P2 of qwen edit model. (Comfy-Org#9412)

* P2 of qwen edit model.

* Typo.

* Fix normal qwen.

* Fix.

* Make the TextEncodeQwenImageEdit also set the ref latent.

If you don't want it to set the ref latent and want to use the
ReferenceLatent node with your custom latent instead just disconnect the
VAE.

Author: comfyanonymous

comfy/clip_model.py

@@ -97,7 +97,7 @@ 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=None, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=torch.float32):
+    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, embeds_info=[]):
         if embeds is not None:
             x = embeds + comfy.ops.cast_to(self.embeddings.position_embedding.weight, dtype=dtype, device=embeds.device)
         else:

comfy/model_base.py

@@ -1325,6 +1325,7 @@ def extra_conds_shapes(self, **kwargs):
 class QwenImage(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLUX, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.qwen_image.model.QwenImageTransformer2DModel)
+        self.memory_usage_factor_conds = ("ref_latents",)
 
     def extra_conds(self, **kwargs):
         out = super().extra_conds(**kwargs)
@@ -1342,3 +1343,10 @@ def extra_conds(self, **kwargs):
             if ref_latents_method is not None:
                 out['ref_latents_method'] = comfy.conds.CONDConstant(ref_latents_method)
         return out
+
+    def extra_conds_shapes(self, **kwargs):
+        out = {}
+        ref_latents = kwargs.get("reference_latents", None)
+        if ref_latents is not None:
+            out['ref_latents'] = list([1, 16, sum(map(lambda a: math.prod(a.size()), ref_latents)) // 16])
+        return out

comfy/sd1_clip.py

@@ -204,17 +204,19 @@ def process_tokens(self, tokens, device):
             tokens_embed = self.transformer.get_input_embeddings()(tokens_embed, out_dtype=torch.float32)
             index = 0
             pad_extra = 0
+            embeds_info = []
             for o in other_embeds:
                 emb = o[1]
                 if torch.is_tensor(emb):
                     emb = {"type": "embedding", "data": emb}
 
+                extra = None
                 emb_type = emb.get("type", None)
                 if emb_type == "embedding":
                     emb = emb.get("data", None)
                 else:
                     if hasattr(self.transformer, "preprocess_embed"):
-                        emb = self.transformer.preprocess_embed(emb, device=device)
+                        emb, extra = self.transformer.preprocess_embed(emb, device=device)
                     else:
                         emb = None
 
@@ -229,6 +231,7 @@ def process_tokens(self, tokens, device):
                     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
+                    embeds_info.append({"type": emb_type, "index": ind, "size": emb_shape, "extra": extra})
                 else:
                     index += -1
                     pad_extra += emb_shape
@@ -243,11 +246,11 @@ def process_tokens(self, tokens, device):
             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
+        return torch.cat(embeds_out), torch.tensor(attention_masks, device=device, dtype=torch.long), num_tokens, embeds_info
 
     def forward(self, tokens):
         device = self.transformer.get_input_embeddings().weight.device
-        embeds, attention_mask, num_tokens = self.process_tokens(tokens, device)
+        embeds, attention_mask, num_tokens, embeds_info = self.process_tokens(tokens, device)
 
         attention_mask_model = None
         if self.enable_attention_masks:
@@ -258,7 +261,7 @@ def forward(self, tokens):
         else:
             intermediate_output = self.layer_idx
 
-        outputs = self.transformer(None, attention_mask_model, embeds=embeds, num_tokens=num_tokens, intermediate_output=intermediate_output, final_layer_norm_intermediate=self.layer_norm_hidden_state, dtype=torch.float32)
+        outputs = self.transformer(None, attention_mask_model, embeds=embeds, num_tokens=num_tokens, intermediate_output=intermediate_output, final_layer_norm_intermediate=self.layer_norm_hidden_state, dtype=torch.float32, embeds_info=embeds_info)
 
         if self.layer == "last":
             z = outputs[0].float()

comfy/text_encoders/bert.py

@@ -116,7 +116,7 @@ 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, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None):
+    def forward(self, input_tokens, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, embeds_info=[]):
         x = self.embeddings(input_tokens, embeds=embeds, dtype=dtype)
         mask = None
         if attention_mask is not None:

comfy/text_encoders/llama.py

@@ -2,12 +2,14 @@
 import torch.nn as nn
 from dataclasses import dataclass
 from typing import Optional, Any
+import math
 
 from comfy.ldm.modules.attention import optimized_attention_for_device
 import comfy.model_management
 import comfy.ldm.common_dit
 
 import comfy.model_management
+from . import qwen_vl
 
 @dataclass
 class Llama2Config:
@@ -100,12 +102,10 @@ def rotate_half(x):
     return torch.cat((-x2, x1), dim=-1)
 
 
-def precompute_freqs_cis(head_dim, seq_len, theta, device=None):
+def precompute_freqs_cis(head_dim, position_ids, theta, device=None):
     theta_numerator = torch.arange(0, head_dim, 2, device=device).float()
     inv_freq = 1.0 / (theta ** (theta_numerator / head_dim))
 
-    position_ids = torch.arange(0, seq_len, device=device).unsqueeze(0)
-
     inv_freq_expanded = inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
     position_ids_expanded = position_ids[:, None, :].float()
     freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
@@ -277,7 +277,7 @@ 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, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None):
+    def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, position_ids=None, embeds_info=[]):
         if embeds is not None:
             x = embeds
         else:
@@ -286,8 +286,11 @@ def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermed
         if self.normalize_in:
             x *= self.config.hidden_size ** 0.5
 
+        if position_ids is None:
+            position_ids = torch.arange(0, x.shape[1], device=x.device).unsqueeze(0)
+
         freqs_cis = precompute_freqs_cis(self.config.head_dim,
-                                         x.shape[1],
+                                         position_ids,
                                          self.config.rope_theta,
                                          device=x.device)
 
@@ -372,8 +375,38 @@ def __init__(self, config_dict, dtype, device, operations):
         self.num_layers = config.num_hidden_layers
 
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.visual = qwen_vl.Qwen2VLVisionTransformer(hidden_size=1280, output_hidden_size=config.hidden_size, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
+    def preprocess_embed(self, embed, device):
+        if embed["type"] == "image":
+            image, grid = qwen_vl.process_qwen2vl_images(embed["data"])
+            return self.visual(image.to(device, dtype=torch.float32), grid), grid
+        return None, None
+
+    def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, embeds_info=[]):
+        grid = None
+        for e in embeds_info:
+            if e.get("type") == "image":
+                grid = e.get("extra", None)
+                position_ids = torch.zeros((3, embeds.shape[1]), device=embeds.device)
+                start = e.get("index")
+                position_ids[:, :start] = torch.arange(0, start, device=embeds.device)
+                end = e.get("size") + start
+                len_max = int(grid.max()) // 2
+                start_next = len_max + start
+                position_ids[:, end:] = torch.arange(start_next, start_next + (embeds.shape[1] - end), device=embeds.device)
+                position_ids[0, start:end] = start
+                max_d = int(grid[0][1]) // 2
+                position_ids[1, start:end] = torch.arange(start, start + max_d, device=embeds.device).unsqueeze(1).repeat(1, math.ceil((end - start) / max_d)).flatten(0)[:end - start]
+                max_d = int(grid[0][2]) // 2
+                position_ids[2, start:end] = torch.arange(start, start + max_d, device=embeds.device).unsqueeze(0).repeat(math.ceil((end - start) / max_d), 1).flatten(0)[:end - start]
+
+        if grid is None:
+            position_ids = None
+
+        return super().forward(x, attention_mask=attention_mask, embeds=embeds, num_tokens=num_tokens, intermediate_output=intermediate_output, final_layer_norm_intermediate=final_layer_norm_intermediate, dtype=dtype, position_ids=position_ids)
+
 class Gemma2_2B(BaseLlama, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()

comfy/text_encoders/qwen_image.py

@@ -15,13 +15,27 @@ class QwenImageTokenizer(sd1_clip.SD1Tokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="qwen25_7b", tokenizer=Qwen25_7BVLITokenizer)
         self.llama_template = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n"
+        self.llama_template_images = "<|im_start|>system\nDescribe the key features of the input image \\(color, shape, size, texture, objects, background\\), then explain how the user's text instruction should alter or modify the image. Generate a new image that meets the user's requirements while maintaining consistency with the original input where appropriate.<|im_end|>\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>{}<|im_end|>\n<|im_start|>assistant\n"
 
-    def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None,**kwargs):
+    def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, images=[], **kwargs):
         if llama_template is None:
-            llama_text = self.llama_template.format(text)
+            if len(images) > 0:
+                llama_text = self.llama_template_images.format(text)
+            else:
+                llama_text = self.llama_template.format(text)
         else:
             llama_text = llama_template.format(text)
-        return super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, **kwargs)
+        tokens = super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, **kwargs)
+        key_name = next(iter(tokens))
+        embed_count = 0
+        qwen_tokens = tokens[key_name]
+        for r in qwen_tokens:
+            for i in range(len(r)):
+                if r[i][0] == 151655:
+                    if len(images) > embed_count:
+                        r[i] = ({"type": "image", "data": images[embed_count], "original_type": "image"},) + r[i][1:]
+                        embed_count += 1
+        return tokens
 
 
 class Qwen25_7BVLIModel(sd1_clip.SDClipModel):