Downtown-Case/aws-prototyping_MegaBeam-Mistral-7B-512K-exl2-8.0bpw

MegaBeam-Mistral-7B-512k Model

MegaBeam-Mistral-7B-512k is a Large-Context LLM that supports 524,288 tokens in its context. MegaBeam-Mistral-7B-512k was trained on Mistral-7B Instruct-v0.2, and can be deployed using various serving frameworks like vLLM and Amazon SageMaker's DJL endpoint.

Evaluations

We evaluated MegaBeam-Mistral-7B-512k on three long-context benchmarks. For each benchmark, we deployed the MegaBeam-Mistral-7B-512k model with vLLM (v0.5.1) on an EC2 instance and obtained LLM responses through the OpenAI API provided by vLLM.

1. Needle In A Haystack - Pressure Testing LLMs

The Arize-ai NIAH varies the target random number and introduces a random city for each question, requiring the LLM to extract the random number from various selected context locations.

MegaBeam-Mistral-7B-512k scored 100% on this NIAH benchmark as shown in this plot.

2. RULER: What’s the Real Context Size of Your Long-Context Language Models?

The RULER benchmark evaluates long-context language models across four task categories - Retrieval, Multi-hop Tracing, Aggregation, and Question Answering - with a total of 13 tasks. RULER goes beyond simple in-context recall by introducing more complex long-context scenarios.

MegaBeam-Mistral-7B-512k scored an average of 88.70 across different context lengths as shown in this table (adapted from the RULER project).

Models	4K	8K	16K	32K	64K	128K	Avg.
MegaBeam-Mistral-7B-512k	93.3	91.8	91.5	88.9	83.7	82.8	88.7
Gemini-1.5-pro	96.7	95.8	96	95.9	95.9	94.4	95.8
GPT-4-1106-preview	96.6	96.3	95.2	93.2	87	81.2	91.6
Llama3.1 (70B)	96.5	95.8	95.4	94.8	88.4	66.6	89.6
Qwen2 (72B)	96.9	96.1	94.9	94.1	79.8	53.7	85.9
Command-R-plus (104B)	95.6	95.2	94.2	92	84.3	63.1	87.4
GLM4 (9B)	94.7	92.8	92.1	89.9	86.7	83.1	89.9
Llama3.1 (8B)	95.5	93.8	91.6	87.4	84.7	77.0	88.3
Command-R (35B)	93.8	93.3	92.4	89.5	84.9	76	88.3
GradientAI/Llama3 (70B)	95.1	94.4	90.8	85.4	82.9	72.1	86.5
Mixtral-8x22B (39B/141B)	95.6	94.9	93.4	90.9	84.7	31.7	81.9
Yi (34B)	93.3	92.2	91.3	87.5	83.2	77.3	87.5
Phi3-medium (14B)	93.3	93.2	91.1	86.8	78.6	46.1	81.5
Mixtral-8x7B (12.9B/46.7B)	94.9	92.1	92.5	85.9	72.4	44.5	80.4
GradientAI/Llama3 (8B)	92.8	90.3	85.7	79.9	76.3	69.5	82.4
FILM-7B (7B)	92.8	88.2	88.1	86.9	70.1	27.1	75.5
Mistral-7B-instruct-v0.2 (7B)	93.6	91.2	87.2	75.4	49	13.8	68.4
Mistral-Nemo	87.8	87.2	87.7	69.0	46.8	19.0	66.2
GLM3 (6B)	87.8	83.4	78.6	69.9	56	42	69.6
LWM (7B)	82.3	78.4	73.7	69.1	68.1	65	72.8

This table shows how `MegaBeam-Mistral-7B-512k` performed on 13 RULER tasks with increasing context lengths.

Task	Category	4096	8192	16384	32768	65536	131072
niah_single_1	Retrieval	100	100	100	100	100	100
niah_single_2	Retrieval	98.6	97.8	98.8	98.2	99.4	99.6
niah_single_3	Retrieval	100	100	100	99.8	100	99.8
niah_multikey_1	Retrieval	98.8	99.6	99.2	99	99.6	99.6
niah_multikey_2	Retrieval	100	100	100	99.8	99.4	98.6
niah_multikey_3	Retrieval	99.8	99.4	99.8	100	98.6	97.8
niah_multivalue	Retrieval	97.1	93.8	91.85	83.5	80.3	71.45
niah_multiquery	Retrieval	99.95	99.9	99.85	99.3	99.55	99.3
vt	Multi-hop Tracing	99.2	97.88	96.44	96.12	91.6	89.08
cwe	Aggregation	98.2	90.62	75.6	52.72	5.9	0.94
fwe	Aggregation	81.47	80.07	95.87	96.33	83.73	96.87
qa_1	Q & A	85.6	82	80.6	83	80.6	77.4
qa_2	Q & A	53.8	52	51.6	48.4	49.2	45.8
average	ALL	93.3	91.8	91.5	88.9	83.7	82.8
Total Average	88.7

3. InfiniteBench: Extending Long Context Evaluation Beyond 100K Tokens

InfiniteBench developed 12 tasks to evaluate an LLM's capability to process, comprehend, and reason with extended contexts, specifically those with over 100,000 tokens.

We combine the InfiniteBench project's evaluation results for SOTA LLMs with MegaBeam-Mistral-7B-512k's result in this table.

Task Name	MegaBeam-Mistral -7B-512k	GPT-4-1106 -preview	YaRN-Mistral -7B	Kimi-Chat	Claude 2	Yi-34B -200K
PassKey	100%	100%	92.71%	98.14%	97.80%	100.00%
Retrv.Num	99.49%	100%	56.61%	95.42%	98.14%	100.00%
Retrv.KV	24.20%	89.00%	< 5%	53.60%	65.40%	< 5%
En.Sum	34.66%	14.73%	9.09%	17.93%	14.45%	< 5%
En.QA	20.32%	22.22%	9.55%	16.52%	11.97%	12.17%
En.MC	61.57%	67.25%	27.95%	72.49%	62.88%	38.43%
En.Dia	10.50%	8.50%	7.50%	11.50%	46.50%	< 5%
Zh.QA	19.54%	25.96%	14.43%	17.93%	9.64%	13.61%
Code.Debug	26.14%	39.59%	< 5%	18.02%	< 5%	< 5%
Code.Run	2%	23.25%	< 5%	< 5%	< 5%	< 5%
Math.Calc	0%	< 5%	< 5%	< 5%	< 5%	< 5%
Math.Find	20%	60.00%	17.14%	12.57%	32.29%	25.71%
Average	34.87%	46.08%	20.41%	34.93%	37.21%	25.41%

Example use case

This example demonstrates MegaBeam-Mistral-7B-512k's long context capability by processing a large file that includes hundreds of files from a single Git repository. This can be useful for onboarding new developers.

Serve MegaBeam-Mistral-7B-512k on EC2 instances

On an AWS g5.48xlarge instance, install vLLM as per vLLM docs.

pip install vllm==0.5.1

Start the server

VLLM_ENGINE_ITERATION_TIMEOUT_S=3600 python3 -m vllm.entrypoints.openai.api_server \
        --model aws-prototyping/MegaBeam-Mistral-7B-512k \
        --tensor-parallel-size 8 \
        --revision g5-48x

Important Note - In the repo revision g5-48x, config.json has been updated to set max_position_embeddings to 288,800, fitting the model's KV cache on a single g5.48xlarge instance with 8 A10 GPUs (24GB RAM per GPU).

On an instance with larger GPU RAM (e.g. p4d.24xlarge), simply remove the revision argument in order to support the full sequence length of 524,288 tokens:

VLLM_ENGINE_ITERATION_TIMEOUT_S=3600 python3 -m vllm.entrypoints.openai.api_server \
        --model aws-prototyping/MegaBeam-Mistral-7B-512k \
        --tensor-parallel-size 8 \

Run the client

from openai import OpenAI

# Modify OpenAI's API key and API base to use vLLM's API server.
openai_api_key = "EMPTY"
openai_api_base = "http://localhost:8000/v1"

client = OpenAI(
    # defaults to os.environ.get("OPENAI_API_KEY")
    api_key=openai_api_key,
    base_url=openai_api_base,
)

models = client.models.list()
model = models.data[0].id

chat_completion = client.chat.completions.create(
        messages = [
            {"role": "user", "content": "What is your favourite condiment?"}, # insert your long context here
            {"role": "assistant", "content": "Well, I'm quite partial to a good squeeze of fresh lemon juice. It adds just the right amount of zesty flavour to whatever I'm cooking up in the kitchen!"},
            {"role": "user", "content": "Do you have mayonnaise recipes?"} # insert your long context here
        ],
        model=model,
)

print("Chat completion results:")
print(chat_completion)

Deploy the model on a SageMaker Endpoint

To deploy MegaBeam-Mistral-7B-512k on a SageMaker endpoint, please follow this SageMaker DJL deployment guide.

Run the following Python code in a SageMaker notebook (with each block running in a separate cell)

import sagemaker
from sagemaker import Model, image_uris, serializers, deserializers

sagemaker_session = sagemaker.Session()
region = sagemaker_session.boto_region_name
role = sagemaker.get_execution_role()

%%writefile serving.properties
engine=Python
option.model_id=aws-prototyping/MegaBeam-Mistral-7B-512k
option.revision=g5-48x
option.dtype=bf16
option.task=text-generation
option.rolling_batch=vllm
option.tensor_parallel_degree=8
option.device_map=auto

%%sh
mkdir mymodel
mv serving.properties mymodel/
tar czvf mymodel.tar.gz mymodel/
rm -rf mymodel

image_uri = image_uris.retrieve(
        framework="djl-deepspeed",
        region=region,
        version="0.27.0"
)

s3_code_prefix = "megaBeam-mistral-7b-512k/code"
bucket = sagemaker_session.default_bucket()  # bucket to house artifacts
code_artifact = sagemaker_session.upload_data("mymodel.tar.gz", bucket, s3_code_prefix)
print(f"S3 Code or Model tar ball uploaded to --- > {code_artifact}")
model = Model(image_uri=image_uri, model_data=code_artifact, role=role)

instance_type = "ml.g5.48xlarge"
endpoint_name = sagemaker.utils.name_from_base("megaBeam-mistral-7b-512k")
model.deploy(initial_instance_count=1,
             instance_type=instance_type,
             endpoint_name=endpoint_name
            )

# our requests and responses will be in json format so we specify the serializer and the deserializer
predictor = sagemaker.Predictor(
    endpoint_name=endpoint_name,
    sagemaker_session=sagemaker_session,
    serializer=serializers.JSONSerializer(),
)

# test the endpoint
input_str = """<s>[INST] What is your favourite condiment? [/INST]
Well, I'm quite partial to a good squeeze of fresh lemon juice. It adds just the right amount of zesty flavour to whatever I'm cooking up in the kitchen!</s> "
[INST] Do you have mayonnaise recipes? [/INST]"""
predictor.predict(
    {"inputs": input_str, "parameters": {"max_new_tokens": 75}}
)

Invoke the model on a SageMaker Endpoint

To use MegaBeam-Mistral-7B-512k on a SageMaker endpoint, please try following this example:

import boto3
import json

def call_endpoint(text:str, endpoint_name:str):
    client = boto3.client("sagemaker-runtime")

    parameters = {
        "max_new_tokens": 450,
        "do_sample": True,
        "temperature": 0.7,
    }

    payload = {"inputs": text, "parameters": parameters}

    response = client.invoke_endpoint(
        EndpointName=endpoint_name, Body=json.dumps(payload), ContentType="application/json"
    )

    output = json.loads(response["Body"].read().decode())

    result = output["generated_text"]
    return result

# please insert your long prompt/document content here
prompt = """<s>[INST] What are the main challenges to support long contexts for a Large Language Model? [/INST]"""

#print(prompt)
endpoint_name = "megaBeam-mistral-7b-512k-2024-05-13-14-23-41-219" # please use a valid endpoint name
result = call_endpoint(prompt, endpoint_name)
print(result)

Limitations

Before using the MegaBeam-Mistral-7B-512k model, it is important to perform your own independent assessment, and take measures to ensure that your use would comply with your own specific quality control practices and standards, and that your use would comply with the local rules, laws, regulations, licenses and terms that apply to you, and your content.

The AWS Contributors

Chen Wu, Yin Song, Eden Duthie