[Bug] save_pretrained_merge does not work as intended for finetuned Gemma-3-4b model.

[jasonkhadka]

I am following unsloth colab notebook for pretriaining. Followed by save_pretrained_merge and reloading the merged with FastVisionModel.from_pretrained(...). The reloaded model has very different output from the non-merged finetuned model.

https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/Gemma3_(4B)-Vision.ipynb#scrollTo=QmUBVEnvCDJv

model, processor = FastVisionModel.from_pretrained(
    # "unsloth/gemma-3-4b-it",
    "unsloth/gemma-3-4b-pt",
    load_in_4bit = False, # Use 4bit to reduce memory use. False for 16bit LoRA.
    use_gradient_checkpointing = "unsloth", # True or "unsloth" for long context
)
# train the model on latex dataset from unsloth notebook
# wrap model to Lora: I am not doing vision FT
model = FastVisionModel.get_peft_model(
    model,
    finetune_vision_layers     = False, # False if not finetuning vision layers
    finetune_language_layers   = True, # False if not finetuning language layers
    finetune_attention_modules = True, # False if not finetuning attention layers
    finetune_mlp_modules       = True, # False if not finetuning MLP layers
    r = 16,                           # The larger, the higher the accuracy, but might overfit
    lora_alpha = 16,                  # Recommended alpha == r at least
    lora_dropout = 0,
    bias = "none",
    random_state = 3407,
    use_rslora = False,               # We support rank stabilized LoRA
    loftq_config = None,               # And LoftQ
    target_modules = "all-linear",    # Optional now! Can specify a list if needed
    modules_to_save=[
        "lm_head",
        "embed_tokens",
    ],
)

FastVisionModel.for_training(model) # Enable for training!
trainer = SFTTrainer(...) 
trainer.train()

# infer and save the model
FastVisionModel.for_inference(model)  # Enable for inference!
out1 = model(**inputs)

# save the merged model, model should be merged and saved in fp16
model.save_pretrained_merged("/tmp/unsloth_finetune", processor,)

# reload the merged model
merged_model, merged_processor = FastVisionModel.from_pretrained(
    "/tmp/unsloth_finetune",
    load_in_4bit = False, # Use 4bit to reduce memory use. False for 16bit LoRA.
    use_gradient_checkpointing = "unsloth", # True or "unsloth" for long context
)

FastVisionModel.for_inference(merged_model)  # Enable for inference!
merged_out = merged_model(**inputs)

print((out1.logits - merged_out.logits).mean())
>>> tensor(-0.1275, device='cuda:0', grad_fn=<MeanBackward0>)

print((out1.logits - merged_out.logits).max())
>> tensor(25.7500, device='cuda:0', grad_fn=<MaxBackward1>)

Unsloth versions:

unsloth                  2025.7.8
unsloth_zoo              2025.7.10

[jasonkhadka]

This issue should have been fixed and was discussed before, but apparently not.
@rolandtannous
#1670 (comment)

[rolandtannous]

we've discussed this at length in a separate discord thread. will turn this into a discussion for now.