TheBloke's picture
Initial GGML model commit
f16b3a8
metadata
inference: false
license: other
TheBlokeAI

June Lee's Wizard Vicuna 13B GGML

These files are GGML format model files for June Lee's Wizard Vicuna 13B.

These are SuperHOT GGMLs with an increased context length. SuperHOT is a new system that employs RoPE to expand context beyond what was originally possible for a model. It was discovered and developed by kaiokendev.

In order to use the increased context length, you can presently use:

Support is also expected to come to llama.cpp, however it is still being worked on and there is currently no ETA for that.

To use the increased context with KoboldCpp and (when supported) llama.cpp, simply use --contextsize to set the desired context, eg --contextsize 4096 or --contextsize 8192.

Repositories available

Compatibility

These GGMLs will work with any llama.cpp-compatible GGML client that supports k-quants.

However the increased context length won't work without specific support. See the note in the introduction for details on using increased context.

Explanation of the new k-quant methods

The new methods available are:

  • GGML_TYPE_Q2_K - "type-1" 2-bit quantization in super-blocks containing 16 blocks, each block having 16 weight. Block scales and mins are quantized with 4 bits. This ends up effectively using 2.5625 bits per weight (bpw)
  • GGML_TYPE_Q3_K - "type-0" 3-bit quantization in super-blocks containing 16 blocks, each block having 16 weights. Scales are quantized with 6 bits. This end up using 3.4375 bpw.
  • GGML_TYPE_Q4_K - "type-1" 4-bit quantization in super-blocks containing 8 blocks, each block having 32 weights. Scales and mins are quantized with 6 bits. This ends up using 4.5 bpw.
  • GGML_TYPE_Q5_K - "type-1" 5-bit quantization. Same super-block structure as GGML_TYPE_Q4_K resulting in 5.5 bpw
  • GGML_TYPE_Q6_K - "type-0" 6-bit quantization. Super-blocks with 16 blocks, each block having 16 weights. Scales are quantized with 8 bits. This ends up using 6.5625 bpw
  • GGML_TYPE_Q8_K - "type-0" 8-bit quantization. Only used for quantizing intermediate results. The difference to the existing Q8_0 is that the block size is 256. All 2-6 bit dot products are implemented for this quantization type.

Refer to the Provided Files table below to see what files use which methods, and how.

Provided files

Name Quant method Bits Size Max RAM required Use case
wizard-vicuna-13b-superhot-8k.ggmlv3.q2_K.bin q2_K 2 5.51 GB 8.01 GB New k-quant method. Uses GGML_TYPE_Q4_K for the attention.vw and feed_forward.w2 tensors, GGML_TYPE_Q2_K for the other tensors.
wizard-vicuna-13b-superhot-8k.ggmlv3.q3_K_L.bin q3_K_L 3 6.93 GB 9.43 GB New k-quant method. Uses GGML_TYPE_Q5_K for the attention.wv, attention.wo, and feed_forward.w2 tensors, else GGML_TYPE_Q3_K
wizard-vicuna-13b-superhot-8k.ggmlv3.q3_K_M.bin q3_K_M 3 6.31 GB 8.81 GB New k-quant method. Uses GGML_TYPE_Q4_K for the attention.wv, attention.wo, and feed_forward.w2 tensors, else GGML_TYPE_Q3_K
wizard-vicuna-13b-superhot-8k.ggmlv3.q3_K_S.bin q3_K_S 3 5.66 GB 8.16 GB New k-quant method. Uses GGML_TYPE_Q3_K for all tensors
wizard-vicuna-13b-superhot-8k.ggmlv3.q4_K_M.bin q4_K_M 4 7.87 GB 10.37 GB New k-quant method. Uses GGML_TYPE_Q6_K for half of the attention.wv and feed_forward.w2 tensors, else GGML_TYPE_Q4_K
wizard-vicuna-13b-superhot-8k.ggmlv3.q4_K_S.bin q4_K_S 4 7.37 GB 9.87 GB New k-quant method. Uses GGML_TYPE_Q4_K for all tensors
wizard-vicuna-13b-superhot-8k.ggmlv3.q5_K_M.bin q5_K_M 5 9.23 GB 11.73 GB New k-quant method. Uses GGML_TYPE_Q6_K for half of the attention.wv and feed_forward.w2 tensors, else GGML_TYPE_Q5_K
wizard-vicuna-13b-superhot-8k.ggmlv3.q5_K_S.bin q5_K_S 5 8.97 GB 11.47 GB New k-quant method. Uses GGML_TYPE_Q5_K for all tensors
wizard-vicuna-13b-superhot-8k.ggmlv3.q6_K.bin q6_K 6 10.68 GB 13.18 GB New k-quant method. Uses GGML_TYPE_Q8_K - 6-bit quantization - for all tensors

Note: the above RAM figures assume no GPU offloading. If layers are offloaded to the GPU, this will reduce RAM usage and use VRAM instead.

How to run in koboldcpp

On Linux I use the following command line to launch the KoboldCpp UI with OpenCL aceleration and a context size of 4096:

python ./koboldcpp.py --stream --unbantokens --threads 8 --usecublas 100 wizard-vicuna-13b-superhot-8k.ggmlv3.q5_0.bin

Change --gpulayers 100 to the number of layers you want/are able to offload to the GPU. Remove it if you don't have GPU acceleration.

For OpenCL acceleration, change --usecublas to --useclblast 0 0. You may need to change the second 0 to 1 if you have both an iGPU and a discrete GPU.

Discord

For further support, and discussions on these models and AI in general, join us at:

TheBloke AI's Discord server

Thanks, and how to contribute.

Thanks to the chirper.ai team!

I've had a lot of people ask if they can contribute. I enjoy providing models and helping people, and would love to be able to spend even more time doing it, as well as expanding into new projects like fine tuning/training.

If you're able and willing to contribute it will be most gratefully received and will help me to keep providing more models, and to start work on new AI projects.

Donaters will get priority support on any and all AI/LLM/model questions and requests, access to a private Discord room, plus other benefits.

Special thanks to: Luke from CarbonQuill, Aemon Algiz, Dmitriy Samsonov.

Patreon special mentions: zynix , ya boyyy, Trenton Dambrowitz, Imad Khwaja, Alps Aficionado, chris gileta, John Detwiler, Willem Michiel, RoA, Mano Prime, Rainer Wilmers, Fred von Graf, Matthew Berman, Ghost , Nathan LeClaire, Iucharbius , Ai Maven, Illia Dulskyi, Joseph William Delisle, Space Cruiser, Lone Striker, Karl Bernard, Eugene Pentland, Greatston Gnanesh, Jonathan Leane, Randy H, Pierre Kircher, Willian Hasse, Stephen Murray, Alex , terasurfer , Edmond Seymore, Oscar Rangel, Luke Pendergrass, Asp the Wyvern, Junyu Yang, David Flickinger, Luke, Spiking Neurons AB, subjectnull, Pyrater, Nikolai Manek, senxiiz, Ajan Kanaga, Johann-Peter Hartmann, Artur Olbinski, Kevin Schuppel, Derek Yates, Kalila, K, Talal Aujan, Khalefa Al-Ahmad, Gabriel Puliatti, John Villwock, WelcomeToTheClub, Daniel P. Andersen, Preetika Verma, Deep Realms, Fen Risland, trip7s trip, webtim, Sean Connelly, Michael Levine, Chris McCloskey, biorpg, vamX, Viktor Bowallius, Cory Kujawski.

Thank you to all my generous patrons and donaters!

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?

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

TheBlokeAI
# 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

Discord

For further support, and discussions on these models and AI in general, join us at:

TheBloke AI's Discord server

Thanks, and how to contribute.

Thanks to the chirper.ai team!

I've had a lot of people ask if they can contribute. I enjoy providing models and helping people, and would love to be able to spend even more time doing it, as well as expanding into new projects like fine tuning/training.

If you're able and willing to contribute it will be most gratefully received and will help me to keep providing more models, and to start work on new AI projects.

Donaters will get priority support on any and all AI/LLM/model questions and requests, access to a private Discord room, plus other benefits.

Patreon special mentions: Aemon Algiz, Dmitriy Samsonov, Nathan LeClaire, Trenton Dambrowitz, Mano Prime, David Flickinger, vamX, Nikolai Manek, senxiiz, Khalefa Al-Ahmad, Illia Dulskyi, Jonathan Leane, Talal Aujan, V. Lukas, Joseph William Delisle, Pyrater, Oscar Rangel, Lone Striker, Luke Pendergrass, Eugene Pentland, Sebastain Graf, Johann-Peter Hartman.

Thank you to all my generous patrons and donaters!

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.

gpt3.5 wizard-vicuna-13b vicuna-13b wizard-7b link
Q1 95 90 85 88 link
Q2 95 97 90 89 link
Q3 85 90 80 65 link
Q4 90 85 80 75 link
Q5 90 85 80 75 link
Q6 92 85 87 88 link
Q7 95 90 85 92 link
Q8 90 85 75 70 link
Q9 92 85 70 60 link
Q10 90 80 75 85 link
Q11 90 85 75 65 link
Q12 85 90 80 88 link
Q13 90 95 88 85 link
Q14 94 89 90 91 link
Q15 90 85 88 87 link
91 88 82 80

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.