Merge pull request #826 from mi804/qwen-image-edit-lowvram

fix qwen-image-edit-lowvram

Author: Artiprocher

diffsynth/pipelines/qwen_image.py

@@ -103,7 +103,7 @@ def enable_vram_management(self, num_persistent_param_in_dit=None, vram_limit=No
         vram_limit = vram_limit - vram_buffer
 
         if self.text_encoder is not None:
-            from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import Qwen2_5_VLRotaryEmbedding, Qwen2RMSNorm
+            from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import Qwen2_5_VLRotaryEmbedding, Qwen2RMSNorm, Qwen2_5_VisionPatchEmbed, Qwen2_5_VisionRotaryEmbedding
             dtype = next(iter(self.text_encoder.parameters())).dtype
             enable_vram_management(
                 self.text_encoder,
@@ -112,6 +112,8 @@ def enable_vram_management(self, num_persistent_param_in_dit=None, vram_limit=No
                     torch.nn.Embedding: AutoWrappedModule,
                     Qwen2_5_VLRotaryEmbedding: AutoWrappedModule,
                     Qwen2RMSNorm: AutoWrappedModule,
+                    Qwen2_5_VisionPatchEmbed: AutoWrappedModule,
+                    Qwen2_5_VisionRotaryEmbedding: AutoWrappedModule,
                 },
                 module_config = dict(
                     offload_dtype=dtype,