diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index 70dd896514..3865857dfe 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -557,7 +557,7 @@ class FastModel(FastBaseModel):
             raise RuntimeError("Unsloth: Cohere's Command model only works on transformers >= 4.50.0." + NIGHTLY)
         elif "csm-1b" in lowered_model_name:
             os.environ["UNSLOTH_DISABLE_STATIC_GENERATION"] = "1" # Sesame fails
-            os.environ["UNSLOTH_FORCE_CUSTOM_DTYPE"] = "torch.float16;if name.endswith(('_proj', 'fc1', 'fc2', 'codebook', 'head')): module.to(torch.float16)"
+            os.environ["UNSLOTH_FORCE_CUSTOM_DTYPE"] = "all;torch.float32;torch.float16;if name.endswith(('_proj', 'fc1', 'fc2', 'codebook', 'head')): module.to(torch.float16)"
         elif 'granite-4' in lowered_model_name:
             # granite-4 rms norms are stored as 16 bit, but we upcast
             os.environ["UNSLOTH_UPCAST_LAYERNORM"] = "1"
@@ -566,6 +566,7 @@ class FastModel(FastBaseModel):
             raise RuntimeError("Unsloth: OLMo-2 only works on transformers >= 4.50.0." + NIGHTLY)
         elif "gemma-3n" in lowered_model_name:
             os.environ["UNSLOTH_DISABLE_STATIC_GENERATION"] = "1"
+            s.environ["UNSLOTH_FORCE_CUSTOM_DTYPE"] = "float16;torch.float16;torch.float16;if name.endswith(('.conv')): module.to(torch.float32)"
             if transformers_version < Version("4.53.0"):
                 raise RuntimeError("Unsloth: Gemma 3N only works on transformers >= 4.53.0" + LATEST)
         else:
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index 5e8c4aec38..be9fcd31eb 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -352,11 +352,20 @@ class FastBaseModel:
         correct_dtype = None
         if os.environ.get("UNSLOTH_FORCE_CUSTOM_DTYPE", "") != "":
             custom_datatype = os.environ["UNSLOTH_FORCE_CUSTOM_DTYPE"]
-            assert custom_datatype.count(";") == 1
-            bnb_compute_dtype, custom_datatype = custom_datatype.split(";", 1)
-            dtype = torch.float32
-            bnb_compute_dtype = eval(bnb_compute_dtype)
-            correct_dtype = bnb_compute_dtype
+            assert custom_datatype.count(";") == 3
+            checker, _dtype, _bnb_compute_dtype, _custom_datatype = custom_datatype.split(";", 3)
+
+            # Allow custom dtypes on all runs
+            allow_all_runs = (checker == "all")
+            # Allow only on float16 datatypes
+            allow_float16_runs = (checker == "float16" and dtype == torch.float16)
+
+            if allow_all_runs or allow_float16_runs:
+                dtype = eval(_dtype)
+                bnb_compute_dtype = eval(_bnb_compute_dtype)
+                correct_dtype = bnb_compute_dtype
+                custom_datatype = _custom_datatype
+            pass
         pass
 
         # Stop SDPA for some archs like Pixtral / Mistral3