P2 of qwen edit model. (#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.

comfy/text_encoders/llama.py

@@ -2,12 +2,14 @@ import torch
 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 @@ class Llama2_(nn.Module):
         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 @@ class Llama2_(nn.Module):
         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 @@ class Qwen25_7BVLI(BaseLlama, torch.nn.Module):
         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__()