Jambatypus-v0.1

This model is a QLoRA fine-tuned version of ai21labs/Jamba-v0.1 on the chargoddard/Open-Platypus-Chat dataset.

It has been trained on 2xA100 80 GB using my LazyAxolotl - Jamba notebook.

This repo contains both the adapter and the merged model in FP16 precision.

I recommend using the ChatML template to use this model.

See axolotl config

axolotl version: 0.4.0

base_model: ai21labs/Jamba-v0.1
trust_remote_code: true

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: chargoddard/Open-Platypus-Chat
    type: sharegpt
chat_template: chatml
dataset_prepared_path:
val_set_size: 0.01
output_dir: ./out

sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true
eval_sample_packing: false

use_wandb: true
wandb_project: axolotl
wandb_entity:
wandb_watch:
wandb_name: Jambatypus-v0.1
wandb_log_model:

adapter: qlora
lora_r: 16
lora_alpha: 32
lora_dropout: 0.05
lora_target_linear: true

low_cpu_mem_usage: true
gradient_accumulation_steps: 8
micro_batch_size: 1
num_epochs: 1
optimizer: adamw_bnb_8bit
adam_beta2: 0.95
adam_epsilon: 0.00001
max_grad_norm: 1.0
lr_scheduler: cosine
learning_rate: 0.0002

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 10
evals_per_epoch: 4
saves_per_epoch: 4
save_total_limit: 2
debug:
deepspeed:
weight_decay: 0.0
special_tokens:

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0002
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 8
• total_train_batch_size: 16
• total_eval_batch_size: 2
• optimizer: Adam with betas=(0.9,0.95) and epsilon=1e-05
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss
0.6274	0.01	1	1.0298
0.44	0.25	42	0.9770
0.4406	0.5	84	0.9653
0.4445	0.75	126	0.9645
0.4609	1.0	168	0.9641

Framework versions

• PEFT 0.10.0
• Transformers 4.40.0.dev0
• Pytorch 2.1.2+cu118
• Datasets 2.18.0
• Tokenizers 0.15.0

💻 Usage

The following code creates a Gradio chat interface with Jambatypus.

!pip install -qqq -U git+https://github.com/huggingface/transformers
!pip install -qqq mamba-ssm causal-conv1d>=1.2.0
!pip install -qqq accelerate bitsandbytes torch datasets peft gradio
!pip install -qqq flash-attn --no-build-isolation


import torch
import gradio as gr
from threading import Thread
from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, TextIteratorStreamer

STOP_TOKEN = "<|im_end|>"

def predict(message, history, system_prompt, temperature, max_new_tokens, top_k, repetition_penalty, top_p):
    # Format history with a given chat template
    stop_token = "<|im_end|>"
    instruction = '<|im_start|>system\n' + system_prompt + '\n<|im_end|>\n'
    for human, assistant in history:
        instruction += '<|im_start|>user\n' + human + '\n<|im_end|>\n<|im_start|>assistant\n' + assistant
    instruction += '\n<|im_start|>user\n' + message + '\n<|im_end|>\n<|im_start|>assistant\n'
    
    streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
    enc = tokenizer([instruction], return_tensors="pt", padding=True, truncation=True)
    input_ids, attention_mask = enc.input_ids, enc.attention_mask

    generate_kwargs = dict(
        {"input_ids": input_ids.to(device), "attention_mask": attention_mask.to(device)},
        streamer=streamer,
        do_sample=True,
        temperature=temperature,
        max_new_tokens=max_new_tokens,
        top_k=top_k,
        repetition_penalty=repetition_penalty,
        top_p=top_p
    )
    t = Thread(target=model.generate, kwargs=generate_kwargs)
    t.start()
    outputs = []
    for new_token in streamer:
        if STOP_TOKEN in new_token:
            outputs.append(new_token[:-len(stop_token)-1])
            yield "".join(outputs)
            break
        outputs.append(new_token)
        yield "".join(outputs)


# Load model
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
tokenizer = AutoTokenizer.from_pretrained("ai21labs/Jamba-v0.1")

# 4-bit precision quant config
quantization_config = BitsAndBytesConfig(
    load_in_8bit=True,
    llm_int8_skip_modules=["mamba"]
)
# Load model and tokenizer with ChatML format
model = AutoModelForCausalLM.from_pretrained(
    "ai21labs/Jamba-v0.1",
    trust_remote_code=True,
    torch_dtype=torch.cuda.is_bf16_supported() and torch.bfloat16 or torch.float16,
    attn_implementation="flash_attention_2",
    low_cpu_mem_usage=True,
    quantization_config=quantization_config
)
config = PeftConfig.from_pretrained("mlabonne/Jambatypus-v0.1")
model = PeftModel.from_pretrained(model, "mlabonne/Jambatypus-v0.1")

# Create Gradio interface
gr.ChatInterface(
    predict,
    title="Jambatypus",
    description="Chat with Jambatypus!",
    examples=[
        ["Can you solve the equation 2x + 3 = 11 for x?"],
        ["Write an epic poem about Ancient Rome."],
        ["Who was the first person to walk on the Moon?"],
        ["Use a list comprehension to create a list of squares for numbers from 1 to 10."],
        ["Recommend some popular science fiction books."],
        ["Can you write a short story about a time-traveling detective?"]
    ],
    additional_inputs_accordion=gr.Accordion(label="⚙️ Parameters", open=False),
    additional_inputs=[
        gr.Textbox("Perform the task to the best of your ability.", label="System prompt"),
        gr.Slider(0, 1, 0.8, label="Temperature"),
        gr.Slider(128, 4096, 1024, label="Max new tokens"),
        gr.Slider(1, 80, 40, label="Top K sampling"),
        gr.Slider(0, 2, 1.1, label="Repetition penalty"),
        gr.Slider(0, 1, 0.95, label="Top P sampling"),
    ],
    theme=gr.themes.Soft(primary_hue="green"),
).queue().launch(share=True)