diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index 1a8fff9ada..c13b2286f3 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -986,8 +986,8 @@ def _unsloth_pre_compute_loss(self, model, inputs, *args, **kwargs):
             "Read more on gradient accumulation issues here: https://unsloth.ai/blog/gradient"
         )
     pass
-    with torch.autocast(device_type = "cuda", dtype = torch.float32):
-        outputs = self._old_compute_loss(model, inputs, *args, **kwargs)
+    # with torch.autocast(device_type = "cuda", dtype = torch.float32):
+    outputs = self._old_compute_loss(model, inputs, *args, **kwargs)
     return outputs
 pass
 
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index 73497e70a7..26e9edffd3 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -65,6 +65,9 @@ __all__ = [
     "FastBaseModel",
 ]
 
+global FORCE_FLOAT32
+FORCE_FLOAT32 = ["gemma3"]
+
 
 def unsloth_base_fast_generate(
     self,
@@ -178,6 +181,14 @@ class FastBaseModel:
 
         assert(dtype == torch.float16 or dtype == torch.bfloat16 or dtype == torch.float32)
 
+        global FORCE_FLOAT32
+        os.environ["UNSLOTH_FORCE_FLOAT32"] = "0"
+        for disable_name in FORCE_FLOAT32:
+            if disable_name.lower() == model_type_arch.lower() and dtype == torch.float16:
+                print(f"Unsloth: Using float16 precision for {model_type_arch} won't work! Using float32.")
+                os.environ["UNSLOTH_FORCE_FLOAT32"] = "1"
+                break
+
         bnb_config = None
         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.")