Fix Gemma3 4B training on transformers 5.x (token_type_ids) (#4017)

* Inject token_type_ids for Gemma3 multimodal training on transformers 5.x

In transformers 5.x, create_causal_mask_mapping() raises ValueError when
is_training=True and token_type_ids is None. When doing text-only SFT on
Gemma3 4B (a multimodal model), the dataset_utils detection for
_needs_token_type_ids can miss because:
- The model is wrapped in PeftModel, so type(model).__module__ points to
  peft.peft_model instead of transformers
- The processing_class is a tokenizer (not Gemma3Processor), so the
  fallback MRO check resolves to a module without create_causal_mask_mapping

This adds a fallback in _unsloth_pre_compute_loss that injects
token_type_ids=zeros when:
1. token_type_ids is not already in inputs
2. The inner model config has model_type "gemma3"
3. The model's module has create_causal_mask_mapping (transformers 5.x)
4. The model is in training mode

On transformers 4.x, create_causal_mask_mapping does not exist so this
check is inert.

Depends on: unslothai/unsloth-zoo#488

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: Daniel Hanchen <danielhanchen@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

unsloth/models/_utils.py

@@ -1754,6 +1754,20 @@ def _unsloth_pre_compute_loss(self, model, inputs, *args, **kwargs):
             "Using gradient accumulation will be very slightly less accurate.\n"
             "Read more on gradient accumulation issues here: https://unsloth.ai/blog/gradient"
         )
+    # Gemma3 multimodal models in transformers 5.x require token_type_ids during training.
+    # For text-only SFT, token_type_ids should be all zeros (no image tokens).
+    if "token_type_ids" not in inputs and "input_ids" in inputs:
+        _inner = model
+        for _attr in ("base_model", "model", "model"):
+            _inner = getattr(_inner, _attr, _inner)
+        if getattr(getattr(_inner, "config", None), "model_type", "") in ("gemma3",):
+            import sys as _sys
+
+            _mod = _sys.modules.get(type(_inner).__module__)
+            _has_ccm = _mod is not None and hasattr(_mod, "create_causal_mask_mapping")
+            if _has_ccm and _inner.training:
+                inputs["token_type_ids"] = torch.zeros_like(inputs["input_ids"])
+
     outputs = self._old_compute_loss(model, inputs, *args, **kwargs)
     return outputs