diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index fc91178d88..a42b7f38a7 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -109,6 +109,13 @@ FORCE_FLOAT32 = [
     "qwen3_5",  # Qwen3.5 GDN layers produce NaN grad norms in float16 training
 ]
 
+# Models that must use bfloat16 instead of float16.
+# torch.compile backward graphs overflow fp16 intermediates for these models.
+FORCE_BFLOAT16 = [
+    "gemma4,",  # Add comma bc gemma4 will match gemma4_text
+    "gemma4text",  # Gemma4TextModel (standalone text-only Gemma4)
+]
+
 global DISABLE_COMPILE_MODEL_NAMES
 # Must be alphabetically sorted for each entry
 
@@ -1373,6 +1380,24 @@ class FastModel(FastBaseModel):
                 os.environ["UNSLOTH_FORCE_FLOAT32"] = "1"
                 dtype = torch.bfloat16  # Change to bfloat16 loading
                 break
+        # Switch fp16 to bf16 for models whose torch.compile backward overflows fp16
+        global FORCE_BFLOAT16
+        for disable_name in FORCE_BFLOAT16:
+            if (
+                (
+                    disable_name.lower()
+                    == model_type_arch.lower().replace("-", "").replace("_", "")
+                    or disable_name.lower() in model_types_all
+                )
+                and dtype == torch.float16
+                and SUPPORTS_BFLOAT16
+            ):
+                logger.warning_once(
+                    f"Unsloth: {model_type_arch} does not support float16 training. "
+                    f"Switching to bfloat16."
+                )
+                dtype = torch.bfloat16
+                break
         # Apply gradient checkpointing with smart heuristics
         use_gradient_checkpointing = apply_unsloth_gradient_checkpointing(
             use_gradient_checkpointing, max_seq_length, dtype