support stepvideo quantized

diffsynth/models/stepvideo_dit.py

@@ -238,7 +238,7 @@ class PixArtAlphaCombinedTimestepSizeEmbeddings(nn.Module):
             self.fps_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
 
     def forward(self, timestep, resolution=None, nframe=None, fps=None):
-        hidden_dtype = next(self.timestep_embedder.parameters()).dtype
+        hidden_dtype = timestep.dtype
 
         timesteps_proj = self.time_proj(timestep)
         timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_dtype))  # (N, D)

diffsynth/pipelines/step_video.py

@@ -181,7 +181,7 @@ class StepVideoPipeline(BasePipeline):
         # Denoise
         self.load_models_to_device(["dit"])
         for progress_id, timestep in enumerate(progress_bar_cmd(self.scheduler.timesteps)):
-            timestep = timestep.unsqueeze(0).to(self.device)
+            timestep = timestep.unsqueeze(0).to(dtype=self.torch_dtype, device=self.device)
             print(f"Step {progress_id + 1} / {len(self.scheduler.timesteps)}")
 
             # Inference