Hugging Face's logo

TheBloke
/
Llama2-70B-OASST-SFT-v10-GGML

Text Generation Transformers
English llama sft text-generation-inference
Model card Files Community
Edit model card
TheBlokeAI

TheBloke's LLM work is generously supported by a grant from andreessen horowitz (a16z)

Llama2 70B SFT v10 - GGML

Description

This repo contains GGML format model files for OpenAssistant's Llama2 70B SFT v10.

Important note regarding GGML files.

The GGML format has now been superseded by GGUF. As of August 21st 2023, 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

GPU acceleration is now available for Llama 2 70B GGML files, with both CUDA (NVidia) and Metal (macOS). The following clients/libraries are known to work with these files, including with GPU acceleration:

  • llama.cpp, commit e76d630 and later.
  • text-generation-webui, the most widely used web UI.
  • KoboldCpp, version 1.37 and later. A powerful GGML web UI, especially good for story telling.
  • LM Studio, a fully featured local GUI with GPU acceleration for both Windows and macOS. Use 0.1.11 or later for macOS GPU acceleration with 70B models.
  • llama-cpp-python, version 0.1.77 and later. A Python library with LangChain support, and OpenAI-compatible API server.
  • ctransformers, version 0.2.15 and later. A Python library with LangChain support, and OpenAI-compatible API server.

Repositories available

Prompt template: ChatML 

<|im_start|>system
{system_message}<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant

Compatibility

Works with llama.cpp commit e76d630 until August 21st, 2023

Will not work with llama.cpp after commit dadbed99e65252d79f81101a392d0d6497b86caa.

For compatibility with latest llama.cpp, please use GGUF files instead.

Or one of the other tools and libraries listed above.

To use in llama.cpp, you must add -gqa 8 argument.

For other UIs and libraries, please check the docs.

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
llama2-70b-oasst-sft-v10.ggmlv3.Q2_K.bin Q2_K 2 28.96 GB 31.46 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.
llama2-70b-oasst-sft-v10.ggmlv3.Q3_K_S.bin Q3_K_S 3 30.09 GB 32.59 GB New k-quant method. Uses GGML_TYPE_Q3_K for all tensors
llama2-70b-oasst-sft-v10.ggmlv3.Q3_K_M.bin Q3_K_M 3 33.39 GB 35.89 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
llama2-70b-oasst-sft-v10.ggmlv3.Q3_K_L.bin Q3_K_L 3 36.49 GB 38.99 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
llama2-70b-oasst-sft-v10.ggmlv3.Q4_0.bin Q4_0 4 38.80 GB 41.30 GB Original quant method, 4-bit.
llama2-70b-oasst-sft-v10.ggmlv3.Q4_K_S.bin Q4_K_S 4 39.18 GB 41.68 GB New k-quant method. Uses GGML_TYPE_Q4_K for all tensors
llama2-70b-oasst-sft-v10.ggmlv3.Q4_K_M.bin Q4_K_M 4 41.69 GB 44.19 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
llama2-70b-oasst-sft-v10.ggmlv3.Q4_1.bin Q4_1 4 43.12 GB 45.62 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.
llama2-70b-oasst-sft-v10.ggmlv3.Q5_0.bin Q5_0 5 47.43 GB 49.93 GB Original quant method, 5-bit. Higher accuracy, higher resource usage and slower inference.
llama2-70b-oasst-sft-v10.ggmlv3.Q5_K_S.bin Q5_K_S 5 47.74 GB 50.24 GB New k-quant method. Uses GGML_TYPE_Q5_K for all tensors
llama2-70b-oasst-sft-v10.ggmlv3.Q5_K_M.bin Q5_K_M 5 49.03 GB 51.53 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

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 or earlier.

For compatibility with latest llama.cpp, please use GGUF files instead.

I use the following command line; adjust for your tastes and needs:

./main -t 10 -ngl 40 -gqa 8 -m llama2-70b-oasst-sft-v10.ggmlv3.q4_K_M.bin --color -c 4096 --temp 0.7 --repeat_penalty 1.1 -n -1 -p "<|im_start|>system\n{system_message}<|im_end|>\n<|im_start|>user\n{prompt}<|im_end|>\n<|im_start|>assistant"

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. If you are fully offloading the model to GPU, use -t 1

Change -ngl 40 to the number of GPU layers you have VRAM for. Use -ngl 100 to offload all layers to VRAM - if you have a 48GB card, or 2 x 24GB, or similar. Otherwise you can partially offload as many as you have VRAM for, on one or more GPUs.

If you want to have a chat-style conversation, replace the -p <PROMPT> argument with -i -ins

Remember the -gqa 8 argument, required for Llama 70B models.

Change -c 4096 to the desired sequence length for this 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.

For other parameters and how to use them, please refer to the llama.cpp documentation

How to run in text-generation-webui

Further instructions here: text-generation-webui/docs/llama.cpp-models.md.

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: 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: OpenAssistant's Llama2 70B SFT v10

Open-Assistant Llama2 70B SFT v10

This model is an Open-Assistant fine-tuning of Meta's Llama2 70B LLM. It was fine-tuned in two stages, first on a mix of synthetic instrunctions and coding tasks and then in a "polishing" stage on the best human demonstrations collected at open-assistant.io up to July 23, 2023 (see Configuration Details below).

Model Details

Prompting / Prompt Template

Due to public demand (see survey) we changed the prompt-template for this model from custom prompter/assistant tokens to OpenAI's chatml standard prompt format. We hope that this leads to greater compatibility with chat inference/frontend applications.

Prompt dialogue template:

"""
<|im_start|>system
{system_message}<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant
"""

The model input can contain multiple conversation turns between user and assistant, e.g.

<|im_start|>user
{prompt 1}<|im_end|>
<|im_start|>assistant
{reply 1}<|im_end|>
<|im_start|>user
{prompt 2}<|im_end|>
<|im_start|>assistant
(...)

The model was partly trained with orca system messages.
For inference we recommend to use the official Llama2 system message:

<|im_start|>system
You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature.

If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information.
<|im_end|>

Credits & Special Thanks

We want to especially thank everyone who contributed in the crowed-sourced Open-Assistant dataset creation on https://open-assistant.io/ - without you this project would not have been possible.

Ethical Considerations and Limitations

Testing conducted to date has been in English, and has not covered, nor could it cover all scenarios. For these reasons, as with all LLMs, the potential outputs of llama2-70b-oasst-sft-v10 cannot be predicted in advance, and the model may in some instances produce inaccurate, biased or other objectionable responses to user prompts. Therefore, before deploying any applications of llama2-70b-oasst-sft-v10, developers should perform safety testing and tuning tailored to their specific applications of the model.

Please see Meta's Responsible Use Guide.

Inference via TGI

An early version of this model had an embedding count of 32,007 which was incompatible to sharding with TGI. In the current version the embeddings and the lm_head weights have been padded to a multiple of 128 (by replicating the emembeddings of the unk-token (id: 0)). Sharded inference with TGI should now work as expected.

Configuration Details

The "pretokenizer" utility used to tokenize the datamix is part of the Open-Assistant github repository and can be found here: model/pretokenizer.

Stage 1 Pretokenizer Configuration

Entries of the dataset with assistant replies shorter than 25 tokens were excluded from training.

oasst_pre10_min25:
  datasets:
    - megacode2:
        fraction: 0.5
        val_split: 0.01
        max_val_set: 1000
    - orca-chat:
        val_split: 0.01
        max_val_set: 1000
    - dolly15k_multilingual:
        val_split: 0.05
        max_val_set: 300
    - oa_leet10k:
        val_split: 0.05
        max_val_set: 250
  output_dir: "output/oasst_pre10_min25"
  filename_prefix: "oasst_pre10"
  min_assistant_tokens: 25

Stage 1 dataset statistics:

# Stats for output/oasst_pre10_min25_llama2

## Stats for 'Subset of InstructionDataset (megacode2)' (466364 samples (50.0%))
-----------------
  Accepted: 398223/466364 (85.4%)
  Accepted tokens: 167676873
  Skipped: 68141 (14.6%)
  Min tokens per sample: 36
  Max tokens per sample: 11810
  Avg tokens per sample: 421.063
-----------------

## Stats for 'Subset of OrcaChat (orca-chat)' (325616 samples (100.0%))
-----------------
  Accepted: 325616/325616 (100.0%)
  Accepted tokens: 178307574
  Skipped: 0 (0.0%)
  Min tokens per sample: 105
  Max tokens per sample: 10408
  Avg tokens per sample: 547.601
-----------------

## Stats for 'Subset of Dolly15kMultilingual' (57020 samples (100.0%))
-----------------
  Accepted: 47494/57020 (83.3%)
  Accepted tokens: 13883177
  Skipped: 9526 (16.7%)
  Min tokens per sample: 34
  Max tokens per sample: 9172
  Avg tokens per sample: 292.314
-----------------

## Stats for 'Subset of InstructionDataset (oa_leet10k)' (22236 samples (100.0%))
-----------------
  Accepted: 22236/22236 (100.0%)
  Accepted tokens: 15905296
  Skipped: 0 (0.0%)
  Min tokens per sample: 168
  Max tokens per sample: 10588
  Avg tokens per sample: 715.295
-----------------

## Stats for 'total' (871236 samples (100.0%))
-----------------
  Accepted: 793569/871236 (91.1%)
  Accepted tokens: 375772920
  Skipped: 77667 (8.9%)
  Min tokens per sample: 34
  Max tokens per sample: 11810
  Avg tokens per sample: 473.523
-----------------

Stage 2 Pretokenizer Configuration 

oasst_top1:
  datasets:
    - oasst_export:
        lang: "bg,ca,cs,da,de,en,es,fr,hr,hu,it,nl,pl,pt,ro,ru,sl,sr,sv,uk"
        input_file_path: 2023-07-23_oasst_ready.tar.gz
        top_k: 1
        val_split: 0.05
  output_dir: "output/oasst_top1_2023-07-23"
  filename_prefix: "oasst_top1"

Stage 2 dataset statistics:

# Stats for output/oasst_top1_2023-07-23_llama2

## Stats for 'ListDataset' (11441 samples (100.0%))
-----------------
  Accepted: 11441/11441 (100.0%)
  Accepted tokens: 5315368
  Skipped: 0 (0.0%)
  Min tokens per sample: 20
  Max tokens per sample: 5407
  Avg tokens per sample: 464.58945896337735
-----------------

## Stats for 'total' (11441 samples (100.0%))
-----------------
  Accepted: 11441/11441 (100.0%)
  Accepted tokens: 5315368
  Skipped: 0 (0.0%)
  Min tokens per sample: 20
  Max tokens per sample: 5407
  Avg tokens per sample: 464.58945896337735
-----------------

Megatron Fine-Tuning Arguments for Stage 1 (Instruction Tuning): 

--tensor_model_parallel_size 8
--pipeline_model_parallel_size 4
--load ./checkpoints/llama2-70b-tp8-pp4
--save ./checkpoints/llama2-70b-tp8-pp4-oasst_pre10
--tensorboard_dir ./checkpoints/llama2-70b-tp8-pp4-oasst_pre10/logging
--data_path ./data/oasst_pre10_min25_llama2/oasst_sft10-train
--model_name llama2
--tokenizer_type SentencePieceTokenizer
--bf16
--global_batch_size 64
--micro_batch_size 2
--vocab_file=./llama2/Llama-2-7b/tokenizer.model
--use_rms_norm
--glu_activation swiglu
--no_tie_embed_logits
--vocab_extra_ids_list "\"<|im_start|>,<|im_end|>\""
--layernorm_epsilon 1e-5
--use_flash_attn
--no_bias_gelu_fusion
--seq_length 4096
--max_position_embeddings 4096
--log_interval 1
--save_interval 500
--eval_interval 50
--eval_iters 10
--hidden_dropout 0.0
--position_embedding_type rotary
--no_bias_dropout_fusion
--use_checkpoint_args
--train_iters 12000
--attention_dropout 0.0
--adam_beta1 0.9
--adam_beta2 0.95
--adam_eps 1e-12
--lr_decay_style cosine
--lr_warmup_iters 100
--lr 1e-5
--min_lr 1e-6
--weight_decay 0.000001
--sequence_parallel
--recompute_granularity selective
--log_timers_to_tensorboard
--rope_scaling_factor 1.0
--wandb_logger

Megatron Fine-Tuning Arguments for Stage 2 (OASST Polishing, LIMA Dropout): 

--tensor_model_parallel_size 8
--pipeline_model_parallel_size 4
--load ./checkpoints/llama2-70b-tp8-pp4-oasst_pre10
--save ./checkpoints/llama2-70b-tp8-pp4-oasst_sft10
--tensorboard_dir ./checkpoints/llama2-70b-tp8-pp4-oasst_sft10/logging
--data_path ./data/oasst_top1_2023-07-23_llama2/oasst_top1-train
--model_name llama2
--tokenizer_type SentencePieceTokenizer
--bf16
--global_batch_size 64
--micro_batch_size 2
--vocab_file=./llama2/Llama-2-7b/tokenizer.model
--use_rms_norm
--glu_activation swiglu
--no_tie_embed_logits
--vocab_extra_ids_list "\"<|im_start|>,<|im_end|>\""
--layernorm_epsilon 1e-5
--use_flash_attn
--no_bias_gelu_fusion
--seq_length 4096
--max_position_embeddings 4096
--log_interval 1
--save_interval 346
--eval_interval 50
--eval_iters 10
--hidden_dropout 0.25
--lima_dropout
--position_embedding_type rotary
--no_bias_dropout_fusion
--use_checkpoint_args
--train_iters 519
--attention_dropout 0.0
--adam_beta1 0.9
--adam_beta2 0.95
--adam_eps 1e-12
--lr_decay_style cosine
--lr_warmup_iters 100
--lr 1e-5
--min_lr 1e-6
--weight_decay 0.000001
--sequence_parallel
--recompute_granularity selective
--log_timers_to_tensorboard
--rope_scaling_factor 1.0
--finetune
--wandb_logger
Downloads last month
7
Inference Examples
Text Generation
Inference API (serverless) has been turned off for this model.

Finetuned from

Datasets used to train TheBloke/Llama2-70B-OASST-SFT-v10-GGML