---
license: gemma
language:
- uk
pipeline_tag: text-generation
tags:
- gemma3
- fp8
- quantized
- ukrainian
- conversational
base_model:
- INSAIT-Institute/MamayLM-Gemma-3-12B-IT-v1.0
library_name: transformers
---

# MamayLM-Gemma-3-12B-IT-v1.0-FP8-Static-Nadiia-v2

## Model description

**MamayLM-Gemma-3-12B-IT-v1.0-FP8-Static-Nadiia-v2** is an FP8 static-quantized version of the instruction-tuned [INSAIT-Institute/MamayLM-Gemma-3-12B-IT-v1.0](https://huggingface.co/INSAIT-Institute/MamayLM-Gemma-3-12B-IT-v1.0). based on Gemma 3 12B model, prepared for efficient inference while preserving quality for Ukrainian-language assistant workloads.

This version was quantized on a **NVIDIA B200** GPU using the private **Nadiia-v2** dataset for calibration.

## Intended use

This model is intended for:

- Ukrainian-language assistant scenarios
- Instruction following
- Customer support and helpdesk-style conversations
- General text generation
- Evaluation of FP8 static quantization quality/performance trade-offs

## Quantization Details

- **Base instruction model:** [INSAIT-Institute/MamayLM-Gemma-3-12B-IT-v1.0](https://huggingface.co/INSAIT-Institute/MamayLM-Gemma-3-12B-IT-v1.0).
- **Quantization format:** FP8 static quantization (E4M3)
- **Calibration dataset:** private dataset `Nadiia-v2`
- **Quantization hardware:** NVIDIA B200 GPU

## Usage

Transformers

```
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "oshkorinova/MamayLM-Gemma-3-12B-IT-v1.0-FP8-Static-Nadiia-v2"

tokenizer = AutoTokenizer.from_pretrained(
    model_id,
    trust_remote_code=True,
)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    trust_remote_code=True,
    torch_dtype="auto",
)

messages = [
    {"role": "user", "content": "Поясни простими словами, що таке квантізація моделі."}
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

with torch.no_grad():
    outputs = model.generate(
        inputs,
        max_new_tokens=200,
        do_sample=True,
        temperature=0.7,
    )

print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

vLLM

```
vllm serve oshkorinova/MamayLM-Gemma-3-12B-IT-v1.0-FP8-Static-Nadiia-v2 \
  --trust-remote-code \
  --dtype auto \
  --max-model-len 8192 \
  --limit-mm-per-prompt '{"image": 0}'
```

Example request:

```
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "oshkorinova/MamayLM-Gemma-3-12B-IT-v1.0-FP8-Static-Nadiia-v2",
    "messages": [
      {"role": "user", "content": "Напиши коротке пояснення, що таке FP8-квантізація."}
    ],
    "temperature": 0.7,
    "max_tokens": 200
  }'
```

## Acknowledgements

We would like to thank:

- [Beyond.pl](https://www.beyond.pl/en/) for providing compute resources used for FP8 quantization  
- [glibesyck](https://huggingface.co/glibesyck) for running the evaluation benchmarks