fix dtype bugs

diffsynth_engine/models/utils.py

@@ -2,7 +2,6 @@
 import torch.nn as nn
 from contextlib import contextmanager
 
-
 # mofified from transformers.modeling_utils
 TORCH_INIT_FUNCTIONS = {
     "uniform_": nn.init.uniform_,

diffsynth_engine/models/vae/vae.py

@@ -167,6 +167,8 @@ def __init__(
         self.conv_norm_out = nn.GroupNorm(num_channels=128, num_groups=32, eps=1e-6, device=device, dtype=dtype)
         self.conv_act = nn.SiLU()
         self.conv_out = nn.Conv2d(128, 3, kernel_size=3, padding=1, device=device, dtype=dtype)
+        self.device = device
+        self.dtype = dtype
 
     def forward(self, sample, tiled=False, tile_size=64, tile_stride=32, **kwargs):
         original_dtype = sample.dtype
@@ -277,6 +279,8 @@ def __init__(
         self.conv_norm_out = nn.GroupNorm(num_channels=512, num_groups=32, eps=1e-6, device=device, dtype=dtype)
         self.conv_act = nn.SiLU()
         self.conv_out = nn.Conv2d(512, 2 * latent_channels, kernel_size=3, padding=1, device=device, dtype=dtype)
+        self.device = device
+        self.dtype = dtype
 
     def forward(self, sample, tiled=False, tile_size=64, tile_stride=32, **kwargs):
         original_dtype = sample.dtype

diffsynth_engine/pipelines/base.py

@@ -8,6 +8,7 @@
 from diffsynth_engine.utils.offload import enable_sequential_cpu_offload
 from diffsynth_engine.utils.gguf import load_gguf_checkpoint
 from diffsynth_engine.utils import logging
+from diffsynth_engine.utils.platform import empty_cache
 
 logger = logging.get_logger(__name__)
 
@@ -144,15 +145,17 @@ def generate_noise(shape, seed=None, device="cpu", dtype=torch.float16):
         return noise
 
     def encode_image(self, image: torch.Tensor) -> torch.Tensor:
+        image = image.to(self.device, self.vae_encoder.dtype)
         latents = self.vae_encoder(
             image, tiled=self.vae_tiled, tile_size=self.vae_tile_size, tile_stride=self.vae_tile_stride
         )
         return latents
 
     def decode_image(self, latent: torch.Tensor) -> torch.Tensor:
         vae_dtype = self.vae_decoder.conv_in.weight.dtype
+        latent = latent.to(self.device, vae_dtype)
         image = self.vae_decoder(
-            latent.to(vae_dtype), tiled=self.vae_tiled, tile_size=self.vae_tile_size, tile_stride=self.vae_tile_stride
+            latent, tiled=self.vae_tiled, tile_size=self.vae_tile_size, tile_stride=self.vae_tile_stride
         )
         return image
 
@@ -233,7 +236,7 @@ def load_models_to_device(self, load_model_names: List[str] | None = None):
             return
         if self.offload_mode == "sequential_cpu_offload":
             # fresh the cuda cache
-            torch.cuda.empty_cache()
+            empty_cache()
             return
 
         # offload unnecessary models to cpu
@@ -248,4 +251,4 @@ def load_models_to_device(self, load_model_names: List[str] | None = None):
             if model is not None and (p := next(model.parameters(), None)) is not None and p.device != self.device:
                 model.to(self.device)
         # fresh the cuda cache
-        torch.cuda.empty_cache()
+        empty_cache()

diffsynth_engine/pipelines/flux_image.py

@@ -25,6 +25,7 @@
 from diffsynth_engine.utils import logging
 from diffsynth_engine.utils.fp8_linear import enable_fp8_linear
 from diffsynth_engine.utils.download import fetch_model
+from diffsynth_engine.utils.platform import empty_cache
 
 logger = logging.get_logger(__name__)
 
@@ -546,6 +547,7 @@ def predict_noise(
             current_step=current_step,
             total_step=total_step,
         )
+        self.load_models_to_device(["dit"])
         noise_pred = self.dit(
             hidden_states=latents,
             timestep=timestep,
@@ -570,15 +572,14 @@ def prepare_latents(
     ):
         # Prepare scheduler
         if input_image is not None:
+            self.load_models_to_device(["vae_encoder"])
             total_steps = num_inference_steps
             sigmas, timesteps = self.noise_scheduler.schedule(
                 total_steps, mu=mu, sigma_min=1 / total_steps, sigma_max=1.0
             )
             t_start = max(total_steps - int(num_inference_steps * denoising_strength), 1)
             sigma_start, sigmas = sigmas[t_start - 1], sigmas[t_start - 1 :]
             timesteps = timesteps[t_start - 1 :]
-
-            self.load_models_to_device(["vae_encoder"])
             noise = latents
             image = self.preprocess_image(input_image).to(device=self.device, dtype=self.dtype)
             latents = self.encode_image(image)
@@ -593,6 +594,7 @@ def prepare_latents(
         return init_latents, latents, sigmas, timesteps
 
     def prepare_masked_latent(self, image: Image.Image, mask: Image.Image | None, height: int, width: int):
+        self.load_models_to_device(["vae_encoder"])
         if mask is None:
             image = image.resize((width, height))
             image = self.preprocess_image(image).to(device=self.device, dtype=self.dtype)
@@ -637,6 +639,8 @@ def predict_multicontrolnet(
         total_step: int,
     ):
         double_block_output_results, single_block_output_results = None, None
+        if len(controlnet_params) > 0:
+            self.load_models_to_device([])
         for param in controlnet_params:
             current_scale = param.scale
             if not (
@@ -645,6 +649,9 @@ def predict_multicontrolnet(
                 # if current_step is not in the control range
                 # skip thie controlnet
                 continue
+            if self.offload_mode == "sequential_cpu_offload" or self.offload_mode == "cpu_offload":
+                empty_cache()
+                param.model.to(self.device)
             double_block_output, single_block_output = param.model(
                 latents,
                 param.image,
@@ -656,6 +663,9 @@ def predict_multicontrolnet(
                 image_ids,
                 text_ids,
             )
+            if self.offload_mode == "sequential_cpu_offload" or self.offload_mode == "cpu_offload":
+                empty_cache()
+                param.model.to("cpu")
             double_block_output_results = accumulate(double_block_output_results, double_block_output)
             single_block_output_results = accumulate(single_block_output_results, single_block_output)
         return double_block_output_results, single_block_output_results
@@ -741,7 +751,7 @@ def __call__(
         )
 
         # Denoise
-        self.load_models_to_device(["dit"])
+        self.load_models_to_device([])
         for i, timestep in enumerate(tqdm(timesteps)):
             timestep = timestep.unsqueeze(0).to(dtype=self.dtype)
             noise_pred = self.predict_noise_with_cfg(

diffsynth_engine/tools/flux_inpainting_tool.py

@@ -13,7 +13,9 @@ def __init__(
         dtype: torch.dtype = torch.bfloat16,
         offload_mode: Optional[str] = None,
     ):
-        self.pipe = FluxImagePipeline.from_pretrained(flux_model_path, device=device, offload_mode=offload_mode)
+        self.pipe = FluxImagePipeline.from_pretrained(
+            flux_model_path, device=device, offload_mode=offload_mode, dtype=dtype
+        )
         self.pipe.load_loras(lora_list)
         self.controlnet = FluxControlNet.from_pretrained(
             fetch_model(

diffsynth_engine/tools/flux_outpainting_tool.py

@@ -13,7 +13,9 @@ def __init__(
         dtype: torch.dtype = torch.bfloat16,
         offload_mode: Optional[str] = None,
     ):
-        self.pipe = FluxImagePipeline.from_pretrained(flux_model_path, device=device, offload_mode=offload_mode)
+        self.pipe = FluxImagePipeline.from_pretrained(
+            flux_model_path, device=device, offload_mode=offload_mode, dtype=dtype
+        )
         self.pipe.load_loras(lora_list)
         self.controlnet = FluxControlNet.from_pretrained(
             fetch_model(

diffsynth_engine/tools/flux_reference_tool.py

@@ -19,7 +19,7 @@ def __init__(
         offload_mode: Optional[str] = None,
     ):
         self.pipe: FluxImagePipeline = FluxImagePipeline.from_pretrained(
-            flux_model_path, load_text_encoder=load_text_encoder, device=device, offload_mode=offload_mode
+            flux_model_path, load_text_encoder=load_text_encoder, device=device, offload_mode=offload_mode, dtype=dtype
         )
         self.pipe.load_loras(lora_list)
         redux_model_path = fetch_model("muse/flux1-redux-dev", path="flux1-redux-dev.safetensors", revision="v1")

diffsynth_engine/tools/flux_replace_tool.py

@@ -20,7 +20,7 @@ def __init__(
         offload_mode: Optional[str] = None,
     ):
         self.pipe: FluxImagePipeline = FluxImagePipeline.from_pretrained(
-            flux_model_path, load_text_encoder=load_text_encoder, device=device, offload_mode=offload_mode
+            flux_model_path, load_text_encoder=load_text_encoder, device=device, offload_mode=offload_mode, dtype=dtype
         )
         self.pipe.load_loras(lora_list)
         redux_model_path = fetch_model("muse/flux1-redux-dev", path="flux1-redux-dev.safetensors", revision="v1")

diffsynth_engine/utils/platform.py

@@ -0,0 +1,12 @@
+import torch
+import gc
+
+# 存放跨平台的工具类
+
+
+def empty_cache():
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+    if torch.mps.is_available():
+        torch.mps.empty_cache()
+    gc.collect()