diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index b8681a5cc9..cdd5f97b98 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -1134,6 +1134,7 @@ pass
 def unsloth_compile_transformers(
     dtype,
     model_name,
+    model_types,
     token                   = None,
     revision                = None,
     trust_remote_code       = False,
@@ -1171,30 +1172,8 @@ def unsloth_compile_transformers(
         )
         return
     pass
-
-    model_types = get_transformers_model_type(
-        model_name        = model_name,
-        token             = token,
-        revision          = revision,
-        trust_remote_code = trust_remote_code,
-    )
-    model_types = ["siglip"] + model_types
-
     if disable: return
 
-    # Set forced float32 env flag
-    os.environ["UNSLOTH_FORCE_FLOAT32"] = "1"
-    do_forced_float32 = False
-    model_type_arch = model_types[1]
-    for disable_name in FORCE_FLOAT32:
-        if (disable_name.lower() == model_type_arch.lower() or \
-            disable_name.lower() in model_name.lower()) and \
-            dtype == torch.float16:
-            os.environ["UNSLOTH_FORCE_FLOAT32"] = "1"
-            do_forced_float32 = True
-            break
-    pass
-
     for model_type in model_types:
         _unsloth_compile_transformers(
             model_type,
diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index fbda4916e1..4d2fc1a300 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -17,6 +17,7 @@ from ._utils import (
     HAS_FLASH_ATTENTION,
     HAS_FLASH_ATTENTION_SOFTCAPPING,
     USE_MODELSCOPE,
+    get_transformers_model_type,
 )
 from .granite import FastGraniteModel
 from .llama   import FastLlamaModel, logger
@@ -462,17 +463,15 @@ class FastModel(FastBaseModel):
         if token is None: token = get_token()
 
         SUPPORTS_BFLOAT16 = is_bfloat16_supported()
-        # if dtype is None:
-        #     dtype = torch.float16 if not SUPPORTS_BFLOAT16 else torch.bfloat16
-        # elif dtype == torch.bfloat16 and not SUPPORTS_BFLOAT16:
-        #     logger.warning_once("Device does not support bfloat16. Will change to float16.")
-        #     dtype = torch.float16
+        if dtype is None:
+            dtype = torch.float16 if not SUPPORTS_BFLOAT16 else torch.bfloat16
+        elif dtype == torch.bfloat16 and not SUPPORTS_BFLOAT16:
+            logger.warning_once("Device does not support bfloat16. Will change to float16.")
+            dtype = torch.float16
         assert(dtype in (torch.float16, torch.bfloat16, torch.float32))
 
         patch_compiled_autograd()
         patch_compiling_bitsandbytes()
-        if use_gradient_checkpointing == "unsloth":
-            patch_unsloth_smart_gradient_checkpointing(dtype = dtype)
 
         if full_finetuning and (load_in_4bit or load_in_8bit):
             print("Unsloth: You selected full finetuning support, but 4bit / 8bit is enabled - disabling LoRA / QLoRA.")
@@ -618,11 +617,38 @@ class FastModel(FastBaseModel):
         else:
             redirector = contextlib.redirect_stdout(open(os.devnull, "w"))
 
+        # Get model types like Gemma3 etc
+        model_types = get_transformers_model_type(
+            model_name        = model_name,
+            token             = token,
+            revision          = revision,
+            trust_remote_code = trust_remote_code,
+        )
+        model_types = ["siglip"] + model_types
+
+        # Set forced float32 env flag
+        os.environ["UNSLOTH_FORCE_FLOAT32"] = "0"
+        do_forced_float32 = False
+        model_type_arch = model_types[1]
+        for disable_name in FORCE_FLOAT32:
+            if (disable_name.lower() == model_type_arch.lower() or \
+                disable_name.lower() in model_name.lower()) and \
+                ((dtype == torch.float16) or not SUPPORTS_BFLOAT16):
+                os.environ["UNSLOTH_FORCE_FLOAT32"] = "1"
+                dtype = torch.bfloat16 # Change to bfloat16 loading
+                break
+        pass
+        # Patch gradient checkpointing
+        if use_gradient_checkpointing == "unsloth":
+            patch_unsloth_smart_gradient_checkpointing(dtype = dtype)
+
         with redirector:
             patch_loss_functions(torch_compile = False)
             model_types = unsloth_compile_transformers(
                 dtype                   = dtype,
                 model_name              = model_name,
+                model_types             = model_types,
+                token                   = token,
                 sdpa_dynamic_mask       = True,
                 sdpa_bool_masks         = True,
                 sdpa_gqa_replace        = True,
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index bb6693e763..d79e9a829b 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -202,12 +202,15 @@ class FastBaseModel:
 
         get_statistics() # For debugging - we use a download counter to see if environments are not breaking 
 
-        # if dtype is None:
-        #     dtype = torch.float16 if not SUPPORTS_BFLOAT16 else torch.bfloat16
-        # elif dtype == torch.bfloat16 and not SUPPORTS_BFLOAT16:
-        #     logger.warning_once("Device does not support bfloat16. Will change to float16.")
-        #     dtype = torch.float16
-
+        if dtype is None:
+            dtype = torch.float16 if not SUPPORTS_BFLOAT16 else torch.bfloat16
+        elif dtype == torch.bfloat16 and not SUPPORTS_BFLOAT16:
+            logger.warning_once("Device does not support bfloat16. Will change to float16.")
+            dtype = torch.float16
+        # Check forced float32
+        if os.environ.get("UNSLOTH_FORCE_FLOAT32", "0") == "1":
+            if dtype == torch.float16: dtype = torch.bfloat16
+        pass
         assert(dtype in (torch.float16, torch.bfloat16, torch.float32))
 
         bnb_compute_dtype = dtype