nm-testing/TinyLlama-1.1B-Chat-v1.0-open_platypus-pruned50-quant-ds

TinyLlama 1.1B Chat 1.0 - DeepSparse

This repo contains model files for TinyLlama 1.1B Chat optimized for DeepSparse, a CPU inference runtime for sparse models.

This model was quantized and pruned with SparseGPT, using SparseML.

Inference

Install DeepSparse LLM for fast inference on CPUs:

pip install deepsparse-nightly[llm]

Run in a Python pipeline:

from deepsparse import TextGeneration

prompt = "How to make banana bread?"
formatted_prompt =  f"<|im_start|>user\n{prompt}<|im_end|>\n<|im_start|>assistant\n"

model = TextGeneration(model_path="hf:nm-testing/TinyLlama-1.1B-Chat-v1.0-pruned50-quant-ds")
print(model(formatted_prompt, max_new_tokens=200).generations[0].text)

"""


"""

Prompt template

<|im_start|>user\n
{prompt}<|im_end|>\n
<|im_start|>assistant\n

Sparsification

For details on how this model was sparsified, see the recipe.yaml in this repo and follow the instructions below.

git clone https://github.com/neuralmagic/sparseml
pip install -e "sparseml[transformers]"
python sparseml/src/sparseml/transformers/sparsification/obcq/obcq.py TinyLlama/TinyLlama-1.1B-Chat-v1.0 open_platypus --precision float16  --recipe recipe.yaml --save True

Sparse Finetuning

Continue training the sparse model to improve accuracy:

from sparseml.transformers.finetune.text_generation import run_train


model = "./obcq_deployment"
teacher_model = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
dataset_name = "open_platypus"
concatenate_data = False
output_dir = "./output_finetune"
recipe = "recipe.yaml"
num_train_epochs=2
overwrite_output_dir = True
splits = {
    "train": "train[:50%]",
}

run_train(
    model_name_or_path=model,
    distill_teacher=teacher_model,
    dataset_name=dataset_name,
    output_dir=output_dir,
    recipe=recipe,
    num_train_epochs=num_train_epochs,
    overwrite_output_dir=overwrite_output_dir,
    concatenate_data = concatenate_data,
    splits = splits
)

Export Model

Export the model while injecting the KV Cache

sparseml.export --task text-generation output_finetune/

Follow the instructions on our One Shot With SparseML page for a step-by-step guide for performing one-shot quantization of large language models.

Slack

For further support, and discussions on these models and AI in general, join Neural Magic's Slack Community