metadata

language:
  - en
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|>

# 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/stablelm-7b-sft-v7-epoch-3

pip install hf-hub-ctranslate2>=2.0.8 ctranslate2>=3.14.0

Converted on 2023-06-02 using

ct2-transformers-converter --model OpenAssistant/stablelm-7b-sft-v7-epoch-3 --output_dir /home/michael/tmp-ct2fast-stablelm-7b-sft-v7-epoch-3 --force --copy_files tokenizer.json README.md tokenizer_config.json generation_config.json special_tokens_map.json .gitattributes --quantization int8_float16 --trust_remote_code

Checkpoint compatible to ctranslate2>=3.14.0 and hf-hub-ctranslate2>=2.0.8

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

from hf_hub_ctranslate2 import TranslatorCT2fromHfHub, GeneratorCT2fromHfHub
from transformers import AutoTokenizer

model_name = "michaelfeil/ct2fast-stablelm-7b-sft-v7-epoch-3"
# use either TranslatorCT2fromHfHub or GeneratorCT2fromHfHub here, depending on model.
model = GeneratorCT2fromHfHub(
        # load in int8 on CUDA
        model_name_or_path=model_name,
        device="cuda",
        compute_type="int8_float16",
        # tokenizer=AutoTokenizer.from_pretrained("OpenAssistant/stablelm-7b-sft-v7-epoch-3")
)
outputs = model.generate(
    text=["def fibonnaci(", "User: How are you doing? Bot:"],
    max_length=64,
    include_prompt_in_result=False
)
print(outputs)

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 StableLM-7B SFT-7 Model

This is the 7th iteration English supervised-fine-tuning (SFT) model of the Open-Assistant project. It is based on a StableLM 7B that was fine-tuned on human demonstrations of assistant conversations collected through the https://open-assistant.io/ human feedback web app before April 12, 2023.

Model Details

Developed by: Open-Assistant Contributors
Model type: Transformer-based Language Model
Language: English
Finetuned from: stabilityai/stablelm-base-alpha-7b
Code: Open-Assistant/model/model_training
Demo: TODO
License: Creative Commons license (CC BY-SA-4.0)
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.

Dev Details

wandb: https://wandb.ai/open-assistant/supervised-finetuning/runs/08dfhyuc
base model: stabilityai/stablelm-base-alpha-7b
checkpoint: 3 epochs (12000 steps)

command: deepspeed trainer_sft.py --configs defaults stablelm-7b oasst-mix --cache_dir /home/ubuntu/data_cache --output_dir .saved/stable-lm-7b-1 --num_train_epochs 4 --deepspeed

data:

oasst-mix:
  save_strategy: epoch
  sort_by_length: false
  use_custom_sampler: false
  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: 2023-04-12_oasst_release_ready_synth.jsonl.gz
    - vicuna:
        val_split: 0.05
        max_val_set: 800
        fraction: 1.0
    - dolly15k:
        val_split: 0.05
        max_val_set: 300
    - grade_school_math_instructions:
        val_split: 0.05
    - code_alpaca:
        val_split: 0.05
        max_val_set: 250

stablelm:

stablelm-7b:
  dtype: fp16
  log_dir: stablelm_log_7b
  model_name: stabilityai/stablelm-base-alpha-7b
  output_dir: stablelm_7b
  max_length: 4096
  warmup_steps: 100
  gradient_checkpointing: true
  gradient_accumulation_steps: 2
  per_device_train_batch_size: 4
  per_device_eval_batch_size: 4
  eval_steps: 100
  save_steps: 500
  num_train_epochs: 4
  save_total_limit: 4
  use_flash_attention: true

zero config:

{
  "fp16": {
    "enabled": "auto",
    "loss_scale": 0,
    "loss_scale_window": 1000,
    "initial_scale_power": 16,
    "hysteresis": 2,
    "min_loss_scale": 1
  },
  "bf16": {
    "enabled": "auto"
  },
  "optimizer": {
    "type": "AdamW",
    "params": {
      "lr": "auto",
      "betas": "auto",
      "eps": "auto",
      "weight_decay": "auto"
    }
  },
  "scheduler": {
    "type": "WarmupDecayLR",
    "params": {
      "warmup_min_lr": "auto",
      "warmup_max_lr": "auto",
      "warmup_num_steps": "auto",
      "total_num_steps": "auto"
    }
  },
  "zero_optimization": {
    "stage": 2,
    "allgather_partitions": true,
    "allgather_bucket_size": 1e9,
    "overlap_comm": false,
    "reduce_scatter": true,
    "reduce_bucket_size": 1e9,
    "contiguous_gradients": true
  },
  "gradient_accumulation_steps": "auto",
  "gradient_clipping": "auto",
  "steps_per_print": 2000,
  "train_batch_size": "auto",
  "train_micro_batch_size_per_gpu": "auto",
  "wall_clock_breakdown": false
}