The instructions in this Learning Path are for any Arm server running Ubuntu 22.04 LTS. You need an Arm server instance with at least four cores and 8GB of RAM to run this example. The instructions have been tested on an AWS Graviton3 c7g.2xlarge instance.
Arm CPUs are widely used in traditional ML and AI use cases. In this Learning Path, you learn how to run generative AI inference-based use cases like a LLM chatbot on Arm-based CPUs. You do this by deploying the
Llama-2-7B-Chat model
on your Arm-based CPU using llama.cpp.
llama.cpp is an open source C/C++ project developed by Georgi Gerganov that enables efficient LLM inference on a variety of hardware - both locally, and in the cloud.
The Llama-2-7B-Chat model from Meta belongs to the Llama 2 model family and is free to use for research and commercial purposes. Before you use the model, visit the Llama website and fill in the form to request access.
Llama 2 collection of models perform general natural language processing (NLP) tasks such as text generation. You can access the base foundation Llama 2 model or select the specialized chat Llama 2 version that is already optimized for back-and-forth dialogue. In this Learning Path, you run the specialized chat model. The Llama 2 family of models range in size from 7 billion to 70 billion parameters. The greater the number of parameters, the more information the model can store. This directly affects how well the model understands language and the model’s general capabilities. LLMs that run efficiently on CPUs typically have lower numbers of parameters. For this example, the 7 billion (7b) model is ideal for retaining quality chatbot capability while also running efficiently on your Arm-based CPU.
Traditionally, the training and inference of LLMs has been done on GPUs using full-precision 32-bit (FP32) or half-precision 16-bit (FP16) data type formats for the model parameter and weights. Recently, a new binary model format called GGUF was introduced by the llama.cpp team. This new GGUF model format uses compression and quantization techniques that remove the dependency on using FP32 and FP16 data type formats. For example, GGUF supports quantization where model weights that are generally stored as FP16 data types are scaled down to 4-bit integers. This significantly reduces the need for computational resources and the amount of RAM required. These advancements made in the model format and the data types used make Arm CPUs a great fit for running LLM inferences.
Install the following packages on your Arm based server instance:
sudo apt update
sudo apt install make cmake -y
You also need to install gcc on your machine:
sudo apt install gcc g++ -y
sudo apt install build-essential -y
You are now ready to start building llama.cpp.
Clone the source repository for llama.cpp:
git clone https://github.com/ggerganov/llama.cpp
By default, llama.cpp builds for CPU only on Linux and Windows. You don’t need to provide any extra switches to build it for the Arm CPU that you run it on.
Run make to build it:
cd llama.cpp
make GGML_NO_LLAMAFILE=1 -j$(nproc)
Check that llama.cpp has built correctly by running the help command:
./llama-cli -h
If llama.cpp has built correctly on your machine, you will see the help options being displayed. A snippet of the output is shown below:
usage: ./llama-cli [options]
general:
-h, --help, --usage print usage and exit
--version show version and build info
-v, --verbose print verbose information
--verbosity N set specific verbosity level (default: 0)
--verbose-prompt print a verbose prompt before generation (default: false)
--no-display-prompt don't print prompt at generation (default: false)
-co, --color colorise output to distinguish prompt and user input from generations (default: false)
-s, --seed SEED RNG seed (default: -1, use random seed for < 0)
-t, --threads N number of threads to use during generation (default: 4)
-tb, --threads-batch N number of threads to use during batch and prompt processing (default: same as --threads)
-td, --threads-draft N number of threads to use during generation (default: same as --threads)
-tbd, --threads-batch-draft N number of threads to use during batch and prompt processing (default: same as --threads-draft)
--draft N number of tokens to draft for speculative decoding (default: 5)
-ps, --p-split N speculative decoding split probability (default: 0.1)
-lcs, --lookup-cache-static FNAME
path to static lookup cache to use for lookup decoding (not updated by generation)
-lcd, --lookup-cache-dynamic FNAME
path to dynamic lookup cache to use for lookup decoding (updated by generation)
-c, --ctx-size N size of the prompt context (default: 0, 0 = loaded from model)
-n, --predict N number of tokens to predict (default: -1, -1 = infinity, -2 = until context filled)
-b, --batch-size N logical maximum batch size (default: 2048)
There are a few different ways you can download the Llama-2-7B Chat model. In this Learning Path, you download the model from Hugging Face.
Use of Llama-2-7B-Chat model is governed by the Meta license. Before you proceed to download the model, please visit the Llama website and fill in the form.
Hugging Face is an open source AI community where you can host your own AI models, train them and collaborate with others in the community. You can browse through the thousands of models that are available for a variety of use cases like NLP, audio, and computer vision.
The huggingface_hub library provides APIs and tools that let you easily download and fine-tune pre-trained models. You will use huggingface-cli to download the
Llama-2-7B-Chat model
.
Install the required Python packages:
sudo apt install python-is-python3 python3-pip python3-venv -y
Create and activate a Python virtual environment:
python -m venv venv
source venv/bin/activate
Your terminal prompt now has the (venv) prefix indicating the virtual environment is active. Use this virtual environment for the remaining commands.
Install the huggingface_hub python library using pip:
pip install huggingface_hub
You can now download the model using the huggingface cli:
huggingface-cli download TheBloke/Llama-2-7b-Chat-GGUF llama-2-7b-chat.Q4_0.gguf --local-dir . --local-dir-use-symlinks False
Before you proceed and run this model, take a quick look at what Q4_0 in the model name denotes.
Q4_0 in the model name refers to the quantization method the model uses. The goal of quantization is to reduce the size of the model (to reduce the memory space required) and faster (to reduce memory bandwidth bottlenecks transferring large amounts of data from memory to a processor). The primary trade-off to keep in mind when reducing a model’s size is maintaining quality of performance. Ideally, a model is quantized to meet size and speed requirements while not having a negative impact on performance.
Llama 2 was originally trained and published using the bfloat16 data type, meaning that each of the 7 billion model parameters takes up 16 bits of memory to store. Putting that into real terms, multiplying 16 bits per parameter by 7 billion parameters, the base foundation llama-2-7b model is just over 13Gb in size.
This model is llama-2-7b-chat.Q4_0.gguf, so what does each component mean in relation to the quantization level? The main thing to note is the number of bits per parameter, which is denoted by ‘Q4’ in this case or 4-bit integer. As a result, by only using 4 bits per parameter for 7 billion parameters, the model drops to be 3.6Gb in size.
Here is a quick lookup to the rest of the quantization parts for the Llama-2 model family as it exists today:
| quantization-method | # of bits per parameter | quantization format (does not apply to quantization method ‘IQ’) | quantization method specifics |
|---|---|---|---|
| Q, IQ, F, FP | 2,3,4,5,6,7,8,16,32 | _0, _1, _K | _XXS, _XS, _S, _M, _L |
Some examples:
Each quantization method has a unique approach to quantizing parameters. The deeper technical details of different quantization methodologies are outside the scope of this guide. The main takeaway is that selecting the right model quantization is critical to running an LLM effectively on your hardware, and the most impactful quantization decision is the number of bits per parameter. You will need also need to check you have enough system memory before deploying larger models or models with higher precision/quantization.
In this guide, you will not use any other quantization methods, because Arm has not made kernel optimizations for other quantization types.
To see improvements for Arm optimized kernels, you need to generate a new weights file with rearranged Q4_0 weights. As of llama.cpp commit 0f1a39f3 , Arm has contributed code for three types of GEMV/GEMM kernels corresponding to three processor types:
To re-quantize optimally for Graviton3, run
./llama-quantize --allow-requantize llama-2-7b-chat.Q4_0.gguf llama-2-7b-chat.Q4_0_8_8.gguf Q4_0_8_8
This will output a new file, llama-2-7b-chat.Q4_0_8_8.gguf, which contains reconfigured weights that allow llama-cli to use SVE 256 and MATMUL_INT8 support.
This requantization is optimal only for Graviton3. For Graviton2, requantization should optimally be done in Q4_0_4_4 format, and for Graviton4, Q4_0_4_8 is the optimal requantization format.
First, run the pre-quantized llama-2-7b-chat model exactly as the weights were downloaded from huggingface:
./llama-cli -m llama-2-7b-chat.Q4_0.gguf -p "Building a visually appealing website can be done in ten simple steps:" -n 64 -t 2
This command will use the downloaded model (-m flag), with the specified prompt (-p flag), and target a 64 token completion (-n flag), using two threads (-t flag).
You will see lots of interesting statistics being printed from llama.cpp about the model and the system, followed by the prompt and completion. The tail of the output from running this model on an AWS Graviton3 c7g.2xlarge instance is shown below:
llm_load_tensors: ggml ctx size = 0.14 MiB
llm_load_tensors: CPU buffer size = 3647.87 MiB
..................................................................................................
llama_new_context_with_model: n_ctx = 4096
llama_new_context_with_model: n_batch = 2048
llama_new_context_with_model: n_ubatch = 512
llama_new_context_with_model: flash_attn = 0
llama_new_context_with_model: freq_base = 10000.0
llama_new_context_with_model: freq_scale = 1
llama_kv_cache_init: CPU KV buffer size = 2048.00 MiB
llama_new_context_with_model: KV self size = 2048.00 MiB, K (f16): 1024.00 MiB, V (f16): 1024.00 MiB
llama_new_context_with_model: CPU output buffer size = 0.12 MiB
llama_new_context_with_model: CPU compute buffer size = 296.01 MiB
llama_new_context_with_model: graph nodes = 1030
llama_new_context_with_model: graph splits = 1
system_info: n_threads = 2 / 8 | AVX = 0 | AVX_VNNI = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 0 | NEON = 1 | SVE = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 0 | SSSE3 = 0 | VSX = 0 | MATMUL_INT8 = 1 | LLAMAFILE = 0 |
sampling:
repeat_last_n = 64, repeat_penalty = 1.000, frequency_penalty = 0.000, presence_penalty = 0.000
top_k = 40, tfs_z = 1.000, top_p = 0.950, min_p = 0.050, typical_p = 1.000, temp = 0.800
mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampling order:
CFG -> Penalties -> top_k -> tfs_z -> typical_p -> top_p -> min_p -> temperature
generate: n_ctx = 4096, n_batch = 2048, n_predict = 64, n_keep = 1
Building a visually appealing website can be done in ten simple steps:
Einzeln, a UX/UI Designer at Designhill, provides a list of ten simple steps to create a visually appealing website. These steps include using high-quality images, choosing a consistent color scheme, and incorporating negative space. Additionally, Using a clean and simple layout, creating a clear hierarchy
llama_print_timings: load time = 1120.85 ms
llama_print_timings: sample time = 2.11 ms / 64 runs ( 0.03 ms per token, 30303.03 tokens per second)
llama_print_timings: prompt eval time = 1998.79 ms / 16 tokens ( 124.92 ms per token, 8.00 tokens per second)
llama_print_timings: eval time = 15991.48 ms / 63 runs ( 253.83 ms per token, 3.94 tokens per second)
llama_print_timings: total time = 17996.97 ms / 79 tokens
The system_info printed from llama.cpp highlights important architectural features present on your hardware that improve the performance of the model execution. In the output shown above from running on an AWS Graviton3 instance, you will see:
The end of the output shows several model timings:
You can compare these timings to the optimized model weights by running
./llama-cli -m llama-2-7b-chat.Q4_0_8_8.gguf -p "Building a visually appealing website can be done in ten simple steps:" -n 64 -t 2
This is the same command as before, but with the model file swapped out for the re-quantized file.
The timings on this one look like:
llama_print_timings: load time = 984.78 ms
llama_print_timings: sample time = 2.12 ms / 64 runs ( 0.03 ms per token, 30245.75 tokens per second)
llama_print_timings: prompt eval time = 463.98 ms / 16 tokens ( 29.00 ms per token, 34.48 tokens per second)
llama_print_timings: eval time = 6890.95 ms / 63 runs ( 109.38 ms per token, 9.14 tokens per second)
llama_print_timings: total time = 7362.13 ms / 79 tokens
As you can see, load time improves, but the biggest improvement can be seen in eval times. The number of tokens per second for prompt eval quadruples, while the speed of inference more than doubles.
You have successfully run a LLM chatbot with Arm optimizations, all running on your Arm AArch64 CPU on your server. You can continue experimenting and trying out the model with different prompts.