Merge pull request #7309 from brkirch/fix-embeddings

Fix embeddings, upscalers, and refactor `--upcast-sampling`

modules/devices.py

@@ -87,6 +87,14 @@ dtype_unet = torch.float16
 unet_needs_upcast = False
 
 
+def cond_cast_unet(input):
+    return input.to(dtype_unet) if unet_needs_upcast else input
+
+
+def cond_cast_float(input):
+    return input.float() if unet_needs_upcast else input
+
+
 def randn(seed, shape):
     torch.manual_seed(seed)
     if device.type == 'mps':
@@ -199,6 +207,3 @@ if has_mps():
         cumsum_needs_bool_fix = not torch.BoolTensor([True,True]).to(device=torch.device("mps"), dtype=torch.int64).equal(torch.BoolTensor([True,False]).to(torch.device("mps")).cumsum(0))
         torch.cumsum = lambda input, *args, **kwargs: ( cumsum_fix(input, orig_cumsum, *args, **kwargs) )
         torch.Tensor.cumsum = lambda self, *args, **kwargs: ( cumsum_fix(self, orig_Tensor_cumsum, *args, **kwargs) )
-        orig_narrow = torch.narrow
-        torch.narrow = lambda *args, **kwargs: ( orig_narrow(*args, **kwargs).clone() )
-

modules/processing.py

@@ -173,8 +173,7 @@ class StableDiffusionProcessing:
         midas_in = torch.from_numpy(transformed["midas_in"][None, ...]).to(device=shared.device)
         midas_in = repeat(midas_in, "1 ... -> n ...", n=self.batch_size)
 
-        conditioning_image = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(source_image.to(devices.dtype_vae) if devices.unet_needs_upcast else source_image))
+        conditioning_image = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(source_image))
-        conditioning_image = conditioning_image.float() if devices.unet_needs_upcast else conditioning_image
         conditioning = torch.nn.functional.interpolate(
             self.sd_model.depth_model(midas_in),
             size=conditioning_image.shape[2:],
@@ -218,7 +217,7 @@ class StableDiffusionProcessing:
         )
 
         # Encode the new masked image using first stage of network.
-        conditioning_image = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(conditioning_image.to(devices.dtype_vae) if devices.unet_needs_upcast else conditioning_image))
+        conditioning_image = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(conditioning_image))
 
         # Create the concatenated conditioning tensor to be fed to `c_concat`
         conditioning_mask = torch.nn.functional.interpolate(conditioning_mask, size=latent_image.shape[-2:])
@@ -229,16 +228,18 @@ class StableDiffusionProcessing:
         return image_conditioning
 
     def img2img_image_conditioning(self, source_image, latent_image, image_mask=None):
+        source_image = devices.cond_cast_float(source_image)
+
         # HACK: Using introspection as the Depth2Image model doesn't appear to uniquely
         # identify itself with a field common to all models. The conditioning_key is also hybrid.
         if isinstance(self.sd_model, LatentDepth2ImageDiffusion):
-            return self.depth2img_image_conditioning(source_image.float() if devices.unet_needs_upcast else source_image)
+            return self.depth2img_image_conditioning(source_image)
 
         if self.sd_model.cond_stage_key == "edit":
             return self.edit_image_conditioning(source_image)
 
         if self.sampler.conditioning_key in {'hybrid', 'concat'}:
-            return self.inpainting_image_conditioning(source_image.float() if devices.unet_needs_upcast else source_image, latent_image, image_mask=image_mask)
+            return self.inpainting_image_conditioning(source_image, latent_image, image_mask=image_mask)
 
         # Dummy zero conditioning if we're not using inpainting or depth model.
         return latent_image.new_zeros(latent_image.shape[0], 5, 1, 1)
@@ -418,7 +419,7 @@ def create_random_tensors(shape, seeds, subseeds=None, subseed_strength=0.0, see
 
 def decode_first_stage(model, x):
     with devices.autocast(disable=x.dtype == devices.dtype_vae):
-        x = model.decode_first_stage(x.to(devices.dtype_vae) if devices.unet_needs_upcast else x)
+        x = model.decode_first_stage(x)
 
     return x
 
@@ -1007,7 +1008,7 @@ class StableDiffusionProcessingImg2Img(StableDiffusionProcessing):
 
         image = torch.from_numpy(batch_images)
         image = 2. * image - 1.
-        image = image.to(device=shared.device, dtype=devices.dtype_vae if devices.unet_needs_upcast else None)
+        image = image.to(shared.device)
 
         self.init_latent = self.sd_model.get_first_stage_encoding(self.sd_model.encode_first_stage(image))
 

modules/realesrgan_model.py

@@ -46,7 +46,7 @@ class UpscalerRealESRGAN(Upscaler):
             scale=info.scale,
             model_path=info.local_data_path,
             model=info.model(),
-            half=not cmd_opts.no_half,
+            half=not cmd_opts.no_half and not cmd_opts.upcast_sampling,
             tile=opts.ESRGAN_tile,
             tile_pad=opts.ESRGAN_tile_overlap,
         )

modules/sd_hijack.py

@@ -173,7 +173,7 @@ class EmbeddingsWithFixes(torch.nn.Module):
         vecs = []
         for fixes, tensor in zip(batch_fixes, inputs_embeds):
             for offset, embedding in fixes:
-                emb = embedding.vec
+                emb = devices.cond_cast_unet(embedding.vec)
                 emb_len = min(tensor.shape[0] - offset - 1, emb.shape[0])
                 tensor = torch.cat([tensor[0:offset + 1], emb[0:emb_len], tensor[offset + 1 + emb_len:]])
 

modules/sd_hijack_unet.py

@@ -55,8 +55,14 @@ class GELUHijack(torch.nn.GELU, torch.nn.Module):
 
 unet_needs_upcast = lambda *args, **kwargs: devices.unet_needs_upcast
 CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.apply_model', apply_model, unet_needs_upcast)
-CondFunc('ldm.modules.diffusionmodules.openaimodel.timestep_embedding', lambda orig_func, *args, **kwargs: orig_func(*args, **kwargs).to(devices.dtype_unet), unet_needs_upcast)
+CondFunc('ldm.modules.diffusionmodules.openaimodel.timestep_embedding', lambda orig_func, timesteps, *args, **kwargs: orig_func(timesteps, *args, **kwargs).to(torch.float32 if timesteps.dtype == torch.int64 else devices.dtype_unet), unet_needs_upcast)
 if version.parse(torch.__version__) <= version.parse("1.13.1"):
     CondFunc('ldm.modules.diffusionmodules.util.GroupNorm32.forward', lambda orig_func, self, *args, **kwargs: orig_func(self.float(), *args, **kwargs), unet_needs_upcast)
     CondFunc('ldm.modules.attention.GEGLU.forward', lambda orig_func, self, x: orig_func(self.float(), x.float()).to(devices.dtype_unet), unet_needs_upcast)
     CondFunc('open_clip.transformer.ResidualAttentionBlock.__init__', lambda orig_func, *args, **kwargs: kwargs.update({'act_layer': GELUHijack}) and False or orig_func(*args, **kwargs), lambda _, *args, **kwargs: kwargs.get('act_layer') is None or kwargs['act_layer'] == torch.nn.GELU)
+
+first_stage_cond = lambda _, self, *args, **kwargs: devices.unet_needs_upcast and self.model.diffusion_model.dtype == torch.float16
+first_stage_sub = lambda orig_func, self, x, **kwargs: orig_func(self, x.to(devices.dtype_vae), **kwargs)
+CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.decode_first_stage', first_stage_sub, first_stage_cond)
+CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.encode_first_stage', first_stage_sub, first_stage_cond)
+CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.get_first_stage_encoding', lambda orig_func, *args, **kwargs: orig_func(*args, **kwargs).float(), first_stage_cond)

webui-macos-env.sh

@@ -10,7 +10,7 @@ then
 fi
 
 export install_dir="$HOME"
-export COMMANDLINE_ARGS="--skip-torch-cuda-test --upcast-sampling --use-cpu interrogate"
+export COMMANDLINE_ARGS="--skip-torch-cuda-test --upcast-sampling --no-half-vae --use-cpu interrogate"
 export TORCH_COMMAND="pip install torch==1.12.1 torchvision==0.13.1"
 export K_DIFFUSION_REPO="https://github.com/brkirch/k-diffusion.git"
 export K_DIFFUSION_COMMIT_HASH="51c9778f269cedb55a4d88c79c0246d35bdadb71"