From 5d7d882ce6184c990246d41eb0e512a8fb7805a7 Mon Sep 17 00:00:00 2001
From: Daniel Han <danielhanchen@gmail.com>
Date: Wed, 1 Apr 2026 06:05:37 -0700
Subject: [PATCH] Fix save_pretrained_merged for full-finetuned models (#4755)

* Fix save_pretrained_merged for full-finetuned models

save_pretrained_merged and push_to_hub_merged silently do nothing when
the model is not a PeftModel (i.e. full finetuning without LoRA).
merge_and_overwrite_lora returns None immediately for non-PeftModel,
and unsloth_generic_save does not check the return value.

Add a non-PeftModel branch in unsloth_generic_save that falls back to
model.save_pretrained / model.push_to_hub. When save_method contains
"16bit", cast weights to bfloat16 (or float16) via a state_dict copy
to honor the user's intent without mutating the live model.

The existing PeftModel (LoRA) code path is unchanged.

* Forward create_pr and revision to tokenizer.push_to_hub

The tokenizer push_to_hub call was missing create_pr and revision,
which could cause the tokenizer to push to the wrong branch or
bypass PR creation when the model push uses them.

* Honor merged_16bit dtype contract for full-finetuned models

Cast state_dict to bfloat16/float16 when save_method contains "16bit"
to match the documented behavior of save_pretrained_merged. Also pass
state_dict and save kwargs consistently to both save_pretrained and
push_to_hub paths.

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

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

* Address review feedback for PR #4755

- Simplify PeftModel isinstance check (PeftModelForCausalLM inherits
  from PeftModel)
- Add is_main_process guard for distributed training
- Forward variant to save_pretrained
- Set tokenizer padding_side to "left" before saving (matches other
  save paths)

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

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

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
---
 unsloth/save.py | 86 +++++++++++++++++++++++++++++++++++++++++--------
 1 file changed, 73 insertions(+), 13 deletions(-)

diff --git a/unsloth/save.py b/unsloth/save.py
index 1759d86fb1..178178980d 100644
--- a/unsloth/save.py
+++ b/unsloth/save.py
@@ -2777,19 +2777,79 @@ def unsloth_generic_save(
     elif save_method == "merged_4bit_forced":
         save_method = "merged_4bit"
 
-    merge_and_overwrite_lora(
-        get_model_name,
-        model = model,
-        tokenizer = tokenizer,
-        save_directory = save_directory,
-        push_to_hub = push_to_hub,
-        private = private,
-        token = token,
-        save_method = save_method,
-        output_dtype = None,
-        low_disk_space_usage = True,
-        use_temp_file = False,
-    )
+    # Full-finetuned models (no LoRA) cannot use merge_and_overwrite_lora
+    # since there are no adapters to merge. Fall back to save_pretrained.
+    # This mirrors the non-PeftModel handling in save_pretrained_torchao
+    # and the GGUF save path.
+    _is_peft = isinstance(model, PeftModel)
+    if not _is_peft:
+        if not is_main_process:
+            return
+
+        # Honor merged_16bit by casting to the target dtype if needed
+        _save_kwargs = dict(
+            safe_serialization = safe_serialization,
+            max_shard_size = max_shard_size,
+            variant = variant,
+        )
+        if "16bit" in save_method:
+            _target_dtype = (
+                torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
+            )
+            _save_kwargs["state_dict"] = {
+                k: v.to(dtype = _target_dtype) if v.is_floating_point() else v
+                for k, v in model.state_dict().items()
+            }
+
+        if push_to_hub:
+            print(f"Unsloth: Pushing full fine-tuned model to '{save_directory}' ...")
+            model.push_to_hub(
+                repo_id = save_directory,
+                token = token,
+                private = private,
+                commit_message = commit_message,
+                create_pr = create_pr,
+                revision = revision,
+                commit_description = commit_description,
+                tags = tags,
+                **_save_kwargs,
+            )
+            if tokenizer is not None:
+                old_padding_side = tokenizer.padding_side
+                tokenizer.padding_side = "left"
+                tokenizer.push_to_hub(
+                    save_directory,
+                    token = token,
+                    private = private,
+                    commit_message = commit_message,
+                    create_pr = create_pr,
+                    revision = revision,
+                )
+                tokenizer.padding_side = old_padding_side
+        else:
+            print(f"Unsloth: Saving full fine-tuned model to '{save_directory}' ...")
+            model.save_pretrained(save_directory, **_save_kwargs)
+            if tokenizer is not None:
+                old_padding_side = tokenizer.padding_side
+                tokenizer.padding_side = "left"
+                tokenizer.save_pretrained(save_directory)
+                tokenizer.padding_side = old_padding_side
+
+        print(f"Unsloth: Model saved successfully to '{save_directory}'")
+    else:
+        merge_and_overwrite_lora(
+            get_model_name,
+            model = model,
+            tokenizer = tokenizer,
+            save_directory = save_directory,
+            push_to_hub = push_to_hub,
+            private = private,
+            token = token,
+            save_method = save_method,
+            output_dtype = None,
+            low_disk_space_usage = True,
+            use_temp_file = False,
+        )
 
     if push_to_hub and datasets:
         try: