self, gate: torch.Tensor) -> torch.Tensor:
-        if gate.device.type != "mps":
-            return F.gelu(gate, approximate=self.approximate)
-        # mps: gelu is not implemented for float16
-        return F.gelu(gate.to(dtype=torch.float32), approximate=self.approximate).to(dtype=gate.dtype)
+        if gate.device.type == "mps" and is_torch_version("<", "2.0.0"):
+            # fp16 gelu not supported on mps before torch 2.0
+            return F.gelu(gate.to(dtype=torch.float32), approximate=self.approximate).to(dtype=gate.dtype)
+        return F.gelu(gate, approximate=self.approximate)
 
     def forward(self, hidden_states):
         hidden_states = self.proj(hidden_states)
@@ -105,10 +105,10 @@ def __init__(self, dim_in: int, dim_out: int, bias: bool = True):
         self.proj = nn.Linear(dim_in, dim_out * 2, bias=bias)
 
     def gelu(self, gate: torch.Tensor) -> torch.Tensor:
-        if gate.device.type != "mps":
-            return F.gelu(gate)
-        # mps: gelu is not implemented for float16
-        return F.gelu(gate.to(dtype=torch.float32)).to(dtype=gate.dtype)
+        if gate.device.type == "mps" and is_torch_version("<", "2.0.0"):
+            # fp16 gelu not supported on mps before torch 2.0
+            return F.gelu(gate.to(dtype=torch.float32)).to(dtype=gate.dtype)
+        return F.gelu(gate)
 
     def forward(self, hidden_states, *args, **kwargs):
         if len(args) > 0 or kwargs.get("scale", None) is not None: