diff --git a/unsloth/__init__.py b/unsloth/__init__.py
index b068d6a5fc..1c168d7a72 100644
--- a/unsloth/__init__.py
+++ b/unsloth/__init__.py
@@ -203,10 +203,14 @@ if DEVICE_TYPE == "cuda":
     del major_version, minor_version
 elif DEVICE_TYPE == "hip":
     SUPPORTS_BFLOAT16 = torch.cuda.is_bf16_supported()
+    def is_bf16_supported():
+        return SUPPORTS_BFLOAT16
 elif DEVICE_TYPE == "xpu":
     # torch.xpu.is_bf16_supported() does not have including_emulation
     # set SUPPORTS_BFLOAT16 as torch.xpu.is_bf16_supported()
     SUPPORTS_BFLOAT16 = torch.xpu.is_bf16_supported()
+    def is_bf16_supported():
+        return SUPPORTS_BFLOAT16
 
 # For Gradio HF Spaces?
 # if "SPACE_AUTHOR_NAME" not in os.environ and "SPACE_REPO_NAME" not in os.environ:
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index 3e6dc8ac5f..15a7dd5515 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -782,6 +782,11 @@ class FastBaseModel:
                 # attn_implementation   = attn_implementation,
                 **kwargs,
             )
+            try:
+                from unsloth_zoo.temporary_patches.misc import patch_deepseek_ocr_masked_scatter
+                patch_deepseek_ocr_masked_scatter()
+            except Exception:
+                pass
             if hasattr(model, "generate"):
                 model.fast_generate = make_fast_generate_wrapper(model.generate)
                 model.fast_generate_batches = error_out_no_vllm
diff --git a/unsloth/save.py b/unsloth/save.py
index fc3b7b8771..11e70a5dee 100644
--- a/unsloth/save.py
+++ b/unsloth/save.py
@@ -1996,7 +1996,15 @@ def unsloth_save_pretrained_gguf(
             is_gpt_oss = is_gpt_oss,  # Pass gpt_oss Flag
         )
     except Exception as e:
-        if IS_KAGGLE_ENVIRONMENT:
+        if os.environ.get("UNSLOTH_GGUF_OFFLINE", "0") == "1":
+            print(
+                "Unsloth: GGUF conversion skipped due to offline mode. "
+                f"Reason: {e}"
+            )
+            all_file_locations = []
+            want_full_precision = None
+            is_vlm_update = False
+        elif IS_KAGGLE_ENVIRONMENT:
             raise RuntimeError(
                 f"Unsloth: GGUF conversion failed in Kaggle environment.\n"
                 f"This is likely due to the 20GB disk space limit.\n"
@@ -2010,6 +2018,17 @@ def unsloth_save_pretrained_gguf(
     gguf_directory = f"{save_directory}_gguf"
     modelfile_location = None
     ollama_success = False
+    if not all_file_locations:
+        # Offline or failed GGUF conversion: return early to avoid index errors
+        return {
+            "save_directory": save_directory,
+            "gguf_directory": gguf_directory,
+            "gguf_files": all_file_locations,
+            "modelfile_location": modelfile_location,
+            "want_full_precision": want_full_precision,
+            "is_vlm": is_vlm_update,
+            "fix_bos_token": fix_bos_token,
+        }
     if all_file_locations:
         try:
             if is_vlm_update: