Merge branch 'main' into main

jiqing-feng · web-flow · commit 432700b6cb80 · 2025-01-21T10:37:40.000+08:00
diff --git a/docs/source/en/api/pipelines/mochi.md b/docs/source/en/api/pipelines/mochi.md
@@ -115,7 +115,7 @@ export_to_video(frames, "mochi.mp4", fps=30)
 
 ## Reproducing the results from the Genmo Mochi repo
 
-The [Genmo Mochi implementation](https://github.com/genmoai/mochi/tree/main) uses different precision values for each stage in the inference process. The text encoder and VAE use `torch.float32`, while the DiT uses `torch.bfloat16` with the [attention kernel](https://pytorch.org/docs/stable/generated/torch.nn.attention.sdpa_kernel.html#torch.nn.attention.sdpa_kernel) set to `EFFICIENT_ATTENTION`. Diffusers pipelines currently do not support setting different `dtypes` for different stages of the pipeline. In order to run inference in the same way as the the original implementation, please refer to the following example.
+The [Genmo Mochi implementation](https://github.com/genmoai/mochi/tree/main) uses different precision values for each stage in the inference process. The text encoder and VAE use `torch.float32`, while the DiT uses `torch.bfloat16` with the [attention kernel](https://pytorch.org/docs/stable/generated/torch.nn.attention.sdpa_kernel.html#torch.nn.attention.sdpa_kernel) set to `EFFICIENT_ATTENTION`. Diffusers pipelines currently do not support setting different `dtypes` for different stages of the pipeline. In order to run inference in the same way as the original implementation, please refer to the following example.
 
 <Tip>
 The original Mochi implementation zeros out empty prompts. However, enabling this option and placing the entire pipeline under autocast can lead to numerical overflows with the T5 text encoder.
diff --git a/examples/community/fresco_v2v.py b/examples/community/fresco_v2v.py
@@ -404,10 +404,11 @@ def forward(
             # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
             # This would be a good case for the `match` statement (Python 3.10+)
             is_mps = sample.device.type == "mps"
+            is_npu = sample.device.type == "npu"
             if isinstance(timestep, float):
-                dtype = torch.float32 if is_mps else torch.float64
+                dtype = torch.float32 if (is_mps or is_npu) else torch.float64
             else:
-                dtype = torch.int32 if is_mps else torch.int64
+                dtype = torch.int32 if (is_mps or is_npu) else torch.int64
             timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
         elif len(timesteps.shape) == 0:
             timesteps = timesteps[None].to(sample.device)
diff --git a/examples/community/matryoshka.py b/examples/community/matryoshka.py
@@ -2806,10 +2806,11 @@ def get_time_embed(
             # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
             # This would be a good case for the `match` statement (Python 3.10+)
             is_mps = sample.device.type == "mps"
+            is_npu = sample.device.type == "npu"
             if isinstance(timestep, float):
-                dtype = torch.float32 if is_mps else torch.float64
+                dtype = torch.float32 if (is_mps or is_npu) else torch.float64
             else:
-                dtype = torch.int32 if is_mps else torch.int64
+                dtype = torch.int32 if (is_mps or is_npu) else torch.int64
             timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
         elif len(timesteps.shape) == 0:
             timesteps = timesteps[None].to(sample.device)
diff --git a/examples/research_projects/pixart/pipeline_pixart_alpha_controlnet.py b/examples/research_projects/pixart/pipeline_pixart_alpha_controlnet.py
@@ -1031,10 +1031,11 @@ def __call__(
                     # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
                     # This would be a good case for the `match` statement (Python 3.10+)
                     is_mps = latent_model_input.device.type == "mps"
+                    is_npu = latent_model_input.device.type == "npu"
                     if isinstance(current_timestep, float):
-                        dtype = torch.float32 if is_mps else torch.float64
+                        dtype = torch.float32 if (is_mps or is_npu) else torch.float64
                     else:
-                        dtype = torch.int32 if is_mps else torch.int64
+                        dtype = torch.int32 if (is_mps or is_npu) else torch.int64
                     current_timestep = torch.tensor([current_timestep], dtype=dtype, device=latent_model_input.device)
                 elif len(current_timestep.shape) == 0:
                     current_timestep = current_timestep[None].to(latent_model_input.device)
diff --git a/examples/research_projects/promptdiffusion/promptdiffusioncontrolnet.py b/examples/research_projects/promptdiffusion/promptdiffusioncontrolnet.py
@@ -258,10 +258,11 @@ def forward(
             # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
             # This would be a good case for the `match` statement (Python 3.10+)
             is_mps = sample.device.type == "mps"
+            is_npu = sample.device.type == "npu"
             if isinstance(timestep, float):
-                dtype = torch.float32 if is_mps else torch.float64
+                dtype = torch.float32 if (is_mps or is_npu) else torch.float64
             else:
-                dtype = torch.int32 if is_mps else torch.int64
+                dtype = torch.int32 if (is_mps or is_npu) else torch.int64
             timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
         elif len(timesteps.shape) == 0:
             timesteps = timesteps[None].to(sample.device)
diff --git a/scripts/convert_consistency_decoder.py b/scripts/convert_consistency_decoder.py
@@ -73,7 +73,7 @@ def _download(url: str, root: str):
                 loop.update(len(buffer))
 
     if insecure_hashlib.sha256(open(download_target, "rb").read()).hexdigest() != expected_sha256:
-        raise RuntimeError("Model has been downloaded but the SHA256 checksum does not not match")
+        raise RuntimeError("Model has been downloaded but the SHA256 checksum does not match")
 
     return download_target
 
diff --git a/src/diffusers/models/controlnets/controlnet.py b/src/diffusers/models/controlnets/controlnet.py
@@ -740,10 +740,11 @@ def forward(
             # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
             # This would be a good case for the `match` statement (Python 3.10+)
             is_mps = sample.device.type == "mps"
+            is_npu = sample.device.type == "npu"
             if isinstance(timestep, float):
-                dtype = torch.float32 if is_mps else torch.float64
+                dtype = torch.float32 if (is_mps or is_npu) else torch.float64
             else:
-                dtype = torch.int32 if is_mps else torch.int64
+                dtype = torch.int32 if (is_mps or is_npu) else torch.int64
             timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
         elif len(timesteps.shape) == 0:
             timesteps = timesteps[None].to(sample.device)
diff --git a/src/diffusers/models/controlnets/controlnet_sparsectrl.py b/src/diffusers/models/controlnets/controlnet_sparsectrl.py
@@ -671,10 +671,11 @@ def forward(
             # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
             # This would be a good case for the `match` statement (Python 3.10+)
             is_mps = sample.device.type == "mps"
+            is_npu = sample.device.type == "npu"
             if isinstance(timestep, float):
-                dtype = torch.float32 if is_mps else torch.float64
+                dtype = torch.float32 if (is_mps or is_npu) else torch.float64
             else:
-                dtype = torch.int32 if is_mps else torch.int64
+                dtype = torch.int32 if (is_mps or is_npu) else torch.int64
             timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
         elif len(timesteps.shape) == 0:
             timesteps = timesteps[None].to(sample.device)
diff --git a/src/diffusers/models/controlnets/controlnet_union.py b/src/diffusers/models/controlnets/controlnet_union.py
@@ -681,10 +681,11 @@ def forward(
             # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
             # This would be a good case for the `match` statement (Python 3.10+)
             is_mps = sample.device.type == "mps"
+            is_npu = sample.device.type == "npu"
             if isinstance(timestep, float):
-                dtype = torch.float32 if is_mps else torch.float64
+                dtype = torch.float32 if (is_mps or is_npu) else torch.float64
             else:
-                dtype = torch.int32 if is_mps else torch.int64
+                dtype = torch.int32 if (is_mps or is_npu) else torch.int64
             timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
         elif len(timesteps.shape) == 0:
             timesteps = timesteps[None].to(sample.device)
diff --git a/src/diffusers/models/controlnets/controlnet_xs.py b/src/diffusers/models/controlnets/controlnet_xs.py
@@ -1088,10 +1088,11 @@ def forward(
             # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
             # This would be a good case for the `match` statement (Python 3.10+)
             is_mps = sample.device.type == "mps"
+            is_npu = sample.device.type == "npu"
             if isinstance(timestep, float):
-                dtype = torch.float32 if is_mps else torch.float64
+                dtype = torch.float32 if (is_mps or is_npu) else torch.float64
             else:
-                dtype = torch.int32 if is_mps else torch.int64
+                dtype = torch.int32 if (is_mps or is_npu) else torch.int64
             timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
         elif len(timesteps.shape) == 0:
             timesteps = timesteps[None].to(sample.device)
diff --git a/src/diffusers/models/unets/unet_2d_condition.py b/src/diffusers/models/unets/unet_2d_condition.py
@@ -915,10 +915,11 @@ def get_time_embed(
             # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
             # This would be a good case for the `match` statement (Python 3.10+)
             is_mps = sample.device.type == "mps"
+            is_npu = sample.device.type == "npu"
             if isinstance(timestep, float):
-                dtype = torch.float32 if is_mps else torch.float64
+                dtype = torch.float32 if (is_mps or is_npu) else torch.float64
             else:
-                dtype = torch.int32 if is_mps else torch.int64
+                dtype = torch.int32 if (is_mps or is_npu) else torch.int64
             timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
         elif len(timesteps.shape) == 0:
             timesteps = timesteps[None].to(sample.device)
diff --git a/src/diffusers/models/unets/unet_3d_condition.py b/src/diffusers/models/unets/unet_3d_condition.py
@@ -624,10 +624,11 @@ def forward(
             # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
             # This would be a good case for the `match` statement (Python 3.10+)
             is_mps = sample.device.type == "mps"
+            is_npu = sample.device.type == "npu"
             if isinstance(timestep, float):
-                dtype = torch.float32 if is_mps else torch.float64
+                dtype = torch.float32 if (is_mps or is_npu) else torch.float64
             else:
-                dtype = torch.int32 if is_mps else torch.int64
+                dtype = torch.int32 if (is_mps or is_npu) else torch.int64
             timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
         elif len(timesteps.shape) == 0:
             timesteps = timesteps[None].to(sample.device)
diff --git a/src/diffusers/models/unets/unet_i2vgen_xl.py b/src/diffusers/models/unets/unet_i2vgen_xl.py
@@ -575,10 +575,11 @@ def forward(
             # TODO: this requires sync between CPU and GPU. So try to pass `timesteps` as tensors if you can
             # This would be a good case for the `match` statement (Python 3.10+)
             is_mps = sample.device.type == "mps"
+            is_npu = sample.device.type == "npu"
             if isinstance(timesteps, float):
-                dtype = torch.float32 if is_mps else torch.float64
+                dtype = torch.float32 if (is_mps or is_npu) else torch.float64
             else:
-                dtype = torch.int32 if is_mps else torch.int64
+                dtype = torch.int32 if (is_mps or is_npu) else torch.int64
             timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
         elif len(timesteps.shape) == 0:
             timesteps = timesteps[None].to(sample.device)
diff --git a/src/diffusers/models/unets/unet_motion_model.py b/src/diffusers/models/unets/unet_motion_model.py
@@ -2114,10 +2114,11 @@ def forward(
             # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
             # This would be a good case for the `match` statement (Python 3.10+)
             is_mps = sample.device.type == "mps"
+            is_npu = sample.device.type == "npu"
             if isinstance(timestep, float):
-                dtype = torch.float32 if is_mps else torch.float64
+                dtype = torch.float32 if (is_mps or is_npu) else torch.float64
             else:
-                dtype = torch.int32 if is_mps else torch.int64
+                dtype = torch.int32 if (is_mps or is_npu) else torch.int64
             timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
         elif len(timesteps.shape) == 0:
             timesteps = timesteps[None].to(sample.device)
diff --git a/src/diffusers/models/unets/unet_spatio_temporal_condition.py b/src/diffusers/models/unets/unet_spatio_temporal_condition.py
@@ -402,10 +402,11 @@ def forward(
             # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
             # This would be a good case for the `match` statement (Python 3.10+)
             is_mps = sample.device.type == "mps"
+            is_npu = sample.device.type == "npu"
             if isinstance(timestep, float):
-                dtype = torch.float32 if is_mps else torch.float64
+                dtype = torch.float32 if (is_mps or is_npu) else torch.float64
             else:
-                dtype = torch.int32 if is_mps else torch.int64
+                dtype = torch.int32 if (is_mps or is_npu) else torch.int64
             timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
         elif len(timesteps.shape) == 0:
             timesteps = timesteps[None].to(sample.device)
diff --git a/src/diffusers/optimization.py b/src/diffusers/optimization.py
@@ -258,7 +258,7 @@ def get_polynomial_decay_schedule_with_warmup(
 
     lr_init = optimizer.defaults["lr"]
     if not (lr_init > lr_end):
-        raise ValueError(f"lr_end ({lr_end}) must be be smaller than initial lr ({lr_init})")
+        raise ValueError(f"lr_end ({lr_end}) must be smaller than initial lr ({lr_init})")
 
     def lr_lambda(current_step: int):
         if current_step < num_warmup_steps:
diff --git a/src/diffusers/pipelines/audioldm2/modeling_audioldm2.py b/src/diffusers/pipelines/audioldm2/modeling_audioldm2.py
@@ -768,10 +768,11 @@ def forward(
             # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
             # This would be a good case for the `match` statement (Python 3.10+)
             is_mps = sample.device.type == "mps"
+            is_npu = sample.device.type == "npu"
             if isinstance(timestep, float):
-                dtype = torch.float32 if is_mps else torch.float64
+                dtype = torch.float32 if (is_mps or is_npu) else torch.float64
             else:
-                dtype = torch.int32 if is_mps else torch.int64
+                dtype = torch.int32 if (is_mps or is_npu) else torch.int64
             timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
         elif len(timesteps.shape) == 0:
             timesteps = timesteps[None].to(sample.device)
diff --git a/src/diffusers/pipelines/deprecated/versatile_diffusion/modeling_text_unet.py b/src/diffusers/pipelines/deprecated/versatile_diffusion/modeling_text_unet.py
@@ -1163,10 +1163,11 @@ def forward(
             # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
             # This would be a good case for the `match` statement (Python 3.10+)
             is_mps = sample.device.type == "mps"
+            is_npu = sample.device.type == "npu"
             if isinstance(timestep, float):
-                dtype = torch.float32 if is_mps else torch.float64
+                dtype = torch.float32 if (is_mps or is_npu) else torch.float64
             else:
-                dtype = torch.int32 if is_mps else torch.int64
+                dtype = torch.int32 if (is_mps or is_npu) else torch.int64
             timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
         elif len(timesteps.shape) == 0:
             timesteps = timesteps[None].to(sample.device)
diff --git a/src/diffusers/pipelines/dit/pipeline_dit.py b/src/diffusers/pipelines/dit/pipeline_dit.py
@@ -187,10 +187,11 @@ def __call__(
                 # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
                 # This would be a good case for the `match` statement (Python 3.10+)
                 is_mps = latent_model_input.device.type == "mps"
+                is_npu = latent_model_input.device.type == "npu"
                 if isinstance(timesteps, float):
-                    dtype = torch.float32 if is_mps else torch.float64
+                    dtype = torch.float32 if (is_mps or is_npu) else torch.float64
                 else:
-                    dtype = torch.int32 if is_mps else torch.int64
+                    dtype = torch.int32 if (is_mps or is_npu) else torch.int64
                 timesteps = torch.tensor([timesteps], dtype=dtype, device=latent_model_input.device)
             elif len(timesteps.shape) == 0:
                 timesteps = timesteps[None].to(latent_model_input.device)
diff --git a/src/diffusers/pipelines/latte/pipeline_latte.py b/src/diffusers/pipelines/latte/pipeline_latte.py
@@ -798,10 +798,11 @@ def __call__(
                     # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
                     # This would be a good case for the `match` statement (Python 3.10+)
                     is_mps = latent_model_input.device.type == "mps"
+                    is_npu = latent_model_input.device.type == "npu"
                     if isinstance(current_timestep, float):
-                        dtype = torch.float32 if is_mps else torch.float64
+                        dtype = torch.float32 if (is_mps or is_npu) else torch.float64
                     else:
-                        dtype = torch.int32 if is_mps else torch.int64
+                        dtype = torch.int32 if (is_mps or is_npu) else torch.int64
                     current_timestep = torch.tensor([current_timestep], dtype=dtype, device=latent_model_input.device)
                 elif len(current_timestep.shape) == 0:
                     current_timestep = current_timestep[None].to(latent_model_input.device)
diff --git a/src/diffusers/pipelines/lumina/pipeline_lumina.py b/src/diffusers/pipelines/lumina/pipeline_lumina.py
@@ -806,10 +806,11 @@ def __call__(
                     # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
                     # This would be a good case for the `match` statement (Python 3.10+)
                     is_mps = latent_model_input.device.type == "mps"
+                    is_npu = latent_model_input.device.type == "npu"
                     if isinstance(current_timestep, float):
-                        dtype = torch.float32 if is_mps else torch.float64
+                        dtype = torch.float32 if (is_mps or is_npu) else torch.float64
                     else:
-                        dtype = torch.int32 if is_mps else torch.int64
+                        dtype = torch.int32 if (is_mps or is_npu) else torch.int64
                     current_timestep = torch.tensor(
                         [current_timestep],
                         dtype=dtype,
diff --git a/src/diffusers/pipelines/pag/pag_utils.py b/src/diffusers/pipelines/pag/pag_utils.py
@@ -158,7 +158,7 @@ def set_pag_applied_layers(
         ),
     ):
         r"""
-        Set the the self-attention layers to apply PAG. Raise ValueError if the input is invalid.
+        Set the self-attention layers to apply PAG. Raise ValueError if the input is invalid.
 
         Args:
             pag_applied_layers (`str` or `List[str]`):
diff --git a/src/diffusers/pipelines/pag/pipeline_pag_pixart_sigma.py b/src/diffusers/pipelines/pag/pipeline_pag_pixart_sigma.py
@@ -807,10 +807,11 @@ def __call__(
                     # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
                     # This would be a good case for the `match` statement (Python 3.10+)
                     is_mps = latent_model_input.device.type == "mps"
+                    is_npu = latent_model_input.device.type == "npu"
                     if isinstance(current_timestep, float):
-                        dtype = torch.float32 if is_mps else torch.float64
+                        dtype = torch.float32 if (is_mps or is_npu) else torch.float64
                     else:
-                        dtype = torch.int32 if is_mps else torch.int64
+                        dtype = torch.int32 if (is_mps or is_npu) else torch.int64
                     current_timestep = torch.tensor([current_timestep], dtype=dtype, device=latent_model_input.device)
                 elif len(current_timestep.shape) == 0:
                     current_timestep = current_timestep[None].to(latent_model_input.device)
diff --git a/src/diffusers/pipelines/pixart_alpha/pipeline_pixart_alpha.py b/src/diffusers/pipelines/pixart_alpha/pipeline_pixart_alpha.py
diff --git a/src/diffusers/pipelines/pixart_alpha/pipeline_pixart_sigma.py b/src/diffusers/pipelines/pixart_alpha/pipeline_pixart_sigma.py
diff --git a/src/diffusers/video_processor.py b/src/diffusers/video_processor.py
