Esperanto/llama-3.2-3B-Instruct-kvc-fp16-onnx

Llama 3.2 3B Instruct with Key-Value-Cache enabled in ONNX fp16 format

• Model creator: Meta-Llama
• Original model: Meta-Llama Llama 3.2 3B Instruct

Description

This repo contains the ONNX files for the ONNX conversion of Llama 3.2 3B Instruct done by Esperanto Technologies. The model is in the fp16 format and has the KVC enabled.

How to download ONNX model and weight files

The easiest way to obtain the model is to clone this whole repo. Alternatively you can download the files is using the huggingface-hub Python library.

pip3 install huggingface-hub>=0.17.1

Then you can download any individual model file to the current directory, at high speed, with a command like this:

huggingface-cli download Esperanto/llama-3.2-3B-Instruct-kvc-fp16-onnx --local-dir llama-3.2-3B-Instruct-kvc-fp16-onnx --local-dir-use-symlinks False

For more documentation on downloading with huggingface-cli, please see: HF -> Hub Python Library -> Download files -> Download from the CLI.

How to run from Python code using ONNXRuntime

This model can easily be ran in a CPU using ONNXRuntime.

First install the packages

pip3 install onnx==1.16.1
pip3 install onnxruntime==1.17.1

Example code: generate text with this model

We define the loop with greedy decoding:

import numpy as np
import onnxruntime
import onnx
from transformers import AutoTokenizer

def generate_text(model_path, prompt, tokenizer, max_gen_tokens, total_sequence, window, context):
    model = onnx.load(model_path)

    #we create the inputs for the first iteration
    input_tensor = tokenizer(prompt, return_tensors="pt")
    prompt_size = len(input_tensor['input_ids'][0])
    actual_input = input_tensor['input_ids']
    if prompt_size < window:
        actual_input = np.concatenate((tokenizer.bos_token_id*np.ones([1, window - prompt_size], dtype = 'int64'),
                                       actual_input), axis=1)
    if prompt_size + max_gen_tokens > total_sequence:
        print("ERROR: Longer total sequence is needed!")
        return
    first_attention = np.concatenate((np.zeros([1, total_sequence - window], dtype = 'int64'),
                                      np.ones((1, window), dtype = 'int64')), axis=1)
    max_gen_tokens += prompt_size #we need to generate on top of parsing the prompt
    inputs_names =[node.name for node in model.graph.input]
    output_names =[node.name for node in model.graph.output]
    n_heads = 8 #gqa-heads of the kvc
    inputs_dict = {}
    inputs_dict['input_ids'] = actual_input[:, :window].reshape(1, window).numpy()
    inputs_dict['attention_mask'] = first_attention
    index_pos = sum(first_attention[0])
    inputs_dict['position_ids'] = np.concatenate((np.zeros([1, total_sequence - index_pos], dtype = 'int64'), np.arange(index_pos, dtype = 'int64').reshape(1, index_pos)), axis=1)
    inputs_dict['tree_attention'] = np.triu(-65504*np.ones(total_sequence), k= 1).astype('float16').reshape(1, 1, total_sequence, total_sequence)
    for name in inputs_names:
        if name == 'input_ids' or name == 'attention_mask' or name == 'position_ids' or name == 'tree_attention': continue
        inputs_dict[name] = np.zeros([1, n_heads, context-window, 128], dtype="float16")
    index = 0
    new_token = np.array([10])
    next_index = window
    old_j = 0
    total_input = actual_input.numpy()

    rt_session = onnxruntime.InferenceSession(model_path)
    ## We run the inferences
    while next_index < max_gen_tokens:
        if new_token.any() == tokenizer.eos_token_id:
            break
        #inference
        output = rt_session.run(output_names, inputs_dict)
        outs_dictionary = {name: content for (name, content) in zip (output_names, output)}
        #we prepare the inputs for the next inference
        for name in inputs_names:
            if name == 'input_ids':
                old_j = next_index
                if next_index < prompt_size:
                    if prompt_size - next_index >= window: next_index += window
                    else: next_index = prompt_size 
                    j = next_index - window
                else:
                    next_index +=1
                    j = next_index - window
                    new_token = outs_dictionary['logits'].argmax(-1).reshape(1, window)
                    total_input = np.concatenate((total_input, new_token[: , -1:]), axis = 1)
                inputs_dict['input_ids']= total_input[:, j:next_index].reshape(1, window)
            elif name == 'attention_mask':
                inputs_dict['attention_mask'] = np.concatenate((np.zeros((1, total_sequence-next_index), dtype = 'int64'), np.ones((1, next_index), dtype = 'int64')), axis=1)
            elif name == 'position_ids':
                inputs_dict['position_ids'] = np.concatenate((np.zeros([1, total_sequence - next_index], dtype = 'int64'), np.arange(next_index, dtype = 'int64').reshape(1, next_index)), axis=1)
            elif name == 'tree_attention': continue
            else:
                old_name = name.replace("past_key_values", "present")
                inputs_dict[name] = outs_dictionary[old_name][:, :, next_index-old_j:context-window+(next_index - old_j), :]

    answer = tokenizer.decode(total_input[0], skip_special_tokens=True, clean_up_tokenization_spaces=False)
    return answer

We now run the inferences:

tokenizer = AutoTokenizer.from_pretrained("Esperanto/llama-3.2-3B-Instruct-kvc-fp16-onnx")
model_path = "llama-3.2-3B-Instruct-kvc-fp16-onnx/model.onnx"

max_gen_tokens = 20    #number of tokens we want tog eneral
total_sequence = 128   #total sequence_length
context = 1024         #the context to extend the kvc
window = 16            #number of tokens we want to parse at the time
messages = [
    {"role": "system", "content": "You are a pirate chatbot who always responds in pirate speak!"},
    {"role": "user", "content": "Who are you?"},
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

generated = generate_text(model_path, prompt, tokenizer, max_gen_tokens, total_sequence, window, context)
print(generated)