InvokeAI 6.10.0rc1__py3-none-any.whl → 6.11.0__py3-none-any.whl

invokeai/backend/quantization/gguf/ggml_tensor.py

@@ -17,21 +17,32 @@ def dequantize_and_run(func, args, kwargs):
     Also casts other floating point tensors to match the compute_dtype of GGMLTensors
     to avoid dtype mismatches in matrix operations.
     """
-    # Find the compute_dtype from any GGMLTensor in the args
+    # Find the compute_dtype and target_device from any GGMLTensor in the args
     compute_dtype = None
+    target_device = None
     for a in args:
         if hasattr(a, "compute_dtype"):
             compute_dtype = a.compute_dtype
+        if isinstance(a, torch.Tensor) and target_device is None:
+            target_device = a.device
+        if compute_dtype is not None and target_device is not None:
             break
-    if compute_dtype is None:
+    if compute_dtype is None or target_device is None:
         for v in kwargs.values():
-            if hasattr(v, "compute_dtype"):
+            if hasattr(v, "compute_dtype") and compute_dtype is None:
                 compute_dtype = v.compute_dtype
+            if isinstance(v, torch.Tensor) and target_device is None:
+                target_device = v.device
+            if compute_dtype is not None and target_device is not None:
                 break
 
     def process_tensor(t):
         if hasattr(t, "get_dequantized_tensor"):
-            return t.get_dequantized_tensor()
+            result = t.get_dequantized_tensor()
+            # Ensure the dequantized tensor is on the target device
+            if target_device is not None and result.device != target_device:
+                result = result.to(target_device)
+            return result
         elif isinstance(t, torch.Tensor) and compute_dtype is not None and t.is_floating_point():
             # Cast other floating point tensors to match the GGUF compute_dtype
             return t.to(compute_dtype)

invokeai/backend/util/vae_working_memory.py

@@ -47,8 +47,6 @@ def estimate_vae_working_memory_sd15_sdxl(
         # If we are running in FP32, then we should account for the likely increase in model size (~250MB).
         working_memory += 250 * 2**20
 
-    print(f"estimate_vae_working_memory_sd15_sdxl: {int(working_memory)}")
-
     return int(working_memory)
 
 

invokeai/backend/z_image/extensions/regional_prompting_extension.py

@@ -66,12 +66,16 @@ class ZImageRegionalPromptingExtension:
     ) -> torch.Tensor | None:
         """Prepare a regional attention mask for Z-Image.
 
-        The mask controls which tokens can attend to each other:
-        - Image tokens within a region attend only to each other
+        This uses an 'unrestricted' image self-attention approach (similar to FLUX):
+        - Image tokens can attend to ALL other image tokens (unrestricted self-attention)
         - Image tokens attend only to their corresponding regional text
         - Text tokens attend only to their corresponding regional image
         - Text tokens attend to themselves
 
+        The unrestricted image self-attention allows the model to maintain global
+        coherence across regions, preventing the generation of separate/disconnected
+        images for each region.
+
         Z-Image sequence order: [img_tokens, txt_tokens]
 
         Args:
@@ -129,12 +133,6 @@ class ZImageRegionalPromptingExtension:
                 # 3. txt attends to corresponding regional img
                 # Reshape mask to (1, img_seq_len) for broadcasting
                 regional_attention_mask[txt_start:txt_end, :img_seq_len] = mask_flat.view(1, img_seq_len)
-
-                # 4. img self-attention within region
-                # mask @ mask.T creates pairwise attention within the masked region
-                regional_attention_mask[:img_seq_len, :img_seq_len] += mask_flat.view(img_seq_len, 1) @ mask_flat.view(
-                    1, img_seq_len
-                )
             else:
                 # Global prompt: allow attention to/from background regions only
                 if background_region_mask is not None:
@@ -152,10 +150,10 @@ class ZImageRegionalPromptingExtension:
                     regional_attention_mask[:img_seq_len, txt_start:txt_end] = 1.0
                     regional_attention_mask[txt_start:txt_end, :img_seq_len] = 1.0
 
-        # Allow background regions to attend to themselves
-        if background_region_mask is not None:
-            bg_mask = background_region_mask.view(img_seq_len, 1)
-            regional_attention_mask[:img_seq_len, :img_seq_len] += bg_mask @ bg_mask.T
+        # 4. Allow unrestricted image self-attention
+        # This is the key difference from the restricted approach - all image tokens
+        # can attend to each other, which helps maintain global coherence across regions
+        regional_attention_mask[:img_seq_len, :img_seq_len] = 1.0
 
         # Convert to boolean mask
         regional_attention_mask = regional_attention_mask > 0.5