keras/llama2_7b_en_int8 · Hugging Face

Model Overview

Llama 2 is a set of large language models published by Meta. Both pretrained and instruction tuned models are available, and range in size from 7 billion to 70 billion parameters. See the model card below for benchmarks, data sources, and intended use cases.

Weights are released under the Llama 2 Community License. Keras model code is released under the Apache 2 License.

Installation

Keras and KerasHub can be installed with:

pip install -U -q keras-hub
pip install -U -q keras>=3

Jax, TensorFlow, and Torch come preinstalled in Kaggle Notebooks. For instructions on installing them in another environment see the Keras Getting Started page.

Presets

The following model checkpoints are provided by the Keras team. Full code examples for each are available below.

Preset name	Parameters	Description
`llama2_7b_en`	6.74B	7 billion parameter, 32-layer, base LLaMA 2 model.
`llama2_7b_en_int8`	6.74B	7 billion parameter, 32-layer, base LLaMA 2 model with activation and weights quantized to int8.
`llama2_instruct_7b_en`	6.74B	7 billion parameter, 32-layer, instruction tuned LLaMA 2 model.
`llama2_instruct_7b_en_int8`	6.74B	7 billion parameter, 32-layer, instruction tuned LLaMA 2 model with activation and weights quantized to int8.

Prompts

Llama-2 "instruct" models are instruction tuned on turn by turn conversations and should be prompted with examples that precisely match the training data. Specifically, you must alternate user and assistant turns that begin and end with special tokens. New lines do matter. See the following for an example:

prompt = """&lt;s&gt;[INST] &lt;

Example Usage

import keras
import keras_hub
import numpy as np

Use generate() to do text generation.

llama_lm = keras_hub.models.LlamaCausalLM.from_preset("llama2_7b_en_int8")
llama_lm.generate("What is Keras?", max_length=500)

# Generate with batched prompts.
llama_lm.generate(["What is Keras?", "Give me your best brownie recipe."], max_length=500)

Compile the generate() function with a custom sampler.

llama_lm = keras_hub.models.LlamaCausalLM.from_preset("llama2_7b_en_int8")
llama_lm.compile(sampler="greedy")
llama_lm.generate("I want to say", max_length=30)

llama_lm.compile(sampler=keras_hub.samplers.BeamSampler(num_beams=2))
llama_lm.generate("I want to say", max_length=30)

Use generate() without preprocessing.

prompt = {
    # `1` maps to the start token followed by "I want to say".
    "token_ids": np.array([[1, 306, 864, 304, 1827, 0, 0, 0, 0, 0]] * 2),
    # Use `"padding_mask"` to indicate values that should not be overridden.
    "padding_mask": np.array([[1, 1, 1, 1, 1, 0, 0, 0, 0, 0]] * 2),
}

llama_lm = keras_hub.models.LlamaCausalLM.from_preset(
    "llama2_7b_en_int8",
    preprocessor=None,
    dtype="bfloat16"
)
llama_lm.generate(prompt)

Call fit() on a single batch.

features = ["The quick brown fox jumped.", "I forgot my homework."]
llama_lm = keras_hub.models.LlamaCausalLM.from_preset("llama2_7b_en_int8")
llama_lm.fit(x=features, batch_size=2)

Call fit() without preprocessing.

x = {
    "token_ids": np.array([[1, 450, 4996, 17354, 1701, 29916, 12500, 287, 29889, 0]] * 2),
    "padding_mask": np.array([[1, 1, 1, 1, 1, 1, 1, 1, 1, 0]] * 2),
}
y = np.array([[450, 4996, 17354, 1701, 29916, 12500, 287, 29889, 0, 0]] * 2)
sw = np.array([[1, 1, 1, 1, 1, 1, 1, 1, 0, 0]] * 2)

llama_lm = keras_hub.models.LlamaCausalLM.from_preset(
    "llama2_7b_en_int8",
    preprocessor=None,
    dtype="bfloat16"
)
llama_lm.fit(x=x, y=y, sample_weight=sw, batch_size=2)

Example Usage with Hugging Face URI

import keras
import keras_hub
import numpy as np