--- 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 () 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('') 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 = "user\n", response_part = "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% |
Reproduce Model Quality Results ## 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}" ```
# 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) |
Reproduce Peak Memory Usage Results 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") ```
# 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) |
Reproduce Model Performance Results ## Setup Get vllm source code: ```Shell git clone git@github.com: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 ```
# 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