Studio: thread vision_image_size into DeepSeek OCR + writable MLX ndarray

- trainer.py: DeepSeek OCR collator now honors the new vision_image_size
  setting as image_size. Falls back to 640 when null. base_size stays at
  1024 and crop_mode stays True so the Gundam preset's dynamic cropping
  of large documents keeps working.
- worker.py: _resize_mlx_vlm_image returns np.array(image, copy=True)
  instead of np.asarray(image). The PIL view from np.asarray is not
  writable, which makes HF VLM processors emit "The given NumPy array
  is not writable, and PyTorch does not support non-writable tensors..."
  when they call torch.from_numpy. copy=True keeps the same shape and
  dtype but produces a writable buffer.

studio/backend/core/training/trainer.py

@@ -3057,10 +3057,23 @@ class UnslothTrainer:
 
                     logger.info("Configuring DeepSeek OCR data collator...\n")
                     FastVisionModel.for_training(self.model)
+                    # Honor user-selected Image Size. Keep base_size + crop_mode
+                    # at the Gundam preset (1024 / True) so dynamic cropping of
+                    # large documents still works.
+                    vision_image_size = training_args.get("vision_image_size")
+                    deepseek_image_size = (
+                        640 if vision_image_size is None
+                        else int(vision_image_size)
+                    )
+                    if vision_image_size is not None:
+                        logger.info(
+                            f"DeepSeek OCR image resize: "
+                            f"{deepseek_image_size} (per-crop tile size)\n"
+                        )
                     data_collator = DeepSeekOCRDataCollator(
                         tokenizer = self.tokenizer,
                         model = self.model,
-                        image_size = 640,
+                        image_size = deepseek_image_size,
                         base_size = 1024,
                         crop_mode = True,
                         train_on_responses_only = training_args.get(

studio/backend/core/training/worker.py

@@ -984,9 +984,10 @@ def _resize_mlx_vlm_image(image, resize):
     if new_size != image.size:
         resampling = getattr(Image, "Resampling", Image).LANCZOS
         image = image.resize(new_size, resampling)
-    # mlx-vlm's internal collator square-resizes PIL images. Return an ndarray
-    # so Studio's max-dimension resize is the final resize, like trainer.py.
-    return np.asarray(image)
+    # mlx-vlm's internal collator square-resizes PIL images. Return a writable
+    # ndarray so Studio's max-dimension resize is the final one (like
+    # trainer.py) and HF processors don't warn on non-writable views.
+    return np.array(image, copy = True)
 
 
 def _resize_mlx_vlm_images(value, resize):