ai21labs/AI21-Jamba-Mini-1.6

Model Information

Built with hybrid SSM-Transformer architecture, the Jamba 1.6 family of models outperform other open, instruction-following foundation models on quality, speed, and long context performance, and rival leading closed models on quality. As open models, Jamba Mini 1.6 (12B active/52B total) and Jamba Large 1.6 (94B active/398B total) are available for private deployment, either in VPC or on-premise, and demonstrate superior performance on the kind of long context tasks that matter most to enterprises, such as RAG workflows and grounded question answering across lengthy documents.

The models are released under the Jamba Open Model License, a permissive license allowing full research use and commercial use under the license terms.

If you need to license the model for your needs, talk to us.

For more details of this model, see the release blog post.

Model Details

• Developed by: AI21
• Model type: Joint Attention and Mamba (Jamba)
• License: Jamba Open Model License
• Context length: 256K
• Knowledge cutoff date: March 5, 2024
• Supported languages: English, Spanish, French, Portuguese, Italian, Dutch, German, Arabic and Hebrew

Results on common benchmarks

Benchmark	Jamba Mini 1.6	Ministral 8B	Llama 3.1 8B	Command R7B
Arena Hard	51.2	41.35	28.17	27.95
CRAG	76.2	52	60	23.1
FinanceBench (FullDoc)	45.4	19.2	28.4	2.8
HELMET LongQA	46.9	33	29.2	9.6
LongBench	32	17.5	17.7	2

LongBench and Arena Hard scores are from official leaderboards for applicable models. Examples that couldn't fit models' context windows were scored accordingly. Due to a 32K context limit in its vLLM deployment, Ministral 8B was evaluated through its official API.

Usage

Prerequisites

In order to run optimized Mamba implementations, you first need to install mamba-ssm and causal-conv1d:

pip install mamba-ssm causal-conv1d>=1.2.0

You also have to have the model on a CUDA device.

Run the model with vLLM

The recommended way to perform efficient inference with Jamba Mini 1.6 is using vLLM. First, make sure to install vLLM (version 0.5.4 or higher is required)

pip install vllm>=0.5.4

In the example below, number_gpus should match the number of GPUs you want to deploy Jamba Mini 1.6 on. A minimum of 2 80GB GPUs is required.

from vllm import LLM, SamplingParams
from transformers import AutoTokenizer

model = "ai21labs/AI21-Jamba-Mini-1.6"
number_gpus = 2

llm = LLM(model=model,
          max_model_len=200*1024,
          tensor_parallel_size=number_gpus)

tokenizer = AutoTokenizer.from_pretrained(model)

messages = [
   {"role": "system", "content": "You are an ancient oracle who speaks in cryptic but wise phrases, always hinting at deeper meanings."},
   {"role": "user", "content": "Hello!"},
]

prompts = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)

sampling_params = SamplingParams(temperature=0.4, top_p=0.95, max_tokens=100) 
outputs = llm.generate(prompts, sampling_params)

generated_text = outputs[0].outputs[0].text
print(generated_text)
#Output: Seek and you shall find. The path is winding, but the journey is enlightening. What wisdom do you seek from the ancient echoes?

With the default BF16 precision on 2 80GB A100 GPUs and default vLLM configuration, you'll be able to perform inference on prompts up to 200K tokens long. On more than 2 80GB GPUs, you can easily fit the full 256K context.

Note: vLLM's main branch has some memory utilization improvements specific to the Jamba architecture that allow using the full 256K context length on 2 80 GPUs. You can build vLLM from source if you wish to make use of them.

ExpertsInt8 quantization

We've developed an innovative and efficient quantization technique, ExpertsInt8, designed for MoE models deployed in vLLM, including Jamba models. Using it, you'll be able to deploy Jamba Mini 1.6 on a single 80GB GPU.

In order to use ExpertsInt8, you need to use vllm version 0.5.5 or higher: pip install vllm>=0.5.5

With default vLLM configuration, you can fit prompts up to 100K on a single 80GB A100 GPU:

import os
os.environ['VLLM_FUSED_MOE_CHUNK_SIZE']='32768'    # This is a workaround a bug in vLLM's fused_moe kernel

from vllm import LLM
llm = LLM(model="ai21labs/AI21-Jamba-Mini-1.6",
          max_model_len=100*1024,
          quantization="experts_int8")

Run the model with transformers

The following example loads Jamba Mini 1.6 to the GPU in BF16 precision, uses optimized FlashAttention2 and Mamba kernels, and parallelizes the model across multiple GPUs using accelerate. Note that in half precision (FP16/BF16), Jamba Mini 1.6 is too large to fit on a single 80GB GPU, so you'll need at least 2 such GPUs.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("ai21labs/AI21-Jamba-Mini-1.6",
                                             torch_dtype=torch.bfloat16,
                                             attn_implementation="flash_attention_2",
                                             device_map="auto")

tokenizer = AutoTokenizer.from_pretrained("ai21labs/AI21-Jamba-Mini-1.6")

messages = [
   {"role": "system", "content": "You are an ancient oracle who speaks in cryptic but wise phrases, always hinting at deeper meanings."},
   {"role": "user", "content": "Hello!"},
]

input_ids = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors='pt').to(model.device)

outputs = model.generate(input_ids, max_new_tokens=216)

# Decode the output
conversation = tokenizer.decode(outputs[0], skip_special_tokens=True)

# Split the conversation to get only the assistant's response
assistant_response = conversation.split(messages[-1]['content'])[1].strip()
print(assistant_response)
# Output: Seek and you shall find. The path is winding, but the journey is enlightening. What wisdom do you seek from the ancient echoes?

Note: Versions 4.44.0 and 4.44.1 of transformers have a bug that restricts the ability to run the Jamba architecture. Make sure you're not using these versions.

Note: If you're having trouble installing mamba-ssm and causal-conv1d for the optimized Mamba kernels, you can run Jamba Mini 1.6 without them, at the cost of extra latency. In order to do that, add the kwarg use_mamba_kernels=False when loading the model via AutoModelForCausalLM.from_pretained().

Load the model in 8-bit

Using 8-bit precision, it is possible to fit up to 140K sequence length on a single 80GB GPU. You can easily quantize the model to 8-bit using bitsandbytes. In order to not degrade model quality, we recommend to exclude the Mamba blocks from the quantization:

from transformers import AutoModelForCausalLM, BitsAndBytesConfig
quantization_config = BitsAndBytesConfig(load_in_8bit=True,
                                         llm_int8_skip_modules=["mamba"])
model = AutoModelForCausalLM.from_pretrained("ai21labs/AI21-Jamba-Mini-1.6",
                                             torch_dtype=torch.bfloat16,
                                             attn_implementation="flash_attention_2",
                                             quantization_config=quantization_config)

Load the model on CPU

If you don't have access to a GPU, you can also load and run Jamba Mini 1.6 on a CPU. Note this will result in poor inference performance.

from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("ai21labs/AI21-Jamba-Mini-1.6",
                                             use_mamba_kernels=False)

Model features

Tool use with Jamba

Jamba Mini 1.6 supports tool use capabilities in accordance with Huggingface's tool use API. The tools defined by the user are inserted into a dedicated section in the chat template which the model was trained to recognize.

Given a conversation that contains tools, the model can output content, tool invocations or both.

Tool usage example

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("ai21labs/AI21-Jamba-Mini-1.6")

messages = [
    {
        "role": "user", 
        "content": "What's the weather like right now in Jerusalem and in London?"
    }
]

tools = [
    {
        'type': 'function', 
        'function': {
            'name': 'get_current_weather', 
            'description': 'Get the current weather', 
            'parameters': {
                'type': 'object', 
                'properties': {
                    'location': {'type': 'string', 'description': 'The city and state, e.g. San Francisco, CA'}, 
                    'format': {'type': 'string', 'enum': ['celsius', 'fahrenheit'], 'description': 'The temperature unit to use. Infer this from the users location.'}
                }, 
                'required': ['location', 'format']
            }
        }
    }
]

prompt = tokenizer.apply_chat_template(
    messages,
    tools=tools,
    tokenize=False,
)

Output:

<tool_calls>[
    {"name": "get_current_weather", "arguments": {"location": "Jerusalem", "format": "celsius"}},
    {"name": "get_current_weather", "arguments": {"location": "celsius", "format": "celsius"}}
]</tool_calls>

Feeding back tool responses into the model

Now that the model has called the tools, we need to feed the tool responses back to the model. In the next call, send the assistant message with the tool_messages field, as shown below, along with additional tool messages (in the corresponding order) that contain the tool outputs.

The arguments field for each tool call can be either a dict or a JSON string.

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("ai21labs/AI21-Jamba-Mini-1.6")

# Note that you must send the tool responses in the same order as the model called the tools:
messages = [
    {
        "role": "user",
        "content": "What's the weather like right now in Jerusalem and in London?"
    },
    {
        "role": "assistant",
        "content": null,
        "tool_calls": [
            {
                "name": "get_current_weather",
                "arguments": "{\"location\": \"Jerusalem\", \"format\": \"celsius\"}"
            },
            {
                "name": "get_current_weather",
                "arguments": "{\"location\": \"London\", \"format\": \"celsius\"}"
            }
        ]
    },
    {
        "role": "tool",
        "content": "The weather in Jerusalem is 18 degrees celsius."
    },
    {
        "role": "tool",
        "content": "The weather in London is 8 degrees celsius."
    }
]

tool_use_prompt = tokenizer.apply_chat_template(
    messages,
    tools=tools,
    tokenize=False,
)

example output:

The weather in Jerusalem is currently 18 degrees Celsius. In London, it is 8 degrees Celsius.

Fine-tuning examples

The examples below use the SFTTrainer from huggingface/trl, so ensure it's installed:

pip install trl

Full Fine-tuning example

To train a full finetune using AWS multi nodes and FSDP configuration, follow the instructions here hf-finetune-sagemaker

LoRA example

Here is an example of fine-tuning with LoRA PEFT, in bfloat16 (requires ~130GB GPU RAM, so e.g. 2xA100 80GB):

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from datasets import load_dataset
from trl import SFTTrainer, SFTConfig
from peft import LoraConfig

tokenizer = AutoTokenizer.from_pretrained("ai21labs/AI21-Jamba-Mini-1.6")
model = AutoModelForCausalLM.from_pretrained(
    "ai21labs/AI21-Jamba-Mini-1.6",
    device_map="auto",
    torch_dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
)

lora_config = LoraConfig(
    r=8,
    target_modules=[
        "embed_tokens",
        "x_proj", "in_proj", "out_proj", # mamba
        "gate_proj", "up_proj", "down_proj", # mlp
        "q_proj", "k_proj", "v_proj", "o_proj", # attention
    ],
    task_type="CAUSAL_LM",
    bias="none",
)

dataset = load_dataset("philschmid/dolly-15k-oai-style", split="train")
training_args = SFTConfig(
    output_dir="/dev/shm/results",
    logging_dir="./logs",
    num_train_epochs=2,
    per_device_train_batch_size=4,
    learning_rate=1e-5,
    logging_steps=10,
    gradient_checkpointing=True,
    max_seq_length=4096,
    save_steps=100,
)
trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    args=training_args,
    peft_config=lora_config,
    train_dataset=dataset,
)
trainer.train()

Note that the dataset in the example uses conversational format (with messages column), so SFTTrainer automatically applies Jamba's chat-template as explained in TRL docs.

QLoRA example

To fit fine-tuning on a single 80GB GPU, you can levarage QLoRA which combines LoRA with the frozen model quantized to 4-bit:

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from datasets import load_dataset
from trl import SFTTrainer, SFTConfig
from peft import LoraConfig

tokenizer = AutoTokenizer.from_pretrained("ai21labs/AI21-Jamba-Mini-1.6")
quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
)
model = AutoModelForCausalLM.from_pretrained(
    "ai21labs/AI21-Jamba-Mini-1.6",
    device_map="auto",
    quantization_config=quantization_config,
    torch_dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
)
lora_config = LoraConfig(
    r=8,
    target_modules=[
        "embed_tokens", "x_proj", "in_proj", "out_proj", # mamba
        "gate_proj", "up_proj", "down_proj",  # mlp
        "q_proj", "k_proj", "v_proj", "o_proj", # attention
    ],
    task_type="CAUSAL_LM",
    bias="none",
)

dataset = load_dataset("philschmid/dolly-15k-oai-style", split="train")
training_args = SFTConfig(
    output_dir="./results",
    logging_dir="./logs",
    num_train_epochs=2,
    per_device_train_batch_size=8,
    learning_rate=1e-5,
    logging_steps=1,
    gradient_checkpointing=True,
    gradient_checkpointing_kwargs={"use_reentrant": False},
    save_steps=100,
    max_seq_length=4096,
)
trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    args=training_args,
    peft_config=lora_config,
    train_dataset=dataset,
)
trainer.train()

Note: the above example reqiures the bitsandbytes package for the 4-bit quantization:

pip install bitsandbytes

About AI21

AI21 builds reliable, practical, and scalable AI solutions for the enterprise. The Jamba models are available in the AI21 Studio and in leading cloud partners. To learn more about how Jamba Mini 1.6 and Jamba Large 1.6 can bring real world value to your organization, let’s talk.