README.md

---
{}
---
# Hymba: A Hybrid-head Architecture for Small Language Models

[[Slide](https://docs.google.com/presentation/d/1uidqBfDy8a149yE1-AKtNnPm1qwa01hp8sOj3_KAoMI/edit#slide=id.g2f73b22dcb8_0_1017)][Technical Report]  **!!! This huggingface repo is still under development.**

Developed by Deep Learning Efficiency Research (DLER) team at NVIDIA Research.


## Hymba: A Novel LM Architecture
- Fuse attention heads and SSM heads within the same layer, offering parallel and complementary processing of the same inputs

<div align="center">
<img src="https://huggingface.co/nvidia/Hymba-1.5B-Instruct/resolve/main/images/module.png" alt="Hymba Module" width="600">
</div>

- Introduce meta tokens that are prepended to the input sequences and interact with all subsequent tokens, thus storing important information and alleviating the burden of "forced-to-attend" in attention

- Integrate with cross-layer KV sharing and global-local attention to further boost memory and computation efficiency

<div align="center">
<img src="https://huggingface.co/nvidia/Hymba-1.5B-Instruct/resolve/main/images/macro_arch.png" alt="Hymba Model" width="600">
</div>


## Hymba: Performance Highlights
- [Hymba-1.5B-Base](https://huggingface.co/nvidia/Hymba-1.5B): Outperform all sub-2B public models, e.g., matching Llama 3.2 3B’s commonsense reasoning accuracy, being 3.49× faster, and reducing cache size by 11.7×

<div align="center">
<img src="https://huggingface.co/nvidia/Hymba-1.5B-Instruct/resolve/main/images/performance1.png" alt="Compare with SoTA Small LMs" width="600">
</div>


- Hymba-1.5B-Instruct: Outperform SOTA small LMs.


<div align="center">
<img src="https://huggingface.co/nvidia/Hymba-1.5B-Instruct/resolve/main/images/instruct_performance.png" alt="Compare with SoTA Small LMs" width="600">
</div>


## Hymba-1.5B-Instruct: Model Usage

We release our Hymba-1.5B-Instruct model and offer the instructions to use our model as follows.

### Step 1: Environment Setup

Since our model employs [FlexAttention](https://pytorch.org/blog/flexattention/), which relies on Pytorch2.5 and other related dependencies, we provide three ways to set up the environment:

- **[Pip]** Install the related packages using our provided `requirement.txt`:
```
pip install -r https://huggingface.co/nvidia/Hymba-1.5B-Instruct/resolve/main/requirements.txt
```

- **[Docker]** We have prepared a docker image with all of Hymba's dependencies installed. You can download our docker image and start a container using the following commands:

```
wget http://10.137.9.244:8000/hymba_docker.tar
docker load -i hymba.tar
docker run --security-opt seccomp=unconfined --gpus all -v /home/$USER:/home/$USER -it hymba:v1 bash
```

- **[Internal Only]** If you are an internal user from NVIDIA and are using the ORD cluster, you can use our prepared `sqsh` file to apply for an interactive node:

   ```
   srun -A nvr_lpr_llm --partition interactive --time 4:00:00 --gpus 8 --container-image /lustre/fsw/portfolios/nvr/users/yongganf/docker/megatron_py25.sqsh --container-mounts=$HOME:/home,/lustre:/lustre  --pty bash
   ```

### Step 2: Chat with Hymba
After setting up the environment, you can use the following script to chat with our Model

```
from transformers import LlamaTokenizer, AutoModelForCausalLM, AutoTokenizer, AutoModel
from huggingface_hub import login
import torch

login()

# Load LLaMA2's tokenizer
tokenizer = LlamaTokenizer.from_pretrained("meta-llama/Llama-2-7b")
        
# Load Hymba-1.5B
model = AutoModelForCausalLM.from_pretrained("nvidia/Hymba-1.5B-Instruct", trust_remote_code=True).cuda().to(torch.bfloat16)

# Chat with our model
def chat_with_model(prompt, model, tokenizer, max_length=64):
    inputs = tokenizer(prompt, return_tensors="pt").to('cuda')
    outputs = model.generate(inputs.input_ids, max_length=max_length, do_sample=False, temperature=0.7, use_cache=True)
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return response

print("Chat with the model (type 'exit' to quit):")
while True:
    print("User:")
    prompt = input()
    if prompt.lower() == "exit":
        break
    
    # Get the model's response
    response = chat_with_model(prompt, model, tokenizer)
    
    print(f"Model: {response}")

```