Skip to content

Llama3-8B

Overview

This tutorial explains how to run the model on vLLM using multiple RBLN NPUs. For this guide, we will use the meta-llama/Meta-Llama-3-8B-Instruct model.

Setup & Installation

Before you begin, ensure that your system environment is properly configured and that all required packages are installed. This includes:

Note

Please note that rebel-compiler requires an RBLN Portal account.

Note

Please note that the meta-llama/Meta-Llama-3-8B-Instruct model on HuggingFace has restricted access. Once access is granted, you can log in using the huggingface-cli command as shown below:

$ huggingface-cli login

    _|    _|  _|    _|    _|_|_|    _|_|_|  _|_|_|  _|      _|    _|_|_|      _|_|_|_|    _|_|      _|_|_|  _|_|_|_|
    _|    _|  _|    _|  _|        _|          _|    _|_|    _|  _|            _|        _|    _|  _|        _|
    _|_|_|_|  _|    _|  _|  _|_|  _|  _|_|    _|    _|  _|  _|  _|  _|_|      _|_|_|    _|_|_|_|  _|        _|_|_|
    _|    _|  _|    _|  _|    _|  _|    _|    _|    _|    _|_|  _|    _|      _|        _|    _|  _|        _|
    _|    _|    _|_|      _|_|_|    _|_|_|  _|_|_|  _|      _|    _|_|_|      _|        _|    _|    _|_|_|  _|_|_|_|

    To login, `huggingface_hub` requires a token generated from [https://huggingface.co/settings/tokens](https://huggingface.co/settings/tokens) .
Token: *****

Execution

Model Compilation

To begin, import the RBLNLlamaForCausalLM class from optimum-rbln. This class's from_pretrained() method downloads the Llama 3 model from the HuggingFace Hub and compiles it using the RBLN Compiler. When exporting the model, specify the following parameters:

  • export: Must be True to compile the model.

  • rbln_batch_size: Defines the batch size for compilation.

  • rbln_max_seq_len: Defines the maximum sequence length.

  • rbln_tensor_parallel_size: Defines the number of NPUs to be used for inference.

After compilation, save the model artifacts to disk using save_pretrained(), creating a directory (e.g., rbln-Llama-3-8B-Instruct) with the compiled model.

Note

Select batch size based on model size and NPU specs. Moreover, vllm-rbln supports Dynamic Batching to ensure optimal throughput and resource utilization. See Dynamic Batching for details.

from optimum.rbln import RBLNLlamaForCausalLM

# Define the HuggingFace model ID
model_id = "meta-llama/Meta-Llama-3-8B-Instruct"

# Compile the model for 4 RBLN NPUs
compiled_model = RBLNLlamaForCausalLM.from_pretrained(
    model_id=model_id,
    export=True,
    rbln_batch_size=4,
    rbln_max_seq_len=8192,
    rbln_tensor_parallel_size=4,
)

compiled_model.save_pretrained("rbln-Llama-3-8B-Instruct")

Inference using vLLM

You can use the compiled model with vLLM. The example below shows how to set up the vLLM engine using a compiled model and run inference.

Note

In the case of regular Paged Attention, not Flash Attention, block_size should match with max_seq_len.

from transformers import AutoTokenizer
from vllm import LLM, SamplingParams

# Make sure the engine configuration
# matches the parameters used during compilation.
model_id = "rbln-Llama-3-8B-Instruct"
max_seq_len = 8192
batch_size = 4

llm = LLM(
    model=model_id,
    device="rbln",
    max_num_seqs=batch_size,
    max_num_batched_tokens=max_seq_len,
    max_model_len=max_seq_len,
    block_size=max_seq_len,
)

tokenizer = AutoTokenizer.from_pretrained(model_id)

sampling_params = SamplingParams(
temperature=0.0,
skip_special_tokens=True,
stop_token_ids=[tokenizer.eos_token_id],
)

conversation = [
    {
        "role": "user",
        "content": "What is the first letter of English alphabets?"
    }
]

chat = tokenizer.apply_chat_template(
    conversation, 
    add_generation_prompt=True,
    tokenize=False
)

outputs = llm.generate(chat, sampling_params)
for output in outputs:
    prompt = output.prompt
    generated_text = output.outputs[0].text
    print(generated_text)

Example Output:

The first letter of the English alphabet is "A".

References