syzymon-long_llama_3b_instruct-gguf

⚠️ ARCHIVED / LEGACY MODEL NOTICE
This repository is part of a legacy collection quantized around 2023. To manage storage quotas and maintain active community projects, some rarely used quantization formats (e.g., Q2_K, Q3_K, Q4_1, Q5_1) have been permanently removed.

Only the most popular and stable formats (Q4_0, Q4_K_M, Q5_K_M, Q6_K, and Q8_0) remain available.

💡 Looking for something modern? If you are starting a new project, we highly recommend using newer architectures (like Llama 3, Mistral, or Qwen) provided by official maintainers or active community members (e.g., Bartowski, TheBloke legacy files, or official organization handles).

⚠️ This repository is no longer actively maintained. Existing files are provided "as is" for archival and legacy hardware purposes.

I'm constantly enhancing these model descriptions to provide you with the most relevant and comprehensive information

long_llama_3b_instruct - GGUF

• Model creator: syzymon
• Original model: long_llama_3b_instruct

OpenLlama is a free reimplementation of the original Llama Model which is licensed under Apache 2 license.

About GGUF format

gguf is the current file format used by the ggml library. A growing list of Software is using it and can therefore use this model. The core project making use of the ggml library is the llama.cpp project by Georgi Gerganov

Quantization variants

There is a bunch of quantized files available to cater to your specific needs. Here's how to choose the best option for you:

Legacy quants

Q4_0, Q4_1, Q5_0, Q5_1 and Q8 are legacy quantization types. Nevertheless, they are fully supported, as there are several circumstances that cause certain model not to be compatible with the modern K-quants.

Note:

Now there's a new option to use K-quants even for previously 'incompatible' models, although this involves some fallback solution that makes them not real K-quants. More details can be found in affected model descriptions. (This mainly refers to Falcon 7b and Starcoder models)

K-quants

K-quants are designed with the idea that different levels of quantization in specific parts of the model can optimize performance, file size, and memory load. So, if possible, use K-quants. With a Q6_K, you'll likely find it challenging to discern a quality difference from the original model - ask your model two times the same question and you may encounter bigger quality differences.

---

Original Model Card:

LongLLaMA: Focused Transformer Training for Context Scaling

<div align="center">

<a href="https://colab.research.google.com/github/CStanKonrad/long_llama/blob/main/long_llama_instruct_colab.ipynb"><img src="https://colab.research.google.com/assets/colab-badge.svg"></a>

</div> <div align="center">

TLDR | Overview | Usage | LongLLaMA performance | Authors | Citation | License | Acknowledgments

</div>

TLDR

This repo contains LongLLaMA-Instruct-3Bv1.1 that is for research purposes only.

LongLLaMA is built upon the foundation of OpenLLaMA and fine-tuned using the Focused Transformer (FoT) method. We release a smaller 3B base variant (not instruction tuned) of the LongLLaMA model on a permissive license (Apache 2.0) and inference code supporting longer contexts on Hugging Face. Our model weights can serve as the drop-in replacement of LLaMA in existing implementations (for short context up to 2048 tokens). Additionally, we provide evaluation results and comparisons against the original OpenLLaMA models. Stay tuned for further updates.

Overview

Base models

Focused Transformer: Contrastive Training for Context Scaling (FoT) presents a simple method for endowing language models with the ability to handle context consisting possibly of millions of tokens while training on significantly shorter input. FoT permits a subset of attention layers to access a memory cache of (key, value) pairs to extend the context length. The distinctive aspect of FoT is its training procedure, drawing from contrastive learning. Specifically, we deliberately expose the memory attention layers to both relevant and irrelevant keys (like negative samples from unrelated documents). This strategy incentivizes the model to differentiate keys connected with semantically diverse values, thereby enhancing their structure. This, in turn, makes it possible to extrapolate the effective context length much beyond what is seen in training.

LongLLaMA is an OpenLLaMA model finetuned with the FoT method, with three layers used for context extension. Crucially, LongLLaMA is able to extrapolate much beyond the context length seen in training: $8k$. E.g., in the passkey retrieval task, it can handle inputs of length $256k$.

<div align="center"> </div>

Instruction/Chat tuning

In the fine_tuning subfolder we provide the code that was used to create LongLLaMA-Instruct-3Bv1.1, an instruction-tuned version of LongLLaMA-3Bv1.1. We used OpenOrca (instructions) and zetavg/ShareGPT-Processed (chat) datasets for tuning.

Usage

See also:

• Colab with LongLLaMA-Instruct-3Bv1.1.
• Colab with an example usage of base LongLLaMA.

Requirements

pip install --upgrade pip
pip install transformers==4.30  sentencepiece accelerate

Loading model

import torch
from transformers import LlamaTokenizer, AutoModelForCausalLM

tokenizer = LlamaTokenizer.from_pretrained("syzymon/long_llama_3b_instruct")
model = AutoModelForCausalLM.from_pretrained("syzymon/long_llama_3b_instruct", 
                                            torch_dtype=torch.float32, 
                                            trust_remote_code=True)

Input handling and generation

LongLLaMA uses the Hugging Face interface, the long input given to the model will be split into context windows and loaded into the memory cache.

prompt = "My name is Julien and I like to"
input_ids = tokenizer(prompt, return_tensors="pt").input_ids
outputs = model(input_ids=input_ids)

During the model call, one can provide the parameter last_context_length (default $1024$), which specifies the number of tokens left in the last context window. Tuning this parameter can improve generation as the first layers do not have access to memory. See details in How LongLLaMA handles long inputs.

generation_output = model.generate(
    input_ids=input_ids,
    max_new_tokens=256,
    num_beams=1,
    last_context_length=1792,
    do_sample=True,
    temperature=1.0,
)
print(tokenizer.decode(generation_output[0]))

Additional configuration

LongLLaMA has several other parameters:

• mem_layers specifies layers endowed with memory (should be either an empty list or a list of all memory layers specified in the description of the checkpoint).
• mem_dtype allows changing the type of memory cache
• mem_attention_grouping can trade off speed for reduced memory usage. When equal to (4, 2048), the memory layers will process at most $4*2048$ queries at once ($4$ heads and $2048$ queries for each head).

import torch
from transformers import LlamaTokenizer, AutoModelForCausalLM

tokenizer = LlamaTokenizer.from_pretrained("syzymon/long_llama_3b_instruct")
model = AutoModelForCausalLM.from_pretrained(
    "syzymon/long_llama_3b_instruct", torch_dtype=torch.float32, 
    mem_layers=[], 
    mem_dtype='bfloat16',
    trust_remote_code=True,
    mem_attention_grouping=(4, 2048),
)

Drop-in use with LLaMA code

LongLLaMA checkpoints can also be used as a drop-in replacement for LLaMA checkpoints in Hugging Face implementation of LLaMA, but in this case, they will be limited to the original context length of $2048$.

from transformers import LlamaTokenizer, LlamaForCausalLM
import torch

tokenizer = LlamaTokenizer.from_pretrained("syzymon/long_llama_3b_instruct")
model = LlamaForCausalLM.from_pretrained("syzymon/long_llama_3b_instruct", torch_dtype=torch.float32)

How LongLLaMA handles long inputs

Inputs over $2048$ tokens are automatically split into windows $w_1, \ldots, w_m$. The first $m-2$ windows contain $2048$ tokens each, $w_{m-1}$ has no more than $2048$ tokens, and $w_m$ contains the number of tokens specified by last_context_length. The model processes the windows one by one extending the memory cache after each. If use_cache is True, the last window will not be loaded to the memory cache but to the local (generation) cache.

The memory cache stores $(key, value)$ pairs for each head of the specified memory layers mem_layers. In addition to this, it stores attention masks.

If use_cache=True (which is the case in generation), LongLLaMA will use two caches: the memory cache for the specified layers and the local (generation) cache for all layers. When the local cache exceeds $2048$ elements, its content is moved to the memory cache for the memory layers.

For simplicity, context extension is realized with a memory cache and full attention in this repo. Replacing this simple mechanism with a KNN search over an external database is possible with systems like Faiss. This potentially would enable further context length scaling. We leave this as a future work.

LongLLaMA performance

We present some illustrative examples of LongLLaMA results. Refer to our paper Focused Transformer: Contrastive Training for Context Scaling for more details.

We manage to achieve good performance on the passkey retrieval task from Landmark Attention: Random-Access Infinite Context Length for Transformers. The code for generating the prompt and running the model is located in examples/passkey.py.

<p align="center" width="100%"> <img src="assets/plot_passkey.png" alt="LongLLaMA" style="width: 70%; min-width: 300px; display: block; margin: auto;"> </p>

Our LongLLaMA 3B model also shows improvements when using long context on two downstream tasks, TREC question classification and WebQS question answering.

<div align="center"> </div>

LongLLaMA retains performance on tasks that do not require long context. We provide a comparison with OpenLLaMA on lm-evaluation-harness in the zero-shot setting.

<div align="center"> </div>

Starting with v1.1 models we have decided to use EleutherAI implementation of lm-evaluation-harness wit a slight modification, that adds <bos> token at beginning of input sequence. The results are provided in the table below.

<div align="center"> </div>

We also provide the results on human-eval. We cut the generated text after either

• "\ndef "
• "\nclass "
• "\nif __name__"

<div align="center"> </div>

Authors

• Szymon Tworkowski
• Konrad Staniszewski
• Mikołaj Pacek
• Henryk Michalewski
• Yuhuai Wu
• Piotr Miłoś

Citation

To cite this work please use

@misc{tworkowski2023focused,
      title={Focused Transformer: Contrastive Training for Context Scaling}, 
      author={Szymon Tworkowski and Konrad Staniszewski and Mikołaj Pacek and Yuhuai Wu and Henryk Michalewski and Piotr Miłoś},
      year={2023},
      eprint={2307.03170},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

License

The code and base models checkpoints are licensed under Apache License, Version 2.0.
The instruction/chat tuned models are for research purposes only.
Some of the examples use external code (see headers of files for copyright notices and licenses).

Acknowledgments

We gratefully acknowledge the TPU Research Cloud program, which was instrumental to our research by providing significant computational resources. We are also grateful to Xinyang Geng and Hao Liu for releasing OpenLLaMA checkpoints and the EasyLM library.

End of original Model File

Please consider to support my work

Coming Soon: I'm in the process of launching a sponsorship/crowdfunding campaign for my work. I'm evaluating Kickstarter, Patreon, or the new GitHub Sponsors platform, and I am hoping for some support and contribution to the continued availability of these kind of models. Your support will enable me to provide even more valuable resources and maintain the models you rely on. Your patience and ongoing support are greatly appreciated as I work to make this page an even more valuable resource for the community.

<center>

</center>