fix: support qwen-image finetune (t2i) (#1700)

Signed-off-by: Harsha Pasham <pashamharsha018@gmail.com>

Author: harshareddy832

tests/unit_tests/flow_matching/test_qwen_image_adapter.py

@@ -20,8 +20,9 @@
 import torch
 
 from nemo_automodel.components.flow_matching.adapters.base import FlowMatchingContext
-from nemo_automodel.components.flow_matching.adapters.qwen_image import QwenImageAdapter
-
+from nemo_automodel.components.flow_matching.adapters.qwen_image import (
+    QwenImageAdapter,
+)
 
 # =============================================================================
 # TestQwenImageAdapterPackLatents
@@ -182,7 +183,8 @@ def test_2d_text_unsqueeze(self):
         noisy = torch.randn(1, 16, 8, 8)
         batch = {"text_embeddings": torch.randn(77, 2048)}  # 2D
         ctx = self._make_context(
-            noisy, batch,
+            noisy,
+            batch,
             timesteps=torch.tensor([500.0]),
             sigma=torch.tensor([0.5]),
         )

tools/diffusion/processors/qwen_image.py

@@ -66,7 +66,7 @@ def load_models(self, model_name: str, device: str) -> Dict[str, Any]:
                 - tokenizer: Qwen2 tokenizer
                 - text_encoder: Qwen2 text encoder
         """
-        from diffusers import AutoencoderKL, QwenImagePipeline
+        from diffusers import QwenImagePipeline
 
         logger.info("[Qwen-Image] Loading models from %s...", model_name)