<!-- header start --> <div style="width: 100%;"> Cloned from TheBloke repo </div> <!-- header end -->

June Lee's Wizard Vicuna 13B fp16

This is fp16 pytorch format model files for June Lee's Wizard Vicuna 13B merged with Kaio Ken's SuperHOT 8K.

Kaio Ken's SuperHOT 13b LoRA is merged on to the base model, and then 8K context can be achieved during inference by using trust_remote_code=True.

Note that config.json has been set to a sequence length of 8192. This can be modified to 4096 if you want to try with a smaller sequence length.

Repositories available

• 4-bit GPTQ models for GPU inference
• 2, 3, 4, 5, 6 and 8-bit GGML models for CPU inference
• Unquantised SuperHOT fp16 model in pytorch format, for GPU inference and for further conversions
• Unquantised base fp16 model in pytorch format, for GPU inference and for further conversions

How to use this model from Python code

First make sure you have Einops installed:

pip3 install auto-gptq

Then run the following code. config.json has been default to a sequence length of 8192, but you can also configure this in your Python code.

The provided modelling code, activated with trust_remote_code=True will automatically set the scale parameter from the configured max_position_embeddings. Eg for 8192, scale is set to 4.

from transformers import AutoConfig, AutoTokenizer, AutoModelForCausalLM, pipeline
import argparse

model_name_or_path = "gsaivinay/wizard-vicuna-13B-SuperHOT-8K-fp16"

use_triton = False

tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, use_fast=True)

config = AutoConfig.from_pretrained(model_name_or_path, trust_remote_code=True)
# Change this to the sequence length you want
config.max_position_embeddings = 8192

model = AutoModelForCausalLM.from_pretrained(model_name_or_path,
        config=config,
        trust_remote_code=True,
        device_map='auto')

# Note: check to confirm if this is correct prompt template is correct for this model!
prompt = "Tell me about AI"
prompt_template=f'''USER: {prompt}
ASSISTANT:'''

print("\n\n*** Generate:")

input_ids = tokenizer(prompt_template, return_tensors='pt').input_ids.cuda()
output = model.generate(inputs=input_ids, temperature=0.7, max_new_tokens=512)
print(tokenizer.decode(output[0]))

# Inference can also be done using transformers' pipeline

print("*** Pipeline:")
pipe = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    max_new_tokens=512,
    temperature=0.7,
    top_p=0.95,
    repetition_penalty=1.15
)

print(pipe(prompt_template)[0]['generated_text'])

Using other UIs: monkey patch

Provided in the repo is llama_rope_scaled_monkey_patch.py, written by @kaiokendev.

It can be theoretically be added to any Python UI or custom code to enable the same result as trust_remote_code=True. I have not tested this, and it should be superseded by using trust_remote_code=True, but I include it for completeness and for interest.

Original model card: Kaio Ken's SuperHOT 8K

SuperHOT Prototype 2 w/ 8K Context

This is a second prototype of SuperHOT, this time 30B with 8K context and no RLHF, using the same technique described in the github blog. Tests have shown that the model does indeed leverage the extended context at 8K.

You will need to use either the monkeypatch or, if you are already using the monkeypatch, change the scaling factor to 0.25 and the maximum sequence length to 8192

Looking for Merged & Quantized Models?

• 30B 4-bit CUDA: tmpupload/superhot-30b-8k-4bit-safetensors
• 30B 4-bit CUDA 128g: tmpupload/superhot-30b-8k-4bit-128g-safetensors

Training Details

I trained the LoRA with the following configuration:

• 1200 samples (~400 samples over 2048 sequence length)
  • learning rate of 3e-4
  • 3 epochs
  • The exported modules are:
  • q_proj
  • k_proj
  • v_proj
  • o_proj
  • no bias
  • Rank = 4
  • Alpha = 8
  • no dropout
  • weight decay of 0.1
  • AdamW beta1 of 0.9 and beta2 0.99, epsilon of 1e-5
  • Trained on 4-bit base model

Original model card: June Lee's Wizard Vicuna 13B

Wizard-Vicuna-13B-HF

This is a float16 HF format repo for junelee's wizard-vicuna 13B.

June Lee's repo was also HF format. The reason I've made this is that the original repo was in float32, meaning it required 52GB disk space, VRAM and RAM.

This model was converted to float16 to make it easier to load and manage.

Repositories available

• 4bit GPTQ models for GPU inference.
• 4bit and 5bit GGML models for CPU inference.
• float16 HF format model for GPU inference.

Original WizardVicuna-13B model card

Github page: https://github.com/melodysdreamj/WizardVicunaLM

WizardVicunaLM

Wizard's dataset + ChatGPT's conversation extension + Vicuna's tuning method

I am a big fan of the ideas behind WizardLM and VicunaLM. I particularly like the idea of WizardLM handling the dataset itself more deeply and broadly, as well as VicunaLM overcoming the limitations of single-turn conversations by introducing multi-round conversations. As a result, I combined these two ideas to create WizardVicunaLM. This project is highly experimental and designed for proof of concept, not for actual usage.

Benchmark

Approximately 7% performance improvement over VicunaLM

Detail

The questions presented here are not from rigorous tests, but rather, I asked a few questions and requested GPT-4 to score them. The models compared were ChatGPT 3.5, WizardVicunaLM, VicunaLM, and WizardLM, in that order.

Principle

We adopted the approach of WizardLM, which is to extend a single problem more in-depth. However, instead of using individual instructions, we expanded it using Vicuna's conversation format and applied Vicuna's fine-tuning techniques.

Turning a single command into a rich conversation is what we've done here.

After creating the training data, I later trained it according to the Vicuna v1.1 training method.

Detailed Method

First, we explore and expand various areas in the same topic using the 7K conversations created by WizardLM. However, we made it in a continuous conversation format instead of the instruction format. That is, it starts with WizardLM's instruction, and then expands into various areas in one conversation using ChatGPT 3.5.

After that, we applied the following model using Vicuna's fine-tuning format.

Training Process

Trained with 8 A100 GPUs for 35 hours.

Weights

You can see the dataset we used for training and the 13b model in the huggingface.

Conclusion

If we extend the conversation to gpt4 32K, we can expect a dramatic improvement, as we can generate 8x more, more accurate and richer conversations.

License

The model is licensed under the LLaMA model, and the dataset is licensed under the terms of OpenAI because it uses ChatGPT. Everything else is free.

Author

JUNE LEE - He is active in Songdo Artificial Intelligence Study and GDG Songdo.