Llama3 on a shoestring Part 2: Upgrading the CPU

Generated locally on a shoestring Stable Diffusion

In Part 1, while llama3 runs the resulting chatbot is frustratingly slow. A key problem is the pre-built docker container which requires the CPU to have support for AVX instructions before it would use the GPU. If you have a GPU you can do without an AVX CPU, but this requires a rebuild from source code like what I did when installing GPU support for tensorflow on the same CPU.

Docker provided a very quick and tempting way to test various LLM models without interfering with other python installs, so it was worth having a quick look at what AVX is.

AVX instructions from 2011 CPUs

 

AVX is Avanced Vector Extensions, first shipped on Intel CPUs in 2011. My AMD Phenom II X4 was bought in 2009 and thus missed the boat. Now the Phenom II uses an AM3+ socket, so there is hope that a later AM3+ CPU might have AVX support. This turned out to be the AMD Bulldozer series. These are sold under the AMD FX-4000 to 8000 series and support AVX.

AMD Bulldozer FX-4000 to FX-8000 series

Incredibly they are still on sale online with a China vendor offering FX-4100 for just RM26.40 (about USD6) up to an FX-6350 for RM148.50 (USD34). That fits my idea of a shoestring budget so I plomped for the mid-range FX-6100 at  RM49.50 (USD11.50).

 

AMD FX-6100 is now just RM49.50

The next thing to do is to check if my equally ancient mainboard supports the FX-6100. This was the Asus M5A78LE. The manual says it does support the FX series. 

And since LLM programs require lots of memory, I might as well push my luck and fill it up. The M5A78LE  takes a maximum of 32GB DDR3 DRAM, twice my current 16GB. I picked up 8GB x 4 Kingston Hyper X Fury Blue (ie 1600MHz) for RM181.5 (USD42) so the whole upgrade cost me RM231 (USD 53).

 

Happily both worked without trouble, and where it previously failed, now the gpu-enabled docker container ran:

$docker run -it --rm --gpus=all -v /home/heong/ollama:/root/.ollama:z -p 11434:11434 --name ollama ollama/ollama

2024/10/12 07:32:48 routes.go:1153: INFO server config env="map[CUDA_VISIBLE_DEVICES: GPU_DEVICE_ORDINAL: HIP_VISIBLE_DEVICES: HSA_OVERRIDE_GFX_VERSION: HTTPS_PROXY: HTTP_PROXY: NO_PRO

XY: OLLAMA_DEBUG:false OLLAMA_FLASH_ATTENTION:false OLLAMA_GPU_OVERHEAD:0 OLLAMA_HOST:http:/

/0.0.0.0:11434 OLLAMA_INTEL_GPU:false OLLAMA_KEEP_ALIVE:5m0s OLLAMA_LLM_LIBRARY: OLLAMA_LOAD_TIMEOUT:5m0s OLLAMA_MAX_LOADED_MODELS:0 OLLAMA_MAX_QUEUE:512 OLLAMA_MODELS:/root/.ollama/models OLLAMA_NOHISTORY:false OLLAMA_NOPRUNE:false OLLAMA_NUM_PARALLEL:0 OLLAMA_ORIGINS:[http://localhost https://localhost http://localhost:* https://localhost:* http://127.0.0.1 https://127.0.0.1 http://127.0.0.1:* https://127.0.0.1:* http://0.0.0.0 https://0.0.0.0 http://0.0.0.0:* https://0.0.0.0:* app://* file://* tauri://*] OLLAMA_SCHED_SPREAD:false OLLAMA_TMPDIR: ROCR_VISIBLE_DEVICES: http_proxy: https_proxy: no_proxy:]"

time=2024-10-12T07:32:48.673Z level=INFO source=images.go:753 msg="total blobs: 11"

time=2024-10-12T07:32:48.820Z level=INFO source=images.go:760 msg="total unused blobs remove

d: 0"

time=2024-10-12T07:32:48.822Z level=INFO source=routes.go:1200 msg="Listening on [::]:11434

(version 0.3.12)"

time=2024-10-12T07:32:48.885Z level=INFO source=common.go:49 msg="Dynamic LLM libraries" runners="[cpu_avx cpu_avx2 cuda_v11 cuda_v12 cpu]"

time=2024-10-12T07:32:48.907Z level=INFO source=gpu.go:199 msg="looking for compatible GPUs"

time=2024-10-12T07:32:49.342Z level=INFO source=types.go:107 msg="inference compute" id=GPU-

49ab809b-7b47-3fd0-60c1-f03c4a8959bd library=cuda variant=v12 compute=8.6 driver=12.6 name="

NVIDIA GeForce RTX 3060" total="11.7 GiB" available="11.2 GiB"

You can query it with curl:

$curl http://localhost:11434/api/generate -d '{"model": "llama2","prompt": "Tell me about Jeeves the butler","stream": true,"options": {"seed": 123,"top_k": 20,"top_p": 0.9,"temperature": 0}}'

And the speed went up quite a bit. 

Llama 3 on a Shoestring Part 1 of 2: 2011-vintage 3GHz AMD Phenom II 16GB RAM and RTX3060 12GB

Llama working at his workstation. This image was generated locally using Stable Diffusion on a 2011 desktop with an Nvidia RTX3060 12GB GPU

Llama 3 is an 'AI model', ie a Large Language Deep Learning model comparable to Google Gemini 3.

 Sean Zheng's excellent post details a very quick way of installing and running Llama3 from a local desktop. He had good results with an Intel i9 with 128GB RAM and an Nvidia RTX 4090 with 24GB VRAM. However, my desktop dates back to 2011 and is just a 3GHz AMD Phenom II with only 16GB DRAM and an Nvidia RTX 3060 GPU with 12GB VRAM. The hope is since the RTX3060 is not too far behind his RTX 4090, Llama3 can run or maybe hobble along in some fashion.

Sean's desktop runs Red Hat's RHEL9.3 but mine runs Ubuntu 22.04LTS. Both of us had already installed Nvidia graphics drivers as well as the CUDA Toolkit. In my case the driver is 560.35.03 and CUDA is 12.6. Sean's method was to run llama3 from a Docker image. This is a excellent sandbox for a beginner like me to try out Llama3, and not risk upsetting other large AI installs like Stable Diffusion or Keras. 

Sean's post is mostly complete, the instructions are replicted here for convenience. First the system updates:

$sudo apt update

$sudo apt upgrade

We then need to update the ubuntu repository for docker:

$sudo apt install apt-transport-https ca-certificates curl software-properties-common

~$curl -fsSL https://download.docker.com/linux/ubuntu/gpg | sudo gpg --dearmor -o /usr/share/keyrings/docker-archive-keyring.gpg

$echo "deb [arch=$(dpkg --print-architecture) signed-by=/usr/share/keyrings/docker-archive-keyring.gpg] https://download.docker.com/linux/ubuntu $(lsb_release -cs) stable" | sudo tee /etc/apt/sources.list.d/docker.list > /dev/null

Then the actual Docker install:

$sudo apt update

$apt-cache policy docker-ce

$sudo apt install docker-ce

And I have a running docker daemon:

$sudo systemctl status docker

�.. docker.service - Docker Application Container Engine

     Loaded: loaded (/lib/systemd/system/docker.service; enabled; vendor preset>

     Active: active (running) since Thu 2024-09-26 11:05:47 +08; 22s ago

TriggeredBy: �.. docker.socket

       Docs: https://docs.docker.com

   Main PID: 56585 (dockerd)

      Tasks: 10

     Memory: 22.2M

        CPU: 729ms

     CGroup: /system.slice/docker.service

             �..�..56585 /usr/bin/dockerd -H fd:// --containerd=/run/containerd/

con>

A quick test seems fine:

$docker run hello-world

Unable to find image 'hello-world:latest' locally

latest: Pulling from library/hello-world

c1ec31eb5944: Pull complete

Digest: sha256:91fb4b041da273d5a3273b6d587d62d518300a6ad268b28628f74997b93171b2

Status: Downloaded newer image for hello-world:latest

Hello from Docker!

This message shows that your installation appears to be working correctly.

Next I just use Docker to pull in ollama:

$docker run -d -v ollama:/root/.ollama -p 11434:1

1434 --name ollama ollama/ollama

Unable to find image 'ollama/ollama:latest' locally

latest: Pulling from ollama/ollama

Digest: sha256:e458178cf2c114a22e1fe954dd9a92c785d1be686578a6c073a60cf259875470

Status: Downloaded newer image for ollama/ollama:latest

c09a5a60d5aa9120175c52f7b13b59420564b126005f4e90da704851bbeb9308

A quick check shows everything seems OK:

$docker ps -a

CONTAINER ID   IMAGE           COMMAND               CREATED         STATUS

              PORTS                                           NAMES

c09a5a60d5aa   ollama/ollama   "/bin/ollama serve"   9 minutes ago   Up 9 minute

s             0.0.0.0:11434->11434/tcp, :::11434->11434/tcp   ollama

75beaa5bac23   hello-world     "/hello"              2 hours ago     Exited (0)

2 hours ago                                                   amazing_ptolemy

OK, now for the GPU version of Ollama. We first stop ollama:

$docker stop c09a5a60d5aa

c09a5a60d5aa

$docker rm c09a5a60d5aa

c09a5a60d5aa

Make the local directory for ollama:

$mkdir ~/ollama

Oops:

$docker run -it --rm --gpus=all -v /home/heong/ollama:/root/.ollama:z -p 11434:11434 --name ollama ollama/ollama

docker: Error response from daemon: could not select device driver "" with capabilities: [[gpu]].

I think that means it cannot find the GPU. From here, I think I need the Nvidia Container Toolkit. The install guide is here.

Update the repository:

$curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey | sudo gpg --dearmor -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg \

  && curl -s -L https://nvidia.github.io/libnvidia-container/stable/deb/nvidia-container-toolkit.list | \

    sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' | \

    sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list

$ sudo apt-get update

Now the actual install:

sudo apt-get install -y nvidia-container-toolkit

The just  restart Docker:

$ sudo systemctl restart docker

Now ollama runs:

$docker run -it --rm --gpus=all -v  /home/heong/ollama:/root/.ollama:z -p 11434:11434 --name ollama ollama/ollama

2024/09/26 13:12:23 routes.go:1153: INFO server config env="map[CUDA_VISIBLE_DEVICES: GPU_DEVICE_ORDINAL: HIP_VISIBLE_DEVICES: HSA_OVERRIDE_GFX_VERSION: HTTPS_PROXY: HTTP_PROXY: NO_PROXY: OLLAMA_DEBUG:false OLLAMA_FLASH_ATTENTION:false OLLAMA_GPU_OVERHEAD:0 OLLAMA_HOST:http://0.0.0.0:11434 OLLAMA_INTEL_GPU:false OLLAMA_KEEP_ALIVE:5m0s OLLAMA_LLM_LIBRARY: OLLAMA_LOAD_TIMEOUT:5m0s OLLAMA_MAX_LOADED_MODELS:0 OLLAMA_MAX_QUEUE:512 OLLAMA_MODELS:/root/.ollama/models OLLAMA_NOHISTORY:false OLLAMA_NOPRUNE:false OLLAMA_NUM_PARALLEL:0 OLLAMA_ORIGINS:[http://localhost https://localhost http://localhost:* https://localhost:* http://127.0.0.1 https://127.0.0.1 http://127.0.0.1:* https://127.0.0.1:* http://0.0.0.0 https://0.0.0.0 http://0.0.0.0:* https://0.0.0.0:* app://* file://* tauri://*] OLLAMA_SCHED_SPREAD:false OLLAMA_TMPDIR: ROCR_VISIBLE_DEVICES: http_proxy: https_proxy: no_proxy:]"

But then it looked like it detected my GPU but refused to use it as my CPU does not have AVX or AVX2 instructions support:

time=2024-09-26T13:12:23.496Z level=WARN source=gpu.go:224 msg="CPU does not have minimum vector extensions, GPU inference disabled" required=avx detected="no vector extensions"

time=2024-09-26T13:12:23.496Z level=INFO source=types.go:107 msg="inference compute" id=0 library=cpu variant="no vector extensions" compute="" driver=0.0 name="" total="15.6 GiB" available="13.2 GiB"

See Part 2 for the CPU upgrade to support AVX.

Now that was a setback, but ollama runs. Let's see if it loads llama 3.

$docker exec -it ollama ollama pull llama3

For good measure lets pull in llama2:

$docker exec -it ollama ollama pull llama3

$docker exec -it ollama ollama list

NAME             ID              SIZE      MODIFIED

llama3:latest    365c0bd3c000    4.7 GB    15 seconds ago

llama2:latest    78e26419b446    3.8 GB    24 hours ago

And indeed llama3 runs on a 2011 AMD CPU with just 16GB RAM:

$docker exec -it ollama ollama run llama3

>>> Send a message (/? for help)

>>> /?

Available Commands:

  /set            Set session variables

  /show           Show model information

  /load <model>   Load a session or model

  /save <model>   Save your current session

  /clear          Clear session context

  /bye            Exit

  /?, /help       Help for a command

  /? shortcuts    Help for keyboard shortcuts

Use """ to begin a multi-line message.

>>> /show info

  Model

    architecture        llama

    parameters          8.0B

    context length      8192

    embedding length    4096

    quantization        Q4_0

  Parameters

    num_keep    24

    stop        "<|start_header_id|>"

    stop        "<|end_header_id|>"

    stop        "<|eot_id|>"

  License

    META LLAMA 3 COMMUNITY LICENSE AGREEMENT

    Meta Llama 3 Version Release Date: April 18, 2024

In response to the prompt

>>> How are you today?

The reply was:

I'm just an AI, I don't have feelings or emotions like humans do. However, 

I am functioning properly and ready to assist with any questions or tasks 

you may have! Is there something specific you'd like to talk about or ask 

for help with?

It was excruciatingly slow, and nvtop show the gpu is indeed not used but ollama seems to be all there. So there you have it, Llama3 running on a 16GB AMD  Phenom II with no GPU.

Happy Trails.

The Little Computer that Could: Motion Detection with Outdoor Camera, Tensorflow and Raspberry Pi Part 3 of 3

 

"I have no spur to prick the sides of my intent, but only vaulting ambition, which o'erleaps itself and falls on the other." - Macbeth

Motion detection with an outdoor camera can be problematic, with wind, shadow and direct sunlight causing multiple false alarms. Suppressing these false alarms can result in genuine alarms being missed. One way is to filter all the alarms through an object recognition program.

The motion detection program of Part 1 will write all alarm image frames to the ./alarms/ directory. A modified object recognition program of Part 2 will inspect each alarm frame and if an object is recognized, will write it to another directory ./alarm/. The alarm filter program is called tiny-yolo_alarmfilter.py. This seemed to work well to start with, but with any new project, time will tell. For starters it seemed to think my dog is a cow, probably because she was sniffing at the grass.

tiny YOLO mislabelling dog as cow

Now my dog does have a temper, but calling her a cow is a little harsh. Even so, both dogs and cows qualify (in my opinion) as valid alarm triggers so it is not a show-stopper for now. It is particularly good at excluding changes in lighting and shadows. Proof positive that the little Raspberry Pi could, and did. 

Recognizing passing vehicles 

Ah, vaulting ambition ... if the Pi can recognize an object, maybe it can also track it. Most alarms are quite passive, except for the loud siren, which tends to annoy the neighbors and should only be used as a last resort. A pan/tilt camera like the Trendnet TV-IP422WN that visibly tracks the object is a lot more menacing, and should scare off the more timid intruders like birds and squirrels. But that is another blog post.

A tensorflow model looks promising, as it can be potentially speeded up with the use of custom hardware like the Coral USB acelerator for about the price of a Raspberry Pi 4.

Coral USB Accelerator

Installing tensorflow proved to be a bit hit and miss, but Katsuya Hyodo's github readme worked for me. This time I started with a squeaky clean version of Raspbian, 2021-05-07-raspios-buster-armhf.img.

Remember to uninstall the dud versions:

# pip3 uninstall tensorflow
# apt-get install -y libhdf5-dev libc-ares-dev libeigen3-dev gccgfortran libgfortran5 libatlas3-base libatlas-base-dev libopenblas-dev libopenblas-base libblas-dev liblapack-dev cython3 openmpi-bin libopenmpi-dev libatlas-base-dev python3-dev
# pip3 install pip --upgrade
# pip3 install keras_applications==1.0.8 --no-deps
# pip3 install keras_preprocessing==1.1.0 --no-deps
# pip3 install h5py==2.9.0
# pip3 install pybind11
# pip3 install -U --user six wheel mock
# wget "https://raw.githubusercontent.com/PINTO0309/Tensorflow-bin/master/tensorflow-1.15.0-cp37-cp37m-linux_armv7l_download.sh"
# sh ./tensorflow-1.15.0-cp37-cp37m-linux_armv7l_download.sh
# pip3 install tensorflow-1.15.0-cp37-cp37m-linux_armv7l.whl

A quick test:

# python3

Python 3.7.3 (default, Jan 22 2021, 20:04:44)

[GCC 8.3.0] on linux

Type "help", "copyright", "credits" or "license" for more information.

>>> import tensorflow

>>> tensorflow.__version__

'1.15.0'

>>>

For object detection I used Edje Electronics. 

Packages tensorflow, libatlas-base-dev, libhdf5-dev, libhdf5-serial-dev I had already installed previously

# apt-get install libjasper-dev

# apt-get install libqtgui4

# apt-get install libqt4-test

I used version 4.4.0.46 because 4.1.0.25 could not be found

# pip3 install opencv-contrib-python==4.4.0.46

# apt-get install protobuf-compiler

# pip install --user Cython

# pip install --user contextlib2

# pip install --user pillow

# pip install --user lxml

# pip install --user matplotlib

Got the tensorflow models

# git clone https://github.com/tensorflow/models.git

Then SSD_Lite:

# wget http://download.tensorflow.org/models/object_detection/ssdlite_mobilenet_v2_c

oco_2018_05_09.tar.gz

# tar -xzvf ssdlite_mobilenet_v2_coco_2018_05_09.tar.gz

The original tensorflow program, ./models/research/object_detection/TensorFlow.py got its images from the default Raspberry Pi camera, so I made a simpler version to take one frame (./341front.jpg in 640x480) at a time, uTensorFlow.py.

Note SSD_Lite misclassified a dog as sheep

The processing time was over 40s on my Raspberry Pi 3. My Coral USB accelerator will take more than a month to arrive, and it needs Tensorflow Lite, for which Edje Electronics has a very promising Tensorflow Lite repository, so why not. Notice this time the commands are as a sudoer user and not root, which I am told is the proper way to do things:

$ sudo pip3 install virtualenv

$ python3 -m venv tflite1-env

$ source tflite1-env/bin/activate

Then comes a whopper of a download:

$ git clone https://github.com/EdjeElectronics/TensorFlow-

Lite-Object-Detection-on-Android-and-Raspberry-Pi.git

$ bash get_pi_requirements.sh

Notice with Tensorflow Lite there is no Tensorflow module:

$ python3

Python 3.7.3 (default, Jan 22 2021, 20:04:44)

[GCC 8.3.0] on linux

Type "help", "copyright", "credits" or "license" for more information.

>>> import tensorflow

Traceback (most recent call last):

  File "<stdin>", line 1, in <module>

ModuleNotFoundError: No module named 'tensorflow'

>>>

$ wget https://storage.googleapis.com/download.tensorflow.org/models/tflite/coco_ssd_mobilenet_v1_1.0_quant_2018_06_29.zip

$ unzip coco_ssd_mobilenet_v1_1.0_quant_2018_06_29.zip

Archive:  coco_ssd_mobilenet_v1_1.0_quant_2018_06_29.zip

  inflating: detect.tflite

  inflating: labelmap.txt

The original program is TFLite_detection_webcam.py but I need a version that works on individual image files, and not a video file, as in tiny-yolo_alarmfilter.py. The progam is tflite_alarmfilter.py, which took no time at all to write in python. You run it thus:

$ python3 tflite_alarmfilter.py --resolution='640x480' --modeldir=.

Processing alarm file 341front_20210804_094635_original.png

Processing alarm file 341front_20210804_094649_original.png

Processing alarm file 341front_20210804_094648_original.png

The framerate is now an astounding 2fps even without the Coral USB Accelerator. The detection seems slightly better with a more accurate bounding box without multiple boxes nested in the same object.

Tensorflow Lite SSDLite MobileNet v2: note accurate bounding box

tiny-YOLO: multiple bounding boxes over same object

I had originally planned to outsource the alarm files filtering to an x86 CPU, but the results with Tensorflow Lite made everything possible on the same Raspberry Pi 3, truly the little CPU that could. 

Happy Trails.

The Little Computer that Could: YOLO Object Recognition and Raspberry Pi Part 2 of 3

 

You Only Look Once

One cure for motion detector false alarms in Part 1 is object recognition; the alarm is only raised if the intruder is recognized. YOLO is fast, trendy (deep neural net!) and will run on a Raspberry Pi. arunponnusamy comes straight to the point.

I already had numpy and openCV from Part 1, so it is just

# wget https://pjreddie.com/media/files/yolov3.weights

Now this is a massive file and I needed a copper Ethernet connection to the Internet to download it. Next you download the zip file from arunponnusamy for the files yolo_opencv.py, yolov3.cfg, yolov3.txt and dog.jpg. Do not use wget on github. And simply run it:

# python3 yolo.py --image dog.jpg --config yolov3.cfg --weights yolov3.weights --classes yolov3.txt

If you get 'OutOfMemoryError' you will probably need to reboot your Pi.

The output should be:

I substituted a frame from my IP Camera, and sure enough it recognized a person correctly:

# python3 yolo.py --image alarm.png --config yolov3.cfg --weights yolov3.weights --classes yolov3.txt

Output of yolov3.weights model

For convenience, see my github repository for copies of arunponnusamy's code. It took my Raspberry Pi 3 some 50 seconds to process the image, which is a little long but would probably keep up with the motion detector triggers of a few frames per hour. But arunponnusamy's reference is Joseph Redmon, who also mentioned a tiny version of the pre-trained neural net model. You will also need to download his yolov3-tiny.cfg, which I got by cloning from his github.

git clone https://github.com/pjreddie/darknet

Joseph's executable is a C program, but the input files are the same so I simply used his tiny model files with arunponnusamy's python script:

$ python3 yolo.py --image alarm.png --config yolov3-tiny.cfg --weights yolov3-tiny.weights --classes yolov3.txt

Note I also reused arunponnusamy's yolov3.txt which is the dictionary of object names. Joseph's dictionary seems to be baked into his executable file. The command finishes in 5 seconds, which is 10 times better. The output bounding box is different, but it did recognize the object as a person.

Output of yolov3-tiny.weights model

The idea is to use YOLO to filter out the false alarms from the motion detection program from Part 1, In Part 3, we will look at tiny-YOLO, Tensorflow, and SSD_Lite. The Raspberry Pi is truly the little computer that could.

Happy Trails.

Raspberry Pi 4 Voice Assistant Part 2 of 4: Google Text to Speech

gTTS: The Empire Strikes Back

In Part 1, one of my goals was to have my laptop issue voice commands to to my Google Home smart speaker. After trying out Mycroft, Jasper seems like the logical next step. The other text to speech systems voice quality were something like the Texas Instruments Speak & Spell products: especially ESP32Talkie. Even Mycroft sounded a bit sad next to my Home Mini speaker.

And then I stumbled upon gTTS, Google Text to Speech.  This python interface to Google's text to speech can be installed using:

pip install gTTS

And requires a program, of only ten lines:

from gtts import gTTS

def text_to_speech(input_name, output_name, language):

    file = open(input_name, 'r')

    content = file.read()

    file.close()

    sound = gTTS(text=content, lang=language)

    sound.save(output_name + '.mp3')

    #https://pypi.org/project/gTTS/

text_to_speech('input_en.txt', 'sound_en', 'en')

#text_to_speech('input_tr.txt', 'sound_tr', 'tr')

print('Done!')

You put your text in a file, input_en.txt:

$cat ./input_en.txt

Hey, Google

$python tts.py

And you get back an mp3 file, sound_en.mp3

And all you need to do now to trigger the Google Home smart speaker is:

$mplayer sound_en.mp3

A typical complete command would be something like:

$mplayer HeyGoogle.mp3; sleep 1; mplayer OfficeLampOn.mp3

For some reason, the other trigger phrase 'OK Google' did not work, but for very little effort I can now integrate disparate IoT devices, be they home brewed, Alexa, or Google Home into one Voice Assistant that rules them all.

Here's what it sounds like:

Happy Trails.

Luke: Vader... Is the dark side stronger?

Yoda: No, no, no. Quicker, easier, more seductive.