Buddhi-128K-Chat (7B) vLLM Inference:

Read release article: 🔗 Introducing Buddhi: Open-Source Chat Model with a 128K Context Window 🔗

Model Description

Buddhi-128k-Chat is a general-purpose first chat model with 128K context length window. It is meticulously fine-tuned on the Mistral 7B Instruct, and optimised to handle an extended context length of up to 128,000 tokens using the innovative YaRN (Yet another Rope Extension) Technique. This enhancement allows Buddhi to maintain a deeper understanding of context in long documents or conversations, making it particularly adept at tasks requiring extensive context retention, such as comprehensive document summarization, detailed narrative generation, and intricate question-answering.

Architecture

The Buddhi-128K-Chat model is fine-tuned on the Mistral-7B Instruct base model. We selected the Mistral 7B Instruct v0.2 as the parent model due to its superior reasoning capabilities. The architecture of the Mistral-7B model includes features like Grouped-Query Attention and Byte-fallback BPE tokenizer. Originally, this model has 32,768 maximum position embeddings. To increase the context size to 128K, we needed to modify the positional embeddings, which is where YaRN comes into play.

In our approach, we utilized the NTK-aware technique, which recommends alternative interpolation techniques for positional interpolation. One experimentation involved Dynamic-YARN, suggesting the dynamic value of the 's' scale factor. This is because during inference, the sequence length changes by 1 after every word prediction. By integrating these position embeddings with the Mistral-7B Instruct base model, we achieved the 128K model.

Additionally, we fine-tuned the model on our dataset to contribute one of the very few 128K chat-based models available in the open-source community with greater reasoning capabilities than all of it.

Hardware requirements:

For 128k Context Length

• 80GB VRAM - A100 Preferred

For 32k Context Length

• 40GB VRAM - A100 Preferred

vLLM - For Faster Inference

Installation

!pip install vllm
!pip install flash_attn # If Flash Attention 2 is supported by your System

Please check out Flash Attention 2 Github Repository for more instructions on how to Install it.

Implementation:

Note: The actual hardware requirements to run the model is roughly around 70GB VRAM. For experimentation, we are limiting the context length to 75K instead of 128K. This make it suitable for testing the model in 30-35 GB VRAM

from vllm import LLM, SamplingParams

llm = LLM(
    model='aiplanet/buddhi-128k-chat-7b',
    trust_remote_code=True,
    dtype = 'bfloat16',
    gpu_memory_utilization=1,
    max_model_len= 75000 
)

prompts = [
  """<s> [INST] Please tell me a joke. [/INST] """,
  """<s> [INST] What is Machine Learning? [/INST] """
]

sampling_params = SamplingParams(
  temperature=0.8,
  top_p=0.95,
  max_tokens=1000
)

outputs = llm.generate(prompts, sampling_params)

for output in outputs:
    prompt = output.prompt
    generated_text = output.outputs[0].text
    print(generated_text)
    print("\n\n")

# we have also attached a colab notebook, that contains: 2 more experimentations: Long Essay and Entire Book

For Output, do check out the colab notebook:

Transformers - Basic Implementation

import torch
import transformers
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16
)

model_name = "aiplanet/Buddhi-128K-Chat"

model = AutoModelForCausalLM.from_pretrained(
  model_name,
  quantization_config=bnb_config,
  device_map="sequential",
  trust_remote_code=True
)

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

prompt = "<s> [INST] Please tell me a small joke. [/INST] "

tokens = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(
  **tokens,
  max_new_tokens=100,
  do_sample=True,
  top_p=0.95,
  temperature=0.8,
)

decoded_output = tokenizer.batch_decode(outputs.detach().cpu().numpy(), skip_special_tokens=True)[0]
print(f"Output:\n{decoded_output[len(prompt):]}")

Output

Output:
Why don't scientists trust atoms?

Because they make up everything.

Prompt Template for Buddi-128-Chat

In order to leverage instruction fine-tuning, your prompt should be surrounded by [INST] and [/INST] tokens. The very first instruction should begin with a begin of sentence id. The next instructions should not. The assistant generation will be ended by the end-of-sentence token id.

"<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]"

Benchmarks

Long Context Benchmark

LongICLBench Banking77

Short Context Benchmark

Get in Touch

You can schedule a 1:1 meeting with our DevRel & Community Team to get started with AI Planet Open Source LLMs and GenAI Stack. Schedule the call here: https://calendly.com/jaintarun

Stay tuned for more updates and be a part of the coding evolution. Join us on this exciting journey as we make AI accessible to all at AI Planet!

Framework versions

• Transformers 4.39.2
• Pytorch 2.2.1+cu121
• Datasets 2.18.0
• Accelerate 0.27.2
• flash_attn 2.5.6

Citation

@misc {Chaitanya890, lucifertrj ,
   author       = { Chaitanya Singhal, Tarun Jain },
   title        = { Buddhi-128k-Chat by AI Planet},
   year         = 2024,
   url          = { https://huggingface.co/aiplanet//Buddhi-128K-Chat },
   publisher    = { Hugging Face }
}