Quickstart | Web UI | Configurations | MacOS | Example notebooks | FAQ

AirLLM optimizes inference memory usage, allowing 70B large language models to run inference on a single 4GB GPU card without quantization, distillation and pruning. And you can run 405B Llama3.1 on 8GB vram now.

• Best AI Game Sprite Generator

• Best AI Facial Expression Editor

[2024/08/20] v2.11.0: Support Qwen2.5

[2024/08/18] v2.10.1 Support CPU inference. Support non sharded models. Thanks @NavodPeiris for the great work!

[2024/07/30] Support Llama3.1 405B (example notebook). Support 8bit/4bit quantization.

[2024/04/20] AirLLM supports Llama3 natively already. Run Llama3 70B on 4GB single GPU.

[2023/12/25] v2.8.2: Support MacOS running 70B large language models.

[2023/12/20] v2.7: Support AirLLMMixtral.

[2023/12/20] v2.6: Added AutoModel, automatically detect model type, no need to provide model class to initialize model.

[2023/12/18] v2.5: added prefetching to overlap the model loading and compute. 10% speed improvement.

[2023/12/03] added support of ChatGLM, QWen, Baichuan, Mistral, InternLM!

[2023/12/02] added support for safetensors. Now support all top 10 models in open llm leaderboard.

[2023/12/01] airllm 2.0. Support compressions: 3x run time speed up!

[2023/11/20] airllm Initial version!

• Quick start
• Web UI (Local App)
• Model Compression
• Configurations
• Run on MacOS
• Example notebooks
• Supported Models
• Acknowledgement
• FAQ

First, install the airllm pip package.

pip install airllm

Then, initialize AirLLMLlama2, pass in the huggingface repo ID of the model being used, or the local path, and inference can be performed similar to a regular transformer model.

(You can also specify the path to save the splitted layered model through layer_shards_saving_path when init AirLLMLlama2.

from airllm import AutoModel

MAX_LENGTH = 128
# could use hugging face model repo id:
model = AutoModel.from_pretrained("garage-bAInd/Platypus2-70B-instruct")

# or use model's local path...
#model = AutoModel.from_pretrained("/home/ubuntu/.cache/huggingface/hub/models--garage-bAInd--Platypus2-70B-instruct/snapshots/b585e74bcaae02e52665d9ac6d23f4d0dbc81a0f")

input_text = [
        'What is the capital of United States?',
        #'I like',
    ]

input_tokens = model.tokenizer(input_text,
    return_tensors="pt", 
    return_attention_mask=False, 
    truncation=True, 
    max_length=MAX_LENGTH, 
    padding=False)
           
generation_output = model.generate(
    input_tokens['input_ids'].cuda(), 
    max_new_tokens=20,
    use_cache=True,
    return_dict_in_generate=True)

output = model.tokenizer.decode(generation_output.sequences[0])

print(output)

Note: During inference, the original model will first be decomposed and saved layer-wise. Please ensure there is sufficient disk space in the huggingface cache directory.

This project also includes a local web interface in app.py for chat, model settings, and Hugging Face downloads.

pip install -r requirements.txt
python app.py

Open http://127.0.0.1:8000.

Dependency note:

• transformers is intentionally pinned to 4.56.2 for runtime compatibility with this app.
• Avoid upgrading transformers to latest unless you are testing compatibility changes.

1. Open the Settings tab.
2. In Model Config, set:
   • AIRLLM_MODEL_BASE_DIR to your local model directory root (for example O:\\AI\\Models).
   • AIRLLM_DEVICE to cuda:0 (or cpu if no CUDA GPU is available).
3. Click Scan Base Directory to detect compatible local model folders.
4. Select a model from Discovered Models (or set AIRLLM_MODEL_PATH manually).
5. Click Load / Reload Model.
6. Go to Chat and send a prompt.

1. Open Settings -> Hugging Face Download.
2. Enter the model repo ID (for example Qwen/Qwen3-Coder-Next).
3. Optionally set revision, target subdir, and allow/ignore patterns.
4. Click Download from Hugging Face.
5. Monitor progress in Download Progress (status, file counts, logs).
6. If enabled, Set downloaded model as active updates AIRLLM_MODEL_PATH automatically after download.

Notes:

• The web UI tracks one active download job at a time.
• Reloading the page reconnects to an in-progress download job.
• AirLLM loading in this UI expects Transformers-style model directories (config.json + weight files), not GGUF-only folders.

• Chat tab: send prompts to the loaded model (Enter sends, Shift+Enter adds a newline).
• Settings -> Model Config: configure AIRLLM_MODEL_ID, AIRLLM_MODEL_PATH, AIRLLM_MODEL_BASE_DIR, and AIRLLM_DEVICE; scan the base directory for compatible local Transformers models.
• Settings -> Hugging Face Download: download a model into the base directory with revision/pattern filters, with live progress logs in the UI.
• Download persistence: if a download is already running, reloading the page reconnects to the active job and restores progress.
• Dark mode: toggle from the top bar.

• AIRLLM_MODEL_ID: model repo id (for example TinyLlama/TinyLlama-1.1B-Chat-v1.0).
• AIRLLM_MODEL_PATH: explicit local model directory; overrides base-dir resolution when set.
• AIRLLM_MODEL_BASE_DIR: root folder used for model scans and HF downloads.
• AIRLLM_DEVICE: runtime device (for example cuda:0 or cpu).
• HF_TOKEN: optional Hugging Face token for gated/private repositories.
• PORT: web server port (default 8000).

When Persist selected values to project .env is enabled in Settings, changes are written back to the local .env file.

By default the app binds to localhost. To expose it on your LAN, run:

uvicorn app:app --host 0.0.0.0 --port 8000

Then open http://<your-machine-ip>:8000 from another device on the same network.

We just added model compression based on block-wise quantization-based model compression. Which can further speed up the inference speed for up to 3x , with almost ignorable accuracy loss! (see more performance evaluation and why we use block-wise quantization in this paper)

• Step 1. make sure you have bitsandbytes installed by pip install -U bitsandbytes
• Step 2. make sure airllm verion later than 2.0.0: pip install -U airllm
• Step 3. when initialize the model, passing the argument compression ('4bit' or '8bit'):

model = AutoModel.from_pretrained("garage-bAInd/Platypus2-70B-instruct",
                     compression='4bit' # specify '8bit' for 8-bit block-wise quantization 
                    )

Quantization normally needs to quantize both weights and activations to really speed things up. Which makes it harder to maintain accuracy and avoid the impact of outliers in all kinds of inputs.

While in our case the bottleneck is mainly at the disk loading, we only need to make the model loading size smaller. So, we get to only quantize the weights' part, which is easier to ensure the accuracy.

When initialize the model, we support the following configurations:

• compression: supported options: 4bit, 8bit for 4-bit or 8-bit block-wise quantization, or by default None for no compression
• profiling_mode: supported options: True to output time consumptions or by default False
• layer_shards_saving_path: optionally another path to save the splitted model
• hf_token: huggingface token can be provided here if downloading gated models like: meta-llama/Llama-2-7b-hf
• prefetching: prefetching to overlap the model loading and compute. By default, turned on. For now, only AirLLMLlama2 supports this.
• delete_original: if you don't have too much disk space, you can set delete_original to true to delete the original downloaded hugging face model, only keep the transformed one to save half of the disk space.

Just install airllm and run the code the same as on linux. See more in Quick Start.

• make sure you installed mlx and torch
• you probably need to install python native see more here
• only Apple silicon is supported

Example [python notebook] (https://github.com/lyogavin/airllm/blob/main/air_llm/examples/run_on_macos.ipynb)

Example colabs here:

from airllm import AutoModel
MAX_LENGTH = 128
model = AutoModel.from_pretrained("THUDM/chatglm3-6b-base")
input_text = ['What is the capital of China?',]
input_tokens = model.tokenizer(input_text,
    return_tensors="pt", 
    return_attention_mask=False, 
    truncation=True, 
    max_length=MAX_LENGTH, 
    padding=True)
generation_output = model.generate(
    input_tokens['input_ids'].cuda(), 
    max_new_tokens=5,
    use_cache= True,
    return_dict_in_generate=True)
model.tokenizer.decode(generation_output.sequences[0])

from airllm import AutoModel
MAX_LENGTH = 128
model = AutoModel.from_pretrained("Qwen/Qwen-7B")
input_text = ['What is the capital of China?',]
input_tokens = model.tokenizer(input_text,
    return_tensors="pt", 
    return_attention_mask=False, 
    truncation=True, 
    max_length=MAX_LENGTH)
generation_output = model.generate(
    input_tokens['input_ids'].cuda(), 
    max_new_tokens=5,
    use_cache=True,
    return_dict_in_generate=True)
model.tokenizer.decode(generation_output.sequences[0])

from airllm import AutoModel
MAX_LENGTH = 128
model = AutoModel.from_pretrained("baichuan-inc/Baichuan2-7B-Base")
#model = AutoModel.from_pretrained("internlm/internlm-20b")
#model = AutoModel.from_pretrained("mistralai/Mistral-7B-Instruct-v0.1")
input_text = ['What is the capital of China?',]
input_tokens = model.tokenizer(input_text,
    return_tensors="pt", 
    return_attention_mask=False, 
    truncation=True, 
    max_length=MAX_LENGTH)
generation_output = model.generate(
    input_tokens['input_ids'].cuda(), 
    max_new_tokens=5,
    use_cache=True,
    return_dict_in_generate=True)
model.tokenizer.decode(generation_output.sequences[0])

A lot of the code are based on SimJeg's great work in the Kaggle exam competition. Big shoutout to SimJeg:

GitHub account @SimJeg, the code on Kaggle, the associated discussion.

safetensors_rust.SafetensorError: Error while deserializing header: MetadataIncompleteBuffer

If you run into this error, most possible cause is you run out of disk space. The process of splitting model is very disk-consuming. See this. You may need to extend your disk space, clear huggingface .cache and rerun.

Most likely you are loading QWen or ChatGLM model with Llama2 class. Try the following:

For QWen model:

from airllm import AutoModel #<----- instead of AirLLMLlama2
AutoModel.from_pretrained(...)

For ChatGLM model:

from airllm import AutoModel #<----- instead of AirLLMLlama2
AutoModel.from_pretrained(...)

Some models are gated models, needs huggingface api token. You can provide hf_token:

model = AutoModel.from_pretrained("meta-llama/Llama-2-7b-hf", #hf_token='HF_API_TOKEN')

Some model's tokenizer doesn't have padding token, so you can set a padding token or simply turn the padding config off:

input_tokens = model.tokenizer(input_text,
   return_tensors="pt", 
   return_attention_mask=False, 
   truncation=True, 
   max_length=MAX_LENGTH, 
   padding=False  #<-----------   turn off padding 
)

If you find AirLLM useful in your research and wish to cite it, please use the following BibTex entry:

@software{airllm2023,
  author = {Gavin Li},
  title = {AirLLM: scaling large language models on low-end commodity computers},
  url = {https://github.com/lyogavin/airllm/},
  version = {0.0},
  year = {2023},
}

Welcomed contributions, ideas and discussions!

If you find it useful, please ⭐ or buy me a coffee! 🙏