---
language:
- en
license: llama2
library_name: transformers
datasets:
- togethercomputer/llama-instruct
model_name: Llama2 7B 32K Instruct
inference: false
model_creator: Together
model_link: https://huggingface.co/togethercomputer/Llama-2-7B-32K-Instruct
model_type: llama
quantized_by: TheBloke
base_model: togethercomputer/Llama-2-7B-32K-Instruct
---
# Llama2 7B 32K Instruct - GGML
- Model creator: [Together](https://huggingface.co/togethercomputer)
- Original model: [Llama2 7B 32K Instruct](https://huggingface.co/togethercomputer/Llama-2-7B-32K-Instruct)
## Description
This repo contains GGML format model files for [Together's Llama2 7B 32K Instruct](https://huggingface.co/togethercomputer/Llama-2-7B-32K-Instruct).
### Important note regarding GGML files.
The GGML format has now been superseded by GGUF. As of August 21st 2023, [llama.cpp](https://github.com/ggerganov/llama.cpp) no longer supports GGML models. Third party clients and libraries are expected to still support it for a time, but many may also drop support.
Please use the GGUF models instead.
### About GGML
GGML files are for CPU + GPU inference using [llama.cpp](https://github.com/ggerganov/llama.cpp) and libraries and UIs which support this format, such as:
* [text-generation-webui](https://github.com/oobabooga/text-generation-webui), the most popular web UI. Supports NVidia CUDA GPU acceleration.
* [KoboldCpp](https://github.com/LostRuins/koboldcpp), a powerful GGML web UI with GPU acceleration on all platforms (CUDA and OpenCL). Especially good for story telling.
* [LM Studio](https://lmstudio.ai/), a fully featured local GUI with GPU acceleration on both Windows (NVidia and AMD), and macOS.
* [LoLLMS Web UI](https://github.com/ParisNeo/lollms-webui), a great web UI with CUDA GPU acceleration via the c_transformers backend.
* [ctransformers](https://github.com/marella/ctransformers), a Python library with GPU accel, LangChain support, and OpenAI-compatible AI server.
* [llama-cpp-python](https://github.com/abetlen/llama-cpp-python), a Python library with GPU accel, LangChain support, and OpenAI-compatible API server.
## Repositories available
* [GPTQ models for GPU inference, with multiple quantisation parameter options.](https://huggingface.co/TheBloke/Llama-2-7B-32K-Instruct-GPTQ)
* [2, 3, 4, 5, 6 and 8-bit GGUF models for CPU+GPU inference](https://huggingface.co/TheBloke/Llama-2-7B-32K-Instruct-GGUF)
* [2, 3, 4, 5, 6 and 8-bit GGML models for CPU+GPU inference (deprecated)](https://huggingface.co/TheBloke/Llama-2-7B-32K-Instruct-GGML)
* [Together's original unquantised fp16 model in pytorch format, for GPU inference and for further conversions](https://huggingface.co/togethercomputer/Llama-2-7B-32K-Instruct)
## Prompt template: Llama2-Instruct-Only
```
[INST]
{prompt}
[\INST]
```
## Compatibility
These quantised GGML files are compatible with llama.cpp between June 6th (commit `2d43387`) and August 21st 2023.
For support with latest llama.cpp, please use GGUF files instead.
The final llama.cpp commit with support for GGML was: [dadbed99e65252d79f81101a392d0d6497b86caa](https://github.com/ggerganov/llama.cpp/commit/dadbed99e65252d79f81101a392d0d6497b86caa)
As of August 23rd 2023 they are still compatible with all UIs, libraries and utilities which use GGML. This may change in the future.
## Explanation of the new k-quant methods
Click to see details
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 |
| ---- | ---- | ---- | ---- | ---- | ----- |
| [llama-2-7b-32k-instruct.ggmlv3.q2_K.bin](https://huggingface.co/TheBloke/Llama-2-7B-32K-Instruct-GGML/blob/main/llama-2-7b-32k-instruct.ggmlv3.q2_K.bin) | q2_K | 2 | 2.87 GB| 5.37 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. |
| [llama-2-7b-32k-instruct.ggmlv3.q3_K_S.bin](https://huggingface.co/TheBloke/Llama-2-7B-32K-Instruct-GGML/blob/main/llama-2-7b-32k-instruct.ggmlv3.q3_K_S.bin) | q3_K_S | 3 | 2.95 GB| 5.45 GB | New k-quant method. Uses GGML_TYPE_Q3_K for all tensors |
| [llama-2-7b-32k-instruct.ggmlv3.q3_K_M.bin](https://huggingface.co/TheBloke/Llama-2-7B-32K-Instruct-GGML/blob/main/llama-2-7b-32k-instruct.ggmlv3.q3_K_M.bin) | q3_K_M | 3 | 3.28 GB| 5.78 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 |
| [llama-2-7b-32k-instruct.ggmlv3.q3_K_L.bin](https://huggingface.co/TheBloke/Llama-2-7B-32K-Instruct-GGML/blob/main/llama-2-7b-32k-instruct.ggmlv3.q3_K_L.bin) | q3_K_L | 3 | 3.60 GB| 6.10 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 |
| [llama-2-7b-32k-instruct.ggmlv3.q4_0.bin](https://huggingface.co/TheBloke/Llama-2-7B-32K-Instruct-GGML/blob/main/llama-2-7b-32k-instruct.ggmlv3.q4_0.bin) | q4_0 | 4 | 3.83 GB| 6.33 GB | Original quant method, 4-bit. |
| [llama-2-7b-32k-instruct.ggmlv3.q4_K_S.bin](https://huggingface.co/TheBloke/Llama-2-7B-32K-Instruct-GGML/blob/main/llama-2-7b-32k-instruct.ggmlv3.q4_K_S.bin) | q4_K_S | 4 | 3.83 GB| 6.33 GB | New k-quant method. Uses GGML_TYPE_Q4_K for all tensors |
| [llama-2-7b-32k-instruct.ggmlv3.q4_K_M.bin](https://huggingface.co/TheBloke/Llama-2-7B-32K-Instruct-GGML/blob/main/llama-2-7b-32k-instruct.ggmlv3.q4_K_M.bin) | q4_K_M | 4 | 4.08 GB| 6.58 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 |
| [llama-2-7b-32k-instruct.ggmlv3.q4_1.bin](https://huggingface.co/TheBloke/Llama-2-7B-32K-Instruct-GGML/blob/main/llama-2-7b-32k-instruct.ggmlv3.q4_1.bin) | q4_1 | 4 | 4.24 GB| 6.74 GB | Original quant method, 4-bit. Higher accuracy than q4_0 but not as high as q5_0. However has quicker inference than q5 models. |
| [llama-2-7b-32k-instruct.ggmlv3.q5_0.bin](https://huggingface.co/TheBloke/Llama-2-7B-32K-Instruct-GGML/blob/main/llama-2-7b-32k-instruct.ggmlv3.q5_0.bin) | q5_0 | 5 | 4.65 GB| 7.15 GB | Original quant method, 5-bit. Higher accuracy, higher resource usage and slower inference. |
| [llama-2-7b-32k-instruct.ggmlv3.q5_K_S.bin](https://huggingface.co/TheBloke/Llama-2-7B-32K-Instruct-GGML/blob/main/llama-2-7b-32k-instruct.ggmlv3.q5_K_S.bin) | q5_K_S | 5 | 4.65 GB| 7.15 GB | New k-quant method. Uses GGML_TYPE_Q5_K for all tensors |
| [llama-2-7b-32k-instruct.ggmlv3.q5_K_M.bin](https://huggingface.co/TheBloke/Llama-2-7B-32K-Instruct-GGML/blob/main/llama-2-7b-32k-instruct.ggmlv3.q5_K_M.bin) | q5_K_M | 5 | 4.78 GB| 7.28 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 |
| [llama-2-7b-32k-instruct.ggmlv3.q5_1.bin](https://huggingface.co/TheBloke/Llama-2-7B-32K-Instruct-GGML/blob/main/llama-2-7b-32k-instruct.ggmlv3.q5_1.bin) | q5_1 | 5 | 5.06 GB| 7.56 GB | Original quant method, 5-bit. Even higher accuracy, resource usage and slower inference. |
| [llama-2-7b-32k-instruct.ggmlv3.q6_K.bin](https://huggingface.co/TheBloke/Llama-2-7B-32K-Instruct-GGML/blob/main/llama-2-7b-32k-instruct.ggmlv3.q6_K.bin) | q6_K | 6 | 5.53 GB| 8.03 GB | New k-quant method. Uses GGML_TYPE_Q8_K for all tensors - 6-bit quantization |
| [llama-2-7b-32k-instruct.ggmlv3.q8_0.bin](https://huggingface.co/TheBloke/Llama-2-7B-32K-Instruct-GGML/blob/main/llama-2-7b-32k-instruct.ggmlv3.q8_0.bin) | q8_0 | 8 | 7.13 GB| 9.63 GB | Original quant method, 8-bit. Almost indistinguishable from float16. High resource use and slow. Not recommended for most users. |
**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 `llama.cpp`
Make sure you are using `llama.cpp` from commit [dadbed99e65252d79f81101a392d0d6497b86caa](https://github.com/ggerganov/llama.cpp/commit/dadbed99e65252d79f81101a392d0d6497b86caa) or earlier.
For compatibility with latest llama.cpp, please use GGUF files instead.
```
./main -t 10 -ngl 32 -m llama-2-7b-32k-instruct.ggmlv3.q4_K_M.bin --color -c 2048 --temp 0.7 --repeat_penalty 1.1 -n -1 -p "[INST]\n{prompt}\n[\INST]"
```
Change `-t 10` to the number of physical CPU cores you have. For example if your system has 8 cores/16 threads, use `-t 8`.
Change `-ngl 32` to the number of layers to offload to GPU. Remove it if you don't have GPU acceleration.
Change `-c 2048` to the desired sequence length for this model. For example, `-c 4096` for a Llama 2 model. For models that use RoPE, add `--rope-freq-base 10000 --rope-freq-scale 0.5` for doubled context, or `--rope-freq-base 10000 --rope-freq-scale 0.25` for 4x context.
If you want to have a chat-style conversation, replace the `-p ` argument with `-i -ins`
For other parameters and how to use them, please refer to [the llama.cpp documentation](https://github.com/ggerganov/llama.cpp/blob/master/examples/main/README.md)
## How to run in `text-generation-webui`
Further instructions here: [text-generation-webui/docs/llama.cpp.md](https://github.com/oobabooga/text-generation-webui/blob/main/docs/llama.cpp.md).
## Discord
For further support, and discussions on these models and AI in general, join us at:
[TheBloke AI's Discord server](https://discord.gg/theblokeai)
## Thanks, and how to contribute.
Thanks to the [chirper.ai](https://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: https://patreon.com/TheBlokeAI
* Ko-Fi: https://ko-fi.com/TheBlokeAI
**Special thanks to**: Aemon Algiz.
**Patreon special mentions**: Russ Johnson, J, alfie_i, Alex, NimbleBox.ai, Chadd, Mandus, Nikolai Manek, Ken Nordquist, ya boyyy, Illia Dulskyi, Viktor Bowallius, vamX, Iucharbius, zynix, Magnesian, Clay Pascal, Pierre Kircher, Enrico Ros, Tony Hughes, Elle, Andrey, knownsqashed, Deep Realms, Jerry Meng, Lone Striker, Derek Yates, Pyrater, Mesiah Bishop, James Bentley, Femi Adebogun, Brandon Frisco, SuperWojo, Alps Aficionado, Michael Dempsey, Vitor Caleffi, Will Dee, Edmond Seymore, usrbinkat, LangChain4j, Kacper Wikieł, Luke Pendergrass, John Detwiler, theTransient, Nathan LeClaire, Tiffany J. Kim, biorpg, Eugene Pentland, Stanislav Ovsiannikov, Fred von Graf, terasurfer, Kalila, Dan Guido, Nitin Borwankar, 阿明, Ai Maven, John Villwock, Gabriel Puliatti, Stephen Murray, Asp the Wyvern, danny, Chris Smitley, ReadyPlayerEmma, S_X, Daniel P. Andersen, Olakabola, Jeffrey Morgan, Imad Khwaja, Caitlyn Gatomon, webtim, Alicia Loh, Trenton Dambrowitz, Swaroop Kallakuri, Erik Bjäreholt, Leonard Tan, Spiking Neurons AB, Luke @flexchar, Ajan Kanaga, Thomas Belote, Deo Leter, RoA, Willem Michiel, transmissions 11, subjectnull, Matthew Berman, Joseph William Delisle, David Ziegler, Michael Davis, Johann-Peter Hartmann, Talal Aujan, senxiiz, Artur Olbinski, Rainer Wilmers, Spencer Kim, Fen Risland, Cap'n Zoog, Rishabh Srivastava, Michael Levine, Geoffrey Montalvo, Sean Connelly, Alexandros Triantafyllidis, Pieter, Gabriel Tamborski, Sam, Subspace Studios, Junyu Yang, Pedro Madruga, Vadim, Cory Kujawski, K, Raven Klaugh, Randy H, Mano Prime, Sebastain Graf, Space Cruiser
Thank you to all my generous patrons and donaters!
And thank you again to a16z for their generous grant.
# Original model card: Together's Llama2 7B 32K Instruct
# Llama-2-7B-32K-Instruct
## Model Description
Llama-2-7B-32K-Instruct is an open-source, long-context chat model finetuned from [Llama-2-7B-32K](https://huggingface.co/togethercomputer/Llama-2-7B-32K), over high-quality instruction and chat data.
We built Llama-2-7B-32K-Instruct with less than 200 lines of Python script using [Together API](https://together.ai/blog/api-announcement), and we also make the [recipe fully available](https://github.com/togethercomputer/Llama-2-7B-32K-Instruct).
We hope that this can enable everyone to finetune their own version of [Llama-2-7B-32K](https://huggingface.co/togethercomputer/Llama-2-7B-32K) — play with [Together API](https://together.ai/blog/api-announcement) and give us feedback!
## Data Collection Details
Llama-2-7B-32K-Instruct is fine-tuned over a combination of two parts:
1. **19K single- and multi-round conversations generated by human instructions and [Llama-2-70B-Chat](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf) outputs**.
We collected the dataset following the distillation paradigm that is used by Alpaca, Vicuna, WizardLM, Orca — producing instructions by querying a powerful LLM (in this case, [Llama-2-70B-Chat](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf)).
The complete dataset is also released [here](https://huggingface.co/datasets/togethercomputer/llama-instruct).
We also share the complete recipe for the data collection process [here](https://github.com/togethercomputer/Llama-2-7B-32K-Instruct).
2. **Long-context Summarization and Long-context QA**.
We follow the recipe of [Llama-2-7B-32K](https://together.ai/blog/Llama-2-7B-32K), and train our model with the [BookSum dataset](https://huggingface.co/datasets/togethercomputer/Long-Data-Collections) and [Multi-document Question Answering](https://arxiv.org/abs/2307.03172).
The final data mixture used for model finetuning is: 19K instruction (50%) + BookSum (25%) + MQA (25%).
## Model Usage
We encourage you to try out this model using the [Together API](https://together.ai/blog/api-announcement). The updated inference stack allows for efficient inference.
To run the model locally, we strongly recommend to install Flash Attention V2, which is necessary to obtain the best performance:
```
# Please update the path of `CUDA_HOME`
export CUDA_HOME=/usr/local/cuda-11.8
pip install transformers==4.31.0
pip install sentencepiece
pip install ninja
pip install flash-attn --no-build-isolation
pip install git+https://github.com/HazyResearch/flash-attention.git#subdirectory=csrc/rotary
```
You can load the model directly from the Hugging Face model hub using
```python
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("togethercomputer/Llama-2-7B-32K-Instruct")
model = AutoModelForCausalLM.from_pretrained("togethercomputer/Llama-2-7B-32K-Instruct",
trust_remote_code=True, torch_dtype=torch.float16)
input_ids = tokenizer.encode("[INST]\nWrite a poem about cats\n[/INST]\n\n", return_tensors="pt")
output = model.generate(input_ids, max_length=128,
temperature=0.7, repetition_penalty=1.1, top_p=0.7, top_k=50)
output_text = tokenizer.decode(output[0], skip_special_tokens=True)
```
The model is also hosted on [Together Playground](https://api.together.xyz/playground). You can simply play with the model by using prompt formatted by:
```
[INST]\n\n[\INST]\n\n
```
For example, if we query the model with
```
[INST]\nWrite a poem about cats\n[\INST]\n\n
```
the model will return
```
[INST]
Write a poem about cats
[/INST]
Cats are mysterious creatures, with their aloof and independent nature. They're also incredibly beautiful, with their sleek fur and piercing eyes. Here's a poem that captures the essence of cats:
Cats, oh cats, how can I describe you?
Your beauty is beyond compare, it seems.
You're graceful and elegant, like a ballerina's dance,
But don't let your charm fool you, for you're not easily tamed.
With your soft purring and playful meows,
You draw us in with your enchanting powers.
We love to watch you play, your tail twirling 'round,
As if you're dancing on air, with no sound.
But don't be fooled by your sweetness, my friend,
For beneath that gentle exterior, lies a fierce defender.
When danger lurks, you'll spring into action,
Protecting those you hold dear, without question.
Solet us admire you, from afar,
For in your own way, you're truly unique, a star.
And though we may never fully understand,
The depths of your soul, we'll always stand, hand in paw, as one.
This poem captures the essence of cats, highlighting their beauty, independence,and protective nature. It also celebrates the special bond between humans and cats, recognizing their unique qualities and the joy they bring to our lives.
```
## Model Evaluation
We evaluate the model from three aspects: 1) [Alpaca Eval](https://tatsu-lab.github.io/alpaca_eval/);
2) [Rouge score over BookSum](https://together.ai/blog/Llama-2-7B-32K); and
3) [Accuracy over Multi-document Question Answering (MQA)](https://together.ai/blog/Llama-2-7B-32K).
We compare with models including
[GPT-3.5-Turbo-16K](https://platform.openai.com/docs/models/gpt-3-5),
[https://huggingface.co/meta-llama/Llama-2-7b-chat-hf](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf),
[Longchat-7b-16k](https://huggingface.co/lmsys/longchat-7b-16k)
and [Longchat-7b-v1.5-32k](https://huggingface.co/lmsys/longchat-7b-v1.5-32k).
We summarize the results below:
* Alpaca Eval
| Model | win_rate | standard_error | n_total | avg_length |
| -------- | ------- | ------- | ------- | ------- |
| Llama-2-7B-Chat-hf | 71.37 | 1.59 | 805 | 1479 |
| Llama-2-7B-32K-Instruct | 70.36 | 1.61 | 803 | 1885 |
| oasst-rlhf-llama-33b | 66.52 | 1.66 | 805 | 1079 |
| text_davinci_003 | 50.00 | 0.00 | 805 | 307|
| falcon-40b-instruct | 45.71 | 1.75 | 805 | 662 |
| alpaca-farm-ppo-human | 41.24 | 1.73 | 805 | 803 |
| alpaca-7b | 26.46 | 1.54 | 805 | 396 |
| text_davinci_001 | 15.17 | 1.24 | 804 | 296 |
* Rouge Score over BookSum
| Model | R1 | R2 | RL |
| -------- | ------- | ------- | ------- |
| Llama-2-7B-Chat-hf | 0.055 | 0.008 | 0.046 |
| Longchat-7b-16k | 0.303 | 0.055 | 0.160 |
| Longchat-7b-v1.5-32k | 0.308 | 0.057 | 0.163 |
| GPT-3.5-Turbo-16K | 0.324 | 0.066 | 0.178 |
| Llama-2-7B-32K-Instruct (ours) | 0.336 | 0.076 | 0.184 |
* Accuracy over MQA
| Model | 20 docs (Avg 2.9K tokens) | 30 docs (Avg 4.4K tokens) | 50 docs (Avg 7.4K tokens) |
| -------- | ------- | ------- | ------- |
| Llama-2-7B-Chat-hf | 0.448 | 0.421 | 0.354 |
| Longchat-7b-16k | 0.510 | 0.473 | 0.428 |
| Longchat-7b-v1.5-32k | 0.534 | 0.516 | 0.479 |
| GPT-3.5-Turbo-16K | 0.622 | 0.609 | 0.577 |
| Llama-2-7B-32K-Instruct (ours) | 0.622 | 0.604 | 0.589 |
## Limitations and Bias
As with all language models, Llama-2-7B-32K-Instruct may generate incorrect or biased content. It's important to keep this in mind when using the model.
## Community
Join us on [Together Discord](https://discord.gg/6ZVDU8tTD4)