README.md

---
base_model: google/gemma-3-12b-it
tags:
- transformers
- torchao
- gemma3
license: apache-2.0
language:
- en
---

# QAT INT4 google/gemma-3-12b-it model

- **Developed by:** pytorch
- **License:** apache-2.0
- **Quantized from Model :** google/gemma-3-12b-it
- **Quantization Method :** QAT INT4
- **Terms of Use**: [Terms][terms]

[gemma-3-12b-it](https://huggingface.co/google/gemma-3-12b-it) fine-tuned with [unsloth](https://github.com/unslothai/unsloth) using quantization-aware training (QAT) from [torchao](https://huggingface.co/docs/transformers/main/en/quantization/torchao), and quantized with int4 weight only quantization, by PyTorch team.
Use it directly or serve using [vLLM](https://docs.vllm.ai/en/latest/) for 66% VRAM reduction (8.34 GB needed) and 1.73x speedup on H100 GPUs.


# Inference with vLLM
Install vllm nightly and torchao nightly to get some recent changes:
```
pip install vllm --pre --extra-index-url https://wheels.vllm.ai/nightly
pip install torchao
```

## Serving
Then we can serve with the following command:
```Shell
# Server
export MODEL=pytorch/gemma-3-12b-it-QAT-INT4
VLLM_DISABLE_COMPILE_CACHE=1 vllm serve $MODEL --tokenizer $MODEL -O3
```

```Shell
# Client
curl http://localhost:8000/v1/chat/completions -H "Content-Type: application/json" -d '{
  "model": "pytorch/gemma-3-12b-it-QAT-INT4",
  "messages": [
    {"role": "user", "content": "Give me a short introduction to large language models."}
  ],
  "temperature": 0.6,
  "top_p": 0.95,
  "top_k": 20,
  "max_tokens": 32768
}'
```

Note: please use `VLLM_DISABLE_COMPILE_CACHE=1` to disable compile cache when running this code, e.g. `VLLM_DISABLE_COMPILE_CACHE=1 python example.py`, since there are some issues with the composability of compile in vLLM and torchao,
this is expected be resolved in pytorch 2.8.

# Inference with Transformers

Install the required packages:
```Shell
pip install git+https://github.com/huggingface/transformers@main
pip install torchao
pip install torch
pip install accelerate
```

Example:
```Py
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "pytorch/gemma-3-12b-it-QAT-INT4"

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

# prepare the model input
prompt = "Give me a short introduction to large language model."
messages = [
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True # Switches between thinking and non-thinking modes. Default is True.
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# conduct text completion
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=32768
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()

# parsing thinking content
try:
    # rindex finding 151668 (</think>)
    index = len(output_ids) - output_ids[::-1].index(151668)
except ValueError:
    index = 0

thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("\n")
content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n")

print("thinking content:", thinking_content)
print("content:", content)
```


# Fine-tuning Recipe

Install the required packages:
```Shell
pip install torch
pip install git+https://github.com/huggingface/transformers@main
pip install --pre torchao --index-url https://download.pytorch.org/whl/nightly/cu128
pip install unsloth
pip install accelerate
```

Use the following code to fine-tune the model
```Py
# Modeled after https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/Qwen3_(14B)-Reasoning-Conversational.ipynb

from unsloth import FastModel
from unsloth.chat_templates import (
    get_chat_template,
    standardize_data_formats,
    standardize_sharegpt,
    train_on_responses_only,
)
from datasets import load_dataset
from trl import SFTConfig, SFTTrainer
import torch


max_seq_length = 2048
dtype = torch.bfloat16


# ==============
#  Model setup |
# ==============

model, tokenizer = FastModel.from_pretrained(
    model_name = "unsloth/gemma-3-12b-it",
    max_seq_length = max_seq_length,
    dtype = dtype,
    load_in_4bit = False,
    full_finetuning = False,
)

model = FastModel.get_peft_model(
    model,
    finetune_vision_layers = False,
    r = 8,
    lora_alpha = 8,
    lora_dropout = 0,
    qat_scheme = "int4",
)

tokenizer = get_chat_template(tokenizer, chat_template="gemma3")


# =============
#  Data setup |
# =============

def format_into_conversation(example):
    choices = ["A", "B", "C", "D"]
    correct_choice = choices[example["answer"]]
    question = "Choose the correct answer for the following question: "
    question += f"{example['question']}\n\n"
    question += "Choices:\n"
    question += f"A. {example['choices'][0]}\n"
    question += f"B. {example['choices'][1]}\n"
    question += f"C. {example['choices'][2]}\n"
    question += f"D. {example['choices'][3]}"
    answer = f"The correct answer is {correct_choice}."
    return {"conversations": [
        {"from": "human", "value": question},
        {"from": "gpt", "value": answer},
    ]}

def formatting_prompts_func(examples):
    convos = examples["conversations"]
    texts = [tokenizer.apply_chat_template(convo, tokenize = False, add_generation_prompt = False).removeprefix('<bos>') for convo in convos]
    return { "text" : texts, }

dataset = load_dataset("cais/mmlu", "all", split="auxiliary_train")
dataset = dataset.map(format_into_conversation)
dataset = dataset.remove_columns(["question", "subject", "choices", "answer"])
dataset = standardize_data_formats(dataset)
dataset = dataset.map(formatting_prompts_func, batched = True,)


# ========
#  Train |
# ========

trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset = dataset,
    dataset_text_field = "text",
    max_seq_length = max_seq_length,
    packing = False,
    args = SFTConfig(
        per_device_train_batch_size = 32,
        gradient_accumulation_steps = 1,
        warmup_steps = 5,
        num_train_epochs = 1,
        max_steps = 100,
        learning_rate = 2e-5,
        logging_steps = 1,
        optim = "adamw_8bit",
        weight_decay = 0.01,
        lr_scheduler_type = "linear",
        seed = 3407,
        output_dir = "outputs",
        report_to = "none",
    ),
)

trainer = train_on_responses_only(
    trainer,
    instruction_part = "<start_of_turn>user\n",
    response_part = "<start_of_turn>model\n",
)

trainer_stats = trainer.train()
```

# Model Quality
We rely on [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) to evaluate the quality of the quantized model. Here we only run on mmlu for sanity check.

| Benchmark                        |                |                                 |
|----------------------------------|----------------|---------------------------------|
|                                  | mmlu accuracy  | Normalized accuracy degradation | 
| **google/gemma-3-12b-it**        |                |                                 |
| bf16                             | 71.51          | -0%                             |
| int4                             | 69.48          | -100%                           |
| **Fine-tuned without QAT**       |                |                                 |
| bf16                             | 71.55          | +2%                             |   
| int4                             | 69.58          | -95%                            |
| **Fine-tuned with QAT**          |                |                                 |
| int4                             | 70.18          | -65.5%                          |


<details>
<summary> Reproduce Model Quality Results </summary>

## language eval
Need to install lm-eval from source:
https://github.com/EleutherAI/lm-evaluation-harness#install

```Shell
export MODEL=google/gemma-3-12b-it # or pytorch/gemma-3-12b-it-QAT-INT4
lm_eval --model hf --model_args pretrained=$MODEL --tasks mmlu --device cuda:0 --batch_size 8
```

## multi-modal eval
Need to install lmms-eval from source:
`pip install git+https://github.com/EvolvingLMMs-Lab/lmms-eval.git`

```Shell
NUM_PROCESSES=8
MAIN_PORT=12345
MODEL_ID=google/gemma-3-12b-it # or pytorch/gemma-3-12b-it-QAT-INT4
TASKS=chartqa  # or tasks from https://github.com/EvolvingLMMs-Lab/lmms-eval/tree/main/lmms_eval/models/simple
BATCH_SIZE=32
OUTPUT_PATH=./logs/

accelerate launch --num_processes "${NUM_PROCESSES}" --main_process_port "${MAIN_PORT}" -m lmms_eval \
  --model gemma3 \
  --model_args "pretrained=${MODEL_ID}" \
  --tasks "${TASKS}" \
  --batch_size "${BATCH_SIZE}" --output_path "${OUTPUT_PATH}"
```
</details>


# Peak Memory Usage

## Results

| Benchmark        |                         |                                     |
|------------------|-------------------------|-------------------------------------|
|                  | google/gemma-3-12b-it   | pytorch/gemma-3-12b-it-QAT-INT4     |
| Peak Memory (GB) | 24.50                   | 8.34 (66% reduction)                |



<details>
<summary> Reproduce Peak Memory Usage Results </summary>

We can use the following code to get a sense of peak memory usage during inference:

```Py
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TorchAoConfig

# use "google/gemma-3-12b-it" or "pytorch/gemma-3-12b-it-QAT-INT4"
model_id = "pytorch/gemma-3-12b-it-QAT-INT4"
quantized_model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype=torch.bfloat16)
tokenizer = AutoTokenizer.from_pretrained(model_id)

torch.cuda.reset_peak_memory_stats()

prompt = "Hey, are you conscious? Can you talk to me?"
messages = [
    {
        "role": "system",
        "content": "",
    },
    {"role": "user", "content": prompt},
]
templated_prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
print("Prompt:", prompt)
print("Templated prompt:", templated_prompt)
inputs = tokenizer(
    templated_prompt,
    return_tensors="pt",
).to("cuda")
generated_ids = quantized_model.generate(**inputs, max_new_tokens=128)
output_text = tokenizer.batch_decode(
    generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print("Response:", output_text[0][len(prompt):])

mem = torch.cuda.max_memory_reserved() / 1e9
print(f"Peak Memory Usage: {mem:.02f} GB")
```

</details>



# Model Performance

## Results (H100 machine)
| Benchmark (Latency)              |                         |                                    |
|----------------------------------|-------------------------|------------------------------------|
|                                  | google/gemma-3-12b-it   | pytorch/gemma-3-12b-it-QAT-INT4    |
| latency (batch_size=1)           | 3.73s                   | 2.16s (1.73x speedup)              |

<details>
<summary> Reproduce Model Performance Results </summary>

## Setup

Get vllm source code:
```Shell
git clone [email protected]:vllm-project/vllm.git
```

Install vllm
```
VLLM_USE_PRECOMPILED=1 pip install --editable .
```

Run the benchmarks under `vllm` root folder:

## benchmark_latency

### baseline
```Shell
vllm bench latency --input-len 256 --output-len 256 --model google/gemma-3-12b-it --batch-size 1
```

### INT4
```Shell
VLLM_DISABLE_COMPILE_CACHE=1 vllm bench latency --input-len 256 --output-len 256 --model pytorch/gemma-3-12b-it-QAT-INT4 --batch-size 1
```
</details>




# Paper: TorchAO: PyTorch-Native Training-to-Serving Model Optimization
The model's quantization is powered by **TorchAO**, a framework presented in the paper [TorchAO: PyTorch-Native Training-to-Serving Model Optimization](https://huggingface.co/papers/2507.16099).

**Abstract:** We present TorchAO, a PyTorch-native model optimization framework leveraging quantization and sparsity to provide an end-to-end, training-to-serving workflow for AI models. TorchAO supports a variety of popular model optimization techniques, including FP8 quantized training, quantization-aware training (QAT), post-training quantization (PTQ), and 2:4 sparsity, and leverages a novel tensor subclass abstraction to represent a variety of widely-used, backend agnostic low precision data types, including INT4, INT8, FP8, MXFP4, MXFP6, and MXFP8. TorchAO integrates closely with the broader ecosystem at each step of the model optimization pipeline, from pre-training (TorchTitan) to fine-tuning (TorchTune, Axolotl) to serving (HuggingFace, vLLM, SGLang, ExecuTorch), connecting an otherwise fragmented space in a single, unified workflow. TorchAO has enabled recent launches of the quantized Llama 3.2 1B/3B and LlamaGuard3-8B models and is open-source at this https URL .

# Resources
*   **Official TorchAO GitHub Repository:** [https://github.com/pytorch/ao](https://github.com/pytorch/ao)
*   **TorchAO Documentation:** [https://docs.pytorch.org/ao/stable/index.html](https://docs.pytorch.org/ao/stable/index.html)


# Disclaimer
PyTorch has not performed safety evaluations or red teamed the quantized models. Performance characteristics, outputs, and behaviors may differ from the original models. Users are solely responsible for selecting appropriate use cases, evaluating and mitigating for accuracy, safety, and fairness, ensuring security, and complying with all applicable laws and regulations.

Nothing contained in this Model Card should be interpreted as or deemed a restriction or modification to the licenses the models are released under, including any limitations of liability or disclaimers of warranties provided therein.


[terms]: https://ai.google.dev/gemma/terms