Make applying embeddings more efficient.

Adding new tokens no longer makes a whole copy of the embeddings weight
which can be massive on certain models.

comfy/clip_model.py

@@ -97,8 +97,12 @@ def __init__(self, config_dict, dtype, device, operations):
         self.encoder = CLIPEncoder(num_layers, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations)
         self.final_layer_norm = operations.LayerNorm(embed_dim, dtype=dtype, device=device)
 
-    def forward(self, input_tokens, attention_mask=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=torch.float32):
-        x = self.embeddings(input_tokens, dtype=dtype)
+    def forward(self, input_tokens=None, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=torch.float32):
+        if embeds is not None:
+            x = embeds + comfy.ops.cast_to(self.embeddings.position_embedding.weight, dtype=dtype, device=embeds.device)
+        else:
+            x = self.embeddings(input_tokens, dtype=dtype)
+
         mask = None
         if attention_mask is not None:
             mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, attention_mask.shape[-1], attention_mask.shape[-1])
@@ -116,7 +120,10 @@ def forward(self, input_tokens, attention_mask=None, intermediate_output=None, f
         if i is not None and final_layer_norm_intermediate:
             i = self.final_layer_norm(i)
 
-        pooled_output = x[torch.arange(x.shape[0], device=x.device), (torch.round(input_tokens).to(dtype=torch.int, device=x.device) == self.eos_token_id).int().argmax(dim=-1),]
+        if num_tokens is not None:
+            pooled_output = x[list(range(x.shape[0])), list(map(lambda a: a - 1, num_tokens))]
+        else:
+            pooled_output = x[torch.arange(x.shape[0], device=x.device), (torch.round(input_tokens).to(dtype=torch.int, device=x.device) == self.eos_token_id).int().argmax(dim=-1),]
         return x, i, pooled_output
 
 class CLIPTextModel(torch.nn.Module):

comfy/sd1_clip.py

@@ -158,71 +158,75 @@ def reset_clip_options(self):
         self.layer_idx = self.options_default[1]
         self.return_projected_pooled = self.options_default[2]
 
-    def set_up_textual_embeddings(self, tokens, current_embeds):
-        out_tokens = []
-        next_new_token = token_dict_size = current_embeds.weight.shape[0]
-        embedding_weights = []
+    def process_tokens(self, tokens, device):
+        end_token = self.special_tokens.get("end", None)
+        if end_token is None:
+            cmp_token = self.special_tokens.get("pad", -1)
+        else:
+            cmp_token = end_token
+
+        embeds_out = []
+        attention_masks = []
+        num_tokens = []
 
         for x in tokens:
+            attention_mask = []
             tokens_temp = []
+            other_embeds = []
+            eos = False
+            index = 0
             for y in x:
                 if isinstance(y, numbers.Integral):
-                    tokens_temp += [int(y)]
-                else:
-                    if y.shape[0] == current_embeds.weight.shape[1]:
-                        embedding_weights += [y]
-                        tokens_temp += [next_new_token]
-                        next_new_token += 1
+                    if eos:
+                        attention_mask.append(0)
                     else:
-                        logging.warning("WARNING: shape mismatch when trying to apply embedding, embedding will be ignored {} != {}".format(y.shape[0], current_embeds.weight.shape[1]))
-            while len(tokens_temp) < len(x):
-                tokens_temp += [self.special_tokens["pad"]]
-            out_tokens += [tokens_temp]
-
-        n = token_dict_size
-        if len(embedding_weights) > 0:
-            new_embedding = self.operations.Embedding(next_new_token + 1, current_embeds.weight.shape[1], device=current_embeds.weight.device, dtype=current_embeds.weight.dtype)
-            new_embedding.weight[:token_dict_size] = current_embeds.weight
-            for x in embedding_weights:
-                new_embedding.weight[n] = x
-                n += 1
-            self.transformer.set_input_embeddings(new_embedding)
-
-        processed_tokens = []
-        for x in out_tokens:
-            processed_tokens += [list(map(lambda a: n if a == -1 else a, x))] #The EOS token should always be the largest one
-
-        return processed_tokens
+                        attention_mask.append(1)
+                    token = int(y)
+                    tokens_temp += [token]
+                    if not eos and token == cmp_token:
+                        if end_token is None:
+                            attention_mask[-1] = 0
+                        eos = True
+                else:
+                    other_embeds.append((index, y))
+                index += 1
+
+            tokens_embed = torch.tensor([tokens_temp], device=device, dtype=torch.long)
+            tokens_embed = self.transformer.get_input_embeddings()(tokens_embed, out_dtype=torch.float32)
+            index = 0
+            pad_extra = 0
+            for o in other_embeds:
+                ind = index + o[0]
+                emb = o[1].view(1, -1, o[1].shape[-1]).to(device=device, dtype=torch.float32)
+                emb_shape = emb.shape[1]
+                if emb.shape[-1] == tokens_embed.shape[-1]:
+                    tokens_embed = torch.cat([tokens_embed[:, :ind], emb, tokens_embed[:, ind:]], dim=1)
+                    attention_mask = attention_mask[:ind] + [1] * emb_shape + attention_mask[ind:]
+                    index += emb_shape - 1
+                else:
+                    index += -1
+                    pad_extra += emb_shape
+                    logging.warning("WARNING: shape mismatch when trying to apply embedding, embedding will be ignored {} != {}".format(emb.shape[-1], tokens_embed.shape[-1]))
 
-    def forward(self, tokens):
-        backup_embeds = self.transformer.get_input_embeddings()
-        device = backup_embeds.weight.device
-        tokens = self.set_up_textual_embeddings(tokens, backup_embeds)
-        tokens = torch.LongTensor(tokens).to(device)
-
-        attention_mask = None
-        if self.enable_attention_masks or self.zero_out_masked or self.return_attention_masks:
-            attention_mask = torch.zeros_like(tokens)
-            end_token = self.special_tokens.get("end", None)
-            if end_token is None:
-                cmp_token = self.special_tokens.get("pad", -1)
-            else:
-                cmp_token = end_token
+            if pad_extra > 0:
+                padd_embed = self.transformer.get_input_embeddings()(torch.tensor([[self.special_tokens["pad"]] * pad_extra], device=device, dtype=torch.long), out_dtype=torch.float32)
+                tokens_embed = torch.cat([tokens_embed, padd_embed], dim=1)
 
-            for x in range(attention_mask.shape[0]):
-                for y in range(attention_mask.shape[1]):
-                    attention_mask[x, y] = 1
-                    if tokens[x, y] == cmp_token:
-                        if end_token is None:
-                            attention_mask[x, y] = 0
-                        break
+            embeds_out.append(tokens_embed)
+            attention_masks.append(attention_mask)
+            num_tokens.append(sum(attention_mask))
+
+        return torch.cat(embeds_out), torch.tensor(attention_masks, device=device, dtype=torch.long), num_tokens
+
+    def forward(self, tokens):
+        device = self.transformer.get_input_embeddings().weight.device
+        embeds, attention_mask, num_tokens = self.process_tokens(tokens, device)
 
         attention_mask_model = None
         if self.enable_attention_masks:
             attention_mask_model = attention_mask
 
-        outputs = self.transformer(tokens, attention_mask_model, intermediate_output=self.layer_idx, final_layer_norm_intermediate=self.layer_norm_hidden_state, dtype=torch.float32)
-        self.transformer.set_input_embeddings(backup_embeds)
+        outputs = self.transformer(None, attention_mask_model, embeds=embeds, num_tokens=num_tokens, intermediate_output=self.layer_idx, final_layer_norm_intermediate=self.layer_norm_hidden_state, dtype=torch.float32)
 
         if self.layer == "last":
             z = outputs[0].float()

comfy/text_encoders/bert.py

@@ -93,8 +93,11 @@ def __init__(self, vocab_size, max_position_embeddings, type_vocab_size, pad_tok
 
         self.LayerNorm = operations.LayerNorm(embed_dim, eps=layer_norm_eps, dtype=dtype, device=device)
 
-    def forward(self, input_tokens, token_type_ids=None, dtype=None):
-        x = self.word_embeddings(input_tokens, out_dtype=dtype)
+    def forward(self, input_tokens, embeds=None, token_type_ids=None, dtype=None):
+        if embeds is not None:
+            x = embeds
+        else:
+            x = self.word_embeddings(input_tokens, out_dtype=dtype)
         x += comfy.ops.cast_to_input(self.position_embeddings.weight[:x.shape[1]], x)
         if token_type_ids is not None:
             x += self.token_type_embeddings(token_type_ids, out_dtype=x.dtype)
@@ -113,8 +116,8 @@ def __init__(self, config_dict, dtype, device, operations):
         self.embeddings = BertEmbeddings(config_dict["vocab_size"], config_dict["max_position_embeddings"], config_dict["type_vocab_size"], config_dict["pad_token_id"], embed_dim, layer_norm_eps, dtype, device, operations)
         self.encoder = BertEncoder(config_dict["num_hidden_layers"], embed_dim, config_dict["intermediate_size"], config_dict["num_attention_heads"], layer_norm_eps, dtype, device, operations)
 
-    def forward(self, input_tokens, attention_mask=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None):
-        x = self.embeddings(input_tokens, dtype=dtype)
+    def forward(self, input_tokens, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None):
+        x = self.embeddings(input_tokens, embeds=embeds, dtype=dtype)
         mask = None
         if attention_mask is not None:
             mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, attention_mask.shape[-1], attention_mask.shape[-1])

comfy/text_encoders/llama.py

@@ -241,8 +241,11 @@ def __init__(self, config, device=None, dtype=None, ops=None):
         self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype)
         # self.lm_head = ops.Linear(config.hidden_size, config.vocab_size, bias=False, device=device, dtype=dtype)
 
-    def forward(self, x, attention_mask=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None):
-        x = self.embed_tokens(x, out_dtype=dtype)
+    def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None):
+        if embeds is not None:
+            x = embeds
+        else:
+            x = self.embed_tokens(x, out_dtype=dtype)
 
         if self.normalize_in:
             x *= self.config.hidden_size ** 0.5

comfy/text_encoders/t5.py

@@ -239,8 +239,11 @@ def get_input_embeddings(self):
     def set_input_embeddings(self, embeddings):
         self.shared = embeddings
 
-    def forward(self, input_ids, *args, **kwargs):
-        x = self.shared(input_ids, out_dtype=kwargs.get("dtype", torch.float32))
+    def forward(self, input_ids, attention_mask, embeds=None, num_tokens=None, **kwargs):
+        if input_ids is None:
+            x = embeds
+        else:
+            x = self.shared(input_ids, out_dtype=kwargs.get("dtype", torch.float32))
         if self.dtype not in [torch.float32, torch.float16, torch.bfloat16]:
             x = torch.nan_to_num(x) #Fix for fp8 T5 base
-        return self.encoder(x, *args, **kwargs)
+        return self.encoder(x, attention_mask=attention_mask, **kwargs)