metadata

license: apache-2.0
language:
  - en
  - de
  - es
  - fr
tags:
  - ctranslate2
  - int8
  - float16
  - sft
pipeline_tag: text-generation
widget:
  - text: >-
      <|prompter|>What is a meme, and what's the history behind this
      word?<|endoftext|><|assistant|>
  - text: <|prompter|>What's the Earth total population<|endoftext|><|assistant|>
  - text: >-
      <|prompter|>Write a story about future of AI
      development<|endoftext|><|assistant|>
datasets:
  - OpenAssistant/oasst1
library_name: transformers

# Fast-Inference with Ctranslate2

Speedup inference while reducing memory by 2x-4x using int8 inference in C++ on CPU or GPU.

quantized version of OpenAssistant/falcon-7b-sft-top1-696

pip install hf-hub-ctranslate2>=2.10.0 ctranslate2>=3.16.0

# from transformers import AutoTokenizer
model_name = "michaelfeil/ct2fast-falcon-7b-sft-top1-696"

from hf_hub_ctranslate2 import GeneratorCT2fromHfHub
model = GeneratorCT2fromHfHub(
        # load in int8 on CUDA
        model_name_or_path=model_name,
        device="cuda",
        compute_type="int8_float16",
        # tokenizer=AutoTokenizer.from_pretrained("{ORG}/{NAME}")
)
outputs = model.generate(
    text=["def fibonnaci(", "User: How are you doing? Bot:"],
    max_length=64,
    include_prompt_in_result=False
)
print(outputs)

Checkpoint compatible to ctranslate2>=3.16.0 and hf-hub-ctranslate2>=2.10.0

compute_type=int8_float16 for device="cuda"
compute_type=int8 for device="cpu"

Converted on 2023-06-16 using

ct2-transformers-converter --model OpenAssistant/falcon-7b-sft-top1-696 --output_dir ~/tmp-ct2fast-falcon-7b-sft-top1-696 --force --copy_files tokenizer.json README.md tokenizer_config.json generation_config.json special_tokens_map.json .gitattributes --quantization int8_float16 --trust_remote_code

Licence and other remarks:

This is just a quantized version. Licence conditions are intended to be idential to original huggingface repo.

Original description

Open-Assistant Falcon 7B SFT OASST-TOP1 Model

This model is a fine-tuning of TII's Falcon 7B LLM. It was trained with 11,123 top-1 (high-quality) demonstrations of the OASST data set (exported on June 2, 2023) with a batch size of 128 for 8 epochs with LIMA style dropout (p=0.2) and a context-length of 2048 tokens.

Model Details

Finetuned from: tiiuae/falcon-7b
Model type: Causal decoder-only transformer language model
Language: English, German, Spanish, French (and limited capabilities in Italian, Portuguese, Polish, Dutch, Romanian, Czech, Swedish);
Weights & Biases: Training log (Checkpoint: 696 steps)
Code: Open-Assistant/model/model_training
Demo: Continuations for 250 random prompts
License: Apache 2.0
Contact: Open-Assistant Discord

Prompting

Input prompt example:

<|prompter|>What is a meme, and what's the history behind this word?<|endoftext|><|assistant|>

The input ends with the <|assistant|> token to signal that the model should start generating the assistant reply.

Sample Code

from transformers import AutoTokenizer
import transformers
import torch

model = "OpenAssistant/falcon-7b-sft-top1-696"

tokenizer = AutoTokenizer.from_pretrained(model)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
    device_map="auto",
)

input_text="<|prompter|>What is a meme, and what's the history behind this word?<|endoftext|><|assistant|>"

sequences = pipeline(
    input_text,
    max_length=500,
    do_sample=True,
    return_full_text=False,
    top_k=10,
    num_return_sequences=1,
    eos_token_id=tokenizer.eos_token_id,
)
for seq in sequences:
    print(f"Result: {seq['generated_text']}")

Configuration Details

Model:

falcon-7b:
  dtype: bf16
  log_dir: "falcon_log_7b"
  learning_rate: 1e-5
  model_name: "tiiuae/falcon-7b"
  deepspeed_config: configs/zero_config.json
  output_dir: falcon
  weight_decay: 0.0
  max_length: 2048
  save_strategy: steps
  eval_steps: 80
  save_steps: 80
  warmup_steps: 20
  gradient_checkpointing: true
  gradient_accumulation_steps: 4
  per_device_train_batch_size: 4
  per_device_eval_batch_size: 8
  num_train_epochs: 8
  save_total_limit: 4
  residual_dropout: 0.2
  residual_dropout_lima: true

Dataset:

oasst-top1:
  # oasst_export: 11123 (100.00%)
  datasets:
    - oasst_export:
        lang: "bg,ca,cs,da,de,en,es,fr,hr,hu,it,nl,pl,pt,ro,ru,sl,sr,sv,uk" # sft-8.0
        input_file_path: 2023-06-02_oasst_all_labels.jsonl.gz
        val_split: 0.05
        top_k: 1

Train command:

deepspeed trainer_sft.py --configs defaults falcon-7b oasst-top1 --cache_dir <data_cache_dir> --output_dir <output_path> --deepspeed

Export command:

python export_model.py --dtype bf16 --hf_repo_name OpenAssistant/falcon-7b-sft-top1 --trust_remote_code --auth_token <auth_token> <output_path> --max_shard_size 2GB