add cache for image prompts

rudalle/__init__.py

@@ -22,4 +22,4 @@ __all__ = [
     'image_prompts',
 ]
 
-__version__ = '0.0.1-rc5'
+__version__ = '0.0.1-rc6'

rudalle/dalle/transformer.py

@@ -206,7 +206,7 @@ class DalleSelfAttention(torch.nn.Module):
         self.num_attention_heads = num_attention_heads
         self.hidden_size_per_attention_head = divide(hidden_size, num_attention_heads)
 
-        self.query_key_value = torch.nn.Linear(hidden_size, 3*hidden_size)
+        self.query_key_value = torch.nn.Linear(hidden_size, 3 * hidden_size)
         self.attention_dropout = torch.nn.Dropout(attention_dropout_prob)
 
         # Output.
@@ -248,12 +248,12 @@ class DalleSelfAttention(torch.nn.Module):
             attention_scores = (attention_scores_scaled - attention_scores_scaled_maxes) * alpha
         return attention_scores
 
-    def forward(self, hidden_states, ltor_mask, has_cache=False, use_cache=False,):
+    def forward(self, hidden_states, ltor_mask, has_cache=False, use_cache=False, ):
         # hidden_states: [b, s, h]
         # ltor_mask: [1, 1, s, s]
         # Attention heads. [b, s, hp]
         if has_cache and use_cache:
-            mixed_x_layer = self.query_key_value(hidden_states[:, -1:, :])
+            mixed_x_layer = self.query_key_value(hidden_states[:, self.past_key.shape[-2]:, :])
         else:
             mixed_x_layer = self.query_key_value(hidden_states)
 
@@ -277,6 +277,10 @@ class DalleSelfAttention(torch.nn.Module):
                 query_layer=query_layer, key_layer=key_layer, ltor_mask=ltor_mask
             )
 
+        if use_cache and has_cache:
+            extra_cache_size = hidden_states.shape[-2] - self.past_key.shape[-2]
+            attention_scores = attention_scores[..., -extra_cache_size:, :]
+
         if use_cache:
             self.past_key = key_layer
             self.past_value = value_layer
@@ -286,9 +290,6 @@ class DalleSelfAttention(torch.nn.Module):
             self.past_output = None
             has_cache = False
 
-        if use_cache and has_cache:
-            attention_scores = attention_scores[..., -1:, :]
-
         # Attention probabilities. [b, np, s, s]
         attention_probs = torch.nn.Softmax(dim=-1)(attention_scores)
 
@@ -338,16 +339,16 @@ class DalleMLP(torch.nn.Module):
     def __init__(self, hidden_size, output_dropout_prob):
         super(DalleMLP, self).__init__()
         # Project to 4h.
-        self.dense_h_to_4h = torch.nn.Linear(hidden_size, 4*hidden_size)
+        self.dense_h_to_4h = torch.nn.Linear(hidden_size, 4 * hidden_size)
         # Project back to h.
-        self.dense_4h_to_h = torch.nn.Linear(4*hidden_size, hidden_size)
+        self.dense_4h_to_h = torch.nn.Linear(4 * hidden_size, hidden_size)
         self.dropout = torch.nn.Dropout(output_dropout_prob)
         # MLP cache
         self.past_x = None
 
     def forward(self, hidden_states, has_cache=False, use_cache=False):
         if has_cache and use_cache:
-            hidden_states = hidden_states[:, -1:]
+            hidden_states = hidden_states[:, self.past_x.shape[-2]:]
 
         # [b, s, 4hp]
         x = self.dense_h_to_4h(hidden_states)

rudalle/image_prompts.py

@@ -18,7 +18,6 @@ class ImagePrompts:
         self.device = device
         img = self._preprocess_img(pil_image)
         self.image_prompts_idx, self.image_prompts = self._get_image_prompts(img, borders, vae, crop_first)
-        self.allow_cache = True
 
     def _preprocess_img(self, pil_img):
         img = torch.tensor(np.array(pil_img.convert('RGB')).transpose(2, 0, 1)) / 255.
@@ -34,9 +33,6 @@ class ImagePrompts:
         else:
             _, _, [_, _, vqg_img] = vae.model.encode(img)
 
-        if borders['right'] + borders['left'] + borders['down'] != 0:
-            self.allow_cache = False  # TODO fix cache in attention
-
         bs, vqg_img_w, vqg_img_h = vqg_img.shape
         mask = torch.zeros(vqg_img_w, vqg_img_h)
         if borders['up'] != 0:

rudalle/pipelines.py

@@ -39,9 +39,6 @@ def generate_images(text, tokenizer, dalle, vae, top_k, top_p, images_num, image
             if image_prompts is not None:
                 prompts_idx, prompts = image_prompts.image_prompts_idx, image_prompts.image_prompts
                 prompts = prompts.repeat(chunk_bs, 1)
-                if use_cache and image_prompts.allow_cache is False:
-                    print('Warning: use_cache changed to False')
-                    use_cache = False
             for idx in tqdm(range(out.shape[1], total_seq_length)):
                 idx -= text_seq_length
                 if image_prompts is not None and idx in prompts_idx: