use .to(device) to avoid copy, within one_unet_in_gpu context

dalle2_pytorch/dalle2_pytorch.py

@@ -1124,7 +1124,7 @@ class DiffusionPriorNetwork(nn.Module):
         learned_queries = repeat(self.learned_query, 'd -> b 1 d', b = batch)
 
         if self.self_cond:
-            learned_queries = torch.cat((image_embed, self_cond), dim = -2)
+            learned_queries = torch.cat((self_cond, learned_queries), dim = -2)
 
         tokens = torch.cat((
             text_encodings,
@@ -1334,10 +1334,7 @@ class DiffusionPrior(nn.Module):
 
             # predict noise
 
-            if self.predict_x_start or self.predict_v:
-                pred_noise = self.noise_scheduler.predict_noise_from_start(image_embed, t = time_cond, x0 = x_start)
-            else:
-                pred_noise = pred
+            pred_noise = self.noise_scheduler.predict_noise_from_start(image_embed, t = time_cond, x0 = x_start)
 
             if time_next < 0:
                 image_embed = x_start
@@ -2496,7 +2493,7 @@ class Decoder(nn.Module):
         dynamic_thres_percentile = 0.95,
         p2_loss_weight_gamma = 0.,                  # p2 loss weight, from https://arxiv.org/abs/2204.00227 - 0 is equivalent to weight of 1 across time - 1. is recommended
         p2_loss_weight_k = 1,
-        ddim_sampling_eta = 1.                      # can be set to 0. for deterministic sampling afaict
+        ddim_sampling_eta = 0.                      # can be set to 0. for deterministic sampling afaict
     ):
         super().__init__()
 
@@ -2730,11 +2727,16 @@ class Decoder(nn.Module):
         if exists(unet_number):
             unet = self.get_unet(unet_number)
 
+        # devices
+
+        cuda, cpu = torch.device('cuda'), torch.device('cpu')
+
         self.cuda()
 
         devices = [module_device(unet) for unet in self.unets]
-        self.unets.cpu()
-        unet.cuda()
+
+        self.unets.to(cpu)
+        unet.to(cuda)
 
         yield
 
@@ -2975,10 +2977,7 @@ class Decoder(nn.Module):
 
                 # predict noise
 
-                if predict_x_start or predict_v:
-                    pred_noise = noise_scheduler.predict_noise_from_start(img, t = time_cond, x0 = x_start)
-                else:
-                    pred_noise = pred
+                pred_noise = noise_scheduler.predict_noise_from_start(img, t = time_cond, x0 = x_start)
 
                 c1 = eta * ((1 - alpha / alpha_next) * (1 - alpha_next) / (1 - alpha)).sqrt()
                 c2 = ((1 - alpha_next) - torch.square(c1)).sqrt()
@@ -3120,7 +3119,8 @@ class Decoder(nn.Module):
         distributed = False,
         inpaint_image = None,
         inpaint_mask = None,
-        inpaint_resample_times = 5
+        inpaint_resample_times = 5,
+        one_unet_in_gpu_at_time = True
     ):
         assert self.unconditional or exists(image_embed), 'image embed must be present on sampling from decoder unless if trained unconditionally'
 
@@ -3143,6 +3143,7 @@ class Decoder(nn.Module):
             assert image.shape[0] == batch_size, 'image must have batch size of {} if starting at unet number > 1'.format(batch_size)
             prev_unet_output_size = self.image_sizes[start_at_unet_number - 2]
             img = resize_image_to(image, prev_unet_output_size, nearest = True)
+
         is_cuda = next(self.parameters()).is_cuda
 
         num_unets = self.num_unets
@@ -3152,7 +3153,7 @@ class Decoder(nn.Module):
             if unet_number < start_at_unet_number:
                 continue  # It's the easiest way to do it
 
-            context = self.one_unet_in_gpu(unet = unet) if is_cuda else null_context()
+            context = self.one_unet_in_gpu(unet = unet) if is_cuda and one_unet_in_gpu_at_time else null_context()
 
             with context:
                 # prepare low resolution conditioning for upsamplers

dalle2_pytorch/version.py

@@ -1 +1 @@
-__version__ = '1.11.4'
+__version__ = '1.12.4'

setup.py

@@ -27,7 +27,7 @@ setup(
     'accelerate',
     'click',
     'open-clip-torch>=2.0.0,<3.0.0',
-    'clip-anytorch>=2.4.0',
+    'clip-anytorch>=2.5.2',
     'coca-pytorch>=0.0.5',
     'ema-pytorch>=0.0.7',
     'einops>=0.4',