do not use autocast for diffusers (#2349)

fixes #2345

ldm/generate.py

@@ -29,7 +29,7 @@ from ldm.invoke.concepts_lib import HuggingFaceConceptsLibrary
 from ldm.invoke.conditioning import get_uc_and_c_and_ec
 from ldm.invoke.devices import choose_torch_device, choose_precision
 from ldm.invoke.generator.inpaint import infill_methods
-from ldm.invoke.globals import global_cache_dir
+from ldm.invoke.globals import global_cache_dir, Globals
 from ldm.invoke.image_util import InitImageResizer
 from ldm.invoke.model_manager import ModelManager
 from ldm.invoke.pngwriter import PngWriter
@@ -201,6 +201,7 @@ class Generate:
             self.precision = 'float32'
         if self.precision == 'auto':
             self.precision = choose_precision(self.device)
+        Globals.full_precision = self.precision=='float32'
 
         # model caching system for fast switching
         self.model_manager = ModelManager(mconfig,self.device,self.precision,max_loaded_models=max_loaded_models)

ldm/invoke/ckpt_generator/base.py

@@ -335,4 +335,5 @@ class CkptGenerator():
             os.makedirs(dirname, exist_ok=True)
         image.save(filepath,'PNG')
 
-
+    def torch_dtype(self)->torch.dtype:
+        return torch.float16 if self.precision == 'float16' else torch.float32

ldm/invoke/ckpt_generator/txt2img.py

@@ -72,16 +72,18 @@ class CkptTxt2Img(CkptGenerator):
         device         = self.model.device
         if self.use_mps_noise or device.type == 'mps':
             x = torch.randn([1,
-                                self.latent_channels,
+                             self.latent_channels,
-                                height // self.downsampling_factor,
+                             height // self.downsampling_factor,
-                                width  // self.downsampling_factor],
+                             width  // self.downsampling_factor],
-                               device='cpu').to(device)
+                            dtype=self.torch_dtype(),
+                            device='cpu').to(device)
         else:
             x = torch.randn([1,
-                                self.latent_channels,
+                             self.latent_channels,
-                                height // self.downsampling_factor,
+                             height // self.downsampling_factor,
-                                width  // self.downsampling_factor],
+                             width  // self.downsampling_factor],
-                               device=device)
+                            dtype=self.torch_dtype(),
+                            device=device)
         if self.perlin > 0.0:
             x = (1-self.perlin)*x + self.perlin*self.get_perlin_noise(width  // self.downsampling_factor, height // self.downsampling_factor)
         return x

ldm/invoke/devices.py

@@ -21,10 +21,19 @@ def choose_precision(device) -> str:
             return 'float16'
     return 'float32'
 
+def torch_dtype(device) -> torch.dtype:
+    if Globals.full_precision:
+        return torch.float32
+    if choose_precision(device) == 'float16':
+        return torch.float16
+    else:
+        return torch.float32
+
 def choose_autocast(precision):
     '''Returns an autocast context or nullcontext for the given precision string'''
     # float16 currently requires autocast to avoid errors like:
     # 'expected scalar type Half but found Float'
+    print(f'DEBUG: choose_autocast() called')
     if precision == 'autocast' or precision == 'float16':
         return autocast
     return nullcontext

ldm/invoke/generator/base.py

@@ -8,6 +8,7 @@ import os
 import os.path as osp
 import random
 import traceback
+from contextlib import nullcontext
 
 import cv2
 import numpy as np
@@ -18,8 +19,6 @@ from einops import rearrange
 from pytorch_lightning import seed_everything
 from tqdm import trange
 
-from ldm.invoke.devices import choose_autocast
-from ldm.models.diffusion.cross_attention_map_saving import AttentionMapSaver
 from ldm.models.diffusion.ddpm import DiffusionWrapper
 from ldm.util import rand_perlin_2d
 
@@ -64,7 +63,7 @@ class Generator:
                  image_callback=None, step_callback=None, threshold=0.0, perlin=0.0,
                  safety_checker:dict=None,
                  **kwargs):
-        scope = choose_autocast(self.precision)
+        scope = nullcontext
         self.safety_checker = safety_checker
         attention_maps_images = []
         attention_maps_callback = lambda saver: attention_maps_images.append(saver.get_stacked_maps_image())
@@ -236,7 +235,8 @@ class Generator:
 
     def get_perlin_noise(self,width,height):
         fixdevice = 'cpu' if (self.model.device.type == 'mps') else self.model.device
-        return torch.stack([rand_perlin_2d((height, width), (8, 8), device = self.model.device).to(fixdevice) for _ in range(self.latent_channels)], dim=0).to(self.model.device)
+        noise = torch.stack([rand_perlin_2d((height, width), (8, 8), device = self.model.device).to(fixdevice) for _ in range(self.latent_channels)], dim=0).to(self.model.device)
+        return noise
 
     def new_seed(self):
         self.seed = random.randrange(0, np.iinfo(np.uint32).max)
@@ -341,3 +341,6 @@ class Generator:
         image.save(filepath,'PNG')
 
 
+    def torch_dtype(self)->torch.dtype:
+        return torch.float16 if self.precision == 'float16' else torch.float32
+

ldm/invoke/generator/txt2img.py

@@ -36,10 +36,9 @@ class Txt2Img(Generator):
                 threshold = ThresholdSettings(threshold, warmup=0.2) if threshold else None)
             .add_scheduler_args_if_applicable(pipeline.scheduler, eta=ddim_eta))
 
-
         def make_image(x_T) -> PIL.Image.Image:
             pipeline_output = pipeline.image_from_embeddings(
-                latents=torch.zeros_like(x_T),
+                latents=torch.zeros_like(x_T,dtype=self.torch_dtype()),
                 noise=x_T,
                 num_inference_steps=steps,
                 conditioning_data=conditioning_data,
@@ -59,16 +58,18 @@ class Txt2Img(Generator):
         input_channels = min(self.latent_channels, 4)
         if self.use_mps_noise or device.type == 'mps':
             x = torch.randn([1,
-                                input_channels,
+                             input_channels,
-                                height // self.downsampling_factor,
+                             height // self.downsampling_factor,
-                                width  // self.downsampling_factor],
+                             width  // self.downsampling_factor],
-                               device='cpu').to(device)
+                            dtype=self.torch_dtype(),
+                            device='cpu').to(device)
         else:
             x = torch.randn([1,
-                                input_channels,
+                             input_channels,
-                                height // self.downsampling_factor,
+                             height // self.downsampling_factor,
-                                width  // self.downsampling_factor],
+                             width  // self.downsampling_factor],
-                               device=device)
+                            dtype=self.torch_dtype(),
+                            device=device)
         if self.perlin > 0.0:
             x = (1-self.perlin)*x + self.perlin*self.get_perlin_noise(width  // self.downsampling_factor, height // self.downsampling_factor)
         return x

ldm/invoke/generator/txt2img2img.py

@@ -90,9 +90,9 @@ class Txt2Img2Img(Generator):
     def get_noise_like(self, like: torch.Tensor):
         device = like.device
         if device.type == 'mps':
-            x = torch.randn_like(like, device='cpu').to(device)
+            x = torch.randn_like(like, device='cpu', dtype=self.torch_dtype()).to(device)
         else:
-            x = torch.randn_like(like, device=device)
+            x = torch.randn_like(like, device=device, dtype=self.torch_dtype())
         if self.perlin > 0.0:
             shape = like.shape
             x = (1-self.perlin)*x + self.perlin*self.get_perlin_noise(shape[3], shape[2])
@@ -117,10 +117,12 @@ class Txt2Img2Img(Generator):
                                 self.latent_channels,
                                 scaled_height // self.downsampling_factor,
                                 scaled_width  // self.downsampling_factor],
-                                device='cpu').to(device)
+                               dtype=self.torch_dtype(),
+                               device='cpu').to(device)
         else:
             return torch.randn([1,
                                 self.latent_channels,
                                 scaled_height // self.downsampling_factor,
                                 scaled_width  // self.downsampling_factor],
-                                device=device)
+                               dtype=self.torch_dtype(),
+                               device=device)

ldm/invoke/globals.py

@@ -43,6 +43,9 @@ Globals.always_use_cpu = False
 # The CLI will test connectivity at startup time.
 Globals.internet_available = True
 
+# whether we are forcing full precision
+Globals.full_precision = False
+
 def global_config_dir()->Path:
     return Path(Globals.root, Globals.config_dir)
 

ldm/invoke/model_manager.py

@@ -349,7 +349,7 @@ class ModelManager(object):
 
         if self.precision == 'float16':
             print('   | Using faster float16 precision')
-            model.to(torch.float16)
+            model = model.to(torch.float16)
         else:
             print('   | Using more accurate float32 precision')
 
@@ -761,7 +761,7 @@ class ModelManager(object):
         for model in legacy_locations:
             source = models_dir /model
             if source.exists():
-                print(f'DEBUG: Moving {models_dir / model} into hub')
+                print(f'** Moving {models_dir / model} into hub')
                 move(models_dir / model, hub)
 
         # anything else gets moved into the diffusers directory

ldm/models/diffusion/cross_attention_control.py

@@ -7,6 +7,7 @@ import torch
 import diffusers
 from torch import nn
 from diffusers.models.unet_2d_condition import UNet2DConditionModel
+from ldm.invoke.devices import torch_dtype
 
 # adapted from bloc97's CrossAttentionControl colab
 # https://github.com/bloc97/CrossAttentionControl
@@ -383,7 +384,7 @@ def inject_attention_function(unet, context: Context):
                 remapped_saved_attention_slice = torch.index_select(saved_attention_slice, -1, index_map)
                 this_attention_slice = suggested_attention_slice
 
-                mask = context.cross_attention_mask
+                mask = context.cross_attention_mask.to(torch_dtype(suggested_attention_slice.device))
                 saved_mask = mask
                 this_mask = 1 - mask
                 attention_slice = remapped_saved_attention_slice * saved_mask + \

ldm/modules/prompt_to_embeddings_converter.py

@@ -4,7 +4,7 @@ import torch
 from transformers import CLIPTokenizer, CLIPTextModel
 
 from ldm.modules.textual_inversion_manager import TextualInversionManager
-
+from ldm.invoke.devices import torch_dtype
 
 class WeightedPromptFragmentsToEmbeddingsConverter():
 
@@ -207,7 +207,7 @@ class WeightedPromptFragmentsToEmbeddingsConverter():
         per_token_weights += [1.0] * pad_length
 
         all_token_ids_tensor = torch.tensor(all_token_ids, dtype=torch.long, device=device)
-        per_token_weights_tensor = torch.tensor(per_token_weights, dtype=torch.float32, device=device)
+        per_token_weights_tensor = torch.tensor(per_token_weights, dtype=torch_dtype(self.text_encoder.device), device=device)
         #print(f"assembled all_token_ids_tensor with shape {all_token_ids_tensor.shape}")
         return all_token_ids_tensor, per_token_weights_tensor
 

ldm/modules/textual_inversion_manager.py

@@ -111,7 +111,6 @@ class TextualInversionManager():
         if ti.trigger_token_id is not None:
             raise ValueError(f"Tokens already injected for textual inversion with trigger '{ti.trigger_string}'")
 
-        print(f'DEBUG: Injecting token {ti.trigger_string}')
         trigger_token_id = self._get_or_create_token_id_and_assign_embedding(ti.trigger_string, ti.embedding[0])
 
         if ti.embedding_vector_length > 1:

ldm/util.py

@@ -8,6 +8,7 @@ from threading import Thread
 from urllib import request
 from tqdm import tqdm
 from pathlib import Path
+from ldm.invoke.devices import torch_dtype
 
 import numpy as np
 import torch
@@ -235,7 +236,8 @@ def rand_perlin_2d(shape, res, device, fade = lambda t: 6*t**5 - 15*t**4 + 10*t*
     n01 = dot(tile_grads([0, -1],[1, None]), [0, -1]).to(device)
     n11 = dot(tile_grads([1, None], [1, None]), [-1,-1]).to(device)
     t = fade(grid[:shape[0], :shape[1]])
-    return math.sqrt(2) * torch.lerp(torch.lerp(n00, n10, t[..., 0]), torch.lerp(n01, n11, t[..., 0]), t[..., 1]).to(device)
+    noise = math.sqrt(2) * torch.lerp(torch.lerp(n00, n10, t[..., 0]), torch.lerp(n01, n11, t[..., 0]), t[..., 1]).to(device)
+    return noise.to(dtype=torch_dtype(device))
 
 def ask_user(question: str, answers: list):
     from itertools import chain, repeat