In this tutorial we will show an end to end workflow deploying large language models (LLMs) to Snapdragon® platforms such as Snapdragon® 8 Elite, Snapdragon® 8 Gen 3 (e.g., Samsung Galaxy S24 family) and Snapdragon® X Elite (e.g. Snapdragon® based Microsoft Surface Pro). We will use Qualcomm AI Hub to compile the models to QNN binaries and run them with Genie from the QNN SDK.
We will use Llama3 8B as a running example. Other LLMs from AI Hub Models will work with the same flow.
We will walk you through the follow steps:
- Get access to Llama 3 weights from Hugging Face.
- Use Qualcomm AI Hub Models to export Llama 3 using AI Hub.
- Prepare assets required by Qualcomm Genie, the inference runtime for LLMs.
- Run the LLM on device with an example prompt on Android / Windows PC with Snapdragon®.
Note that because this is a large model, it may take 1-2 hours to generate required assets.
If you have any questions, please feel free to post on AI Hub Slack channel
| Model name | Minimum Compile QNN SDK version | Supported devices |
|---|---|---|
| Llama-v2-7B-Chat | 2.27.0 | Snapdragon® 8 Elite Snapdragon® 8 Gen 3 Snapdragon® X Elite Snapdragon® X Plus |
| Llama-v3-8B-Chat | 2.27.0 | Snapdragon® 8 Elite Snapdragon® X Elite Snapdragon® X Plus |
| Llama-v3.1-8B-Chat | 2.27.7 | Snapdragon® 8 Elite |
| Llama-v3.1-8B-Chat | 2.28.0 | Snapdragon® X Elite Snapdragon® X Plus |
| Llama-v3.2-3B-Chat | 2.27.7 | Snapdragon® 8 Elite Snapdragon® 8 Gen 3 (Context length 2048) |
| Llama-v3.2-3B-Chat | 2.28.0 | Snapdragon® X Elite Snapdragon® X Plus |
| Baichuan2-7B | 2.27.7 | Snapdragon® 8 Elite |
| Qwen2-7B-Instruct | 2.27.7 | Snapdragon® 8 Elite |
| Mistral-7B-Instruct-v0.3 | 2.27.7 | Snapdragon® 8 Elite |
| Phi-3.5-Mini-Instruct | 2.29.0 | Snapdragon® 8 Elite Snapdragon® X Elite Snapdragon® 8 Gen 3 |
Device requirements:
- At least Genie SDK from QNN SDK 2.29.0 (earlier versions have issues with long prompts).
- 16GB memory or more for 7B+ or 4096 context length models.
- 12GB memory or more for 3B+ models (and you may need to adjust down context length).
Models that require 2.27.7 will be available on Snapdragon® X Elite starting from QNN SDK 2.28.0.
Setting up Hugging Face token is required only for Llama model family. Request model access on Hugging Face for Llama models. For instance, you can apply here to access Llama 3.2 2B model.
Setup Hugging Face token locally by following the instructions here.
Create a virtualenv for qai-hub-models with Python 3.10.
You can also use conda.
For clarity, we recommend creating a virtual env:
python3.10 -m venv llm_on_genie_venvIn shell session, install qai-hub-models under hub_model virtual env
source llm_on_genie_venv/bin/activate
pip install -U "qai_hub_models[llama-v3-8b-chat-quantized]"Replace llama-v3-8b-chat-quantized with the desired llama model from AI Hub
Model.
Note to replace _ with - (e.g. llama_v3_8b_chat_quantized -> llama-v3-8b-chat-quantized)
Ensure at least 80GB of memory (RAM + swap). On Ubuntu you can check it by
free -hIncrease swap size if needed.
We use qai-hub-models to adapt Huggingface Llama models for on-device inference.
Some of the models can be downloaded directly from AI Hub. For Llama, it has to be exported through AI Hub Models.
To generate the Llama assets, we will run a single command that performs the following steps:
-
Download model weights from Hugging Face. You will need to sign the Llama license if you haven't already done so.
-
Upload models to AI Hub for compilation.
-
Download compiled QNN binaries. Note that it's multiple binaries as we split the model.
Make a directory to put in all deployable assets. For this example we use
mkdir -p genie_bundleThe export command below typically takes 1-2 hours. However, it may take 3-4 hours on PyTorch versions earlier than 2.4.0. We recommend upgrading PyTorch first:
pip install torch==2.4.0This version is not yet supported in general by AI Hub Models but will work for the below export command.
Note that the export also requires a lot of memory (RAM + swap) on the host device (for Llama 3, we recommend 80 GB). If we detect that you have less memory than recommended, the export command will print a warning with instructions of how to increase your swap space.
python -m qai_hub_models.models.llama_v3_8b_chat_quantized.export --device "Snapdragon 8 Elite QRD" --skip-inferencing --skip-profiling --output-dir genie_bundleFor Snapdragon 8 Gen 3, please use --device "Snapdragon 8 Gen 3 QRD".
python -m qai_hub_models.models.llama_v3_8b_chat_quantized.export --device "Snapdragon X Elite CRD" --skip-inferencing --skip-profiling --output-dir genie_bundleNote: For older devices, you may need to adjust the context length using
--context-length <context-length>.
The genie_bundle would now contain both the intermediate models (token,
prompt) and the final QNN models (*.bin). Remove the intermediate models to
have a smaller deployable artifact:
# Remove intermediate assets
rm -rf genie_bundle/{prompt,token}Typically we recommend using the same QNN SDK version that AI Hub uses to compile the assets. You can find this version by clicking the job links posted printed by the export command. If the AI Hub version has not yet reached 2.29.0, we do recommend mismatching the compile and runtime versions in order to pick up an important fix (affecting long prompts) in 2.29.0.
However, if the Device Requirements table above requires a newer version than AI Hub uses, please use the newer version.
Go to QNN SDK and follow the installation instructions. Note that the first time after log in you would be redirected to QPM home page. Click on the link again to get to QNN download page.
If you are on a Mac laptop, we recommend using
Docker to install qpm-cli to extract the .qik file.
If successful, you'd see a message like
SUCCESS: Installed qualcomm_ai_engine_direct.Core at /opt/qcom/aistack/qairt/<version>
Set your QNN_SDK_ROOT environment variable to point to this directory. For
instance, on Linux you would run:
export QNN_SDK_ROOT=/opt/qcom/aistack/qairt/<version>Check out the AI Hub Apps repository using Git:
git clone https://github.com/quic/ai-hub-apps.gitNow copy the HTP config template:
cp ai-hub-apps/tutorials/llm_on_genie/configs/htp/htp_backend_ext_config.json.template genie_bundle/htp_backend_ext_config.jsonEdit soc_model and dsp_arch in genie_bundle/htp_backend_ext_config.json
depending on your target device (should be consistent with the --device you
specified in the export command):
| Generation | soc_model |
dsp_arch |
|---|---|---|
| Snapdragon® Gen 2 | 43 | v73 |
| Snapdragon® Gen 3 | 57 | v75 |
| Snapdragon® 8 Elite | 69 | v79 |
| Snapdragon® X Elite | 60 | v73 |
| Snapdragon® X Plus | 60 | v73 |
To download the tokenizer, go to the source model's Hugging Face page and go to "Files
and versions." You can find a Hugging Face link through the model card on
AI Hub. This will take you to the Qualcomm Hugging Face page,
which in turn will have a link to the source Hugging Face page. The file will be named tokenizer.json
and should be downloaded to the genie_bundle directory. The tokenizers are only hosted on the source Hugging Face page.
| Model name | Tokenizer | Notes |
|---|---|---|
| Llama-v2-7B-Chat | tokenizer.json | |
| Llama-v3-8B-Chat | tokenizer.json | |
| Llama-v3.1-8B-Chat | tokenizer.json | |
| Llama-v3.2-3B-Chat | tokenizer.json | |
| Qwen2-7B-Instruct | tokenizer.json | |
| Phi-3.5-Mini-Instruct | tokenizer.json | To see appropriate spaces in the output, remove lines 193-196 (Strip rule) in the tokenizer file. |
| Mistral-7B-Instruct-v0.3 | tokenizer.json |
For Baichuan2-7B, the tokenizer file can be created using the following code:
from transformers import AutoConfig, AutoTokenizer, AutoModelForCausalLM
config = AutoConfig.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_id, config=config, trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=True, trust_remote_code=True)
custom_tokenizer = AutoTokenizer(models.BPE(vocab=tokenizer.get_vocab(), merges=[]))
custom_tokenizer.save("tokenizer.json")Please run (replacing llama_v3_8b_chat_quantized with the desired model id):
cp ai-hub-apps/tutorials/llm_on_genie/configs/genie/llama_v3_8b_chat_quantized.json genie_bundle/genie_config.jsonFor Windows laptops, please set use-mmap to false.
If you customized context length by adding --context-length to the export
command, please open genie_config.json and modify the "size" option (under
"dialog" -> "context") to be consistent.
In genie_bundle/genie_config.json, also ensure that the list of bin files in
ctx-bins matches with the bin files under genie_bundle. Genie will look for
QNN binaries specified here.
Copy Genie's shared libraries and executable to our bundle.
cp $QNN_SDK_ROOT/lib/hexagon-v73/unsigned/* genie_bundle
cp $QNN_SDK_ROOT/lib/aarch64-windows-msvc/* genie_bundle
cp $QNN_SDK_ROOT/bin/aarch64-windows-msvc/genie-t2t-run.exe genie_bundle# For 8 Gen 2
cp $QNN_SDK_ROOT/lib/hexagon-v73/unsigned/* genie_bundle
# For 8 Gen 3
cp $QNN_SDK_ROOT/lib/hexagon-v75/unsigned/* genie_bundle
# For 8 Elite
cp $QNN_SDK_ROOT/lib/hexagon-v79/unsigned/* genie_bundle
# For all devices
cp $QNN_SDK_ROOT/lib/aarch64-android/* genie_bundle
cp $QNN_SDK_ROOT/bin/aarch64-android/genie-t2t-run genie_bundleYou have two options to run the LLM on device:
- Use the
genie-t2t-runCLI command - Use the CLI Windows ChatApp (Windows only)
All the LLMs have different formats. To get sensible output from the LLMs, it is important to use the correct prompt format for the model. These can also be found on the Hugging Face repository for each of the model. Adding samples for a few models here.
| Model name | Sample Prompt |
|---|---|
| Llama-v2-7B-Chat | <s>[INST] <<SYS>>You are a helpful AI Assistant.<</SYS>>[/INST]</s><s>[INST]What is France's capital?[/INST] |
| Llama-v3-8B-Chat Llama-v3.1-8B-Chat Llama-v3.2-3B-Chat |
<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\nWhat is France's capital?<|eot_id|><|start_header_id|>assistant<|end_header_id|> |
| Qwen2-7B-Instruct | <|im_start|>system\nYou are a helpful AI Assistant<|im_end|><|im_start|>What is France's capital?\n<|im_end|>\n<|im_start|>assistant\n |
| Phi-3.5-Mini-Instruct | <|system|>\nYou are a helpful assistant. Be helpful but brief.<|end|>\n<|user|>What is France's capital?\n<|end|>\n<|assistant|>\n |
| Mistral-7B-Instruct-v0.3 | <s>[INST] You are a helpful assistant\n\nTranslate 'Good morning, how are you?' into French.[/INST] |
In Powershell, navigate to the bundle directory and run
./genie-t2t-run.exe -c genie_config.json -p "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\nWhat is France's capital?<|eot_id|><|start_header_id|>assistant<|end_header_id|>"Note that this prompt format is specific to Llama 3.
Copy genie_bundle from the host machine to the target device using ADB and
open up an interactive shell on the target device:
adb push genie_bundle /data/local/tmp
adb shellOn device, navigate to the bundle directory:
cd /data/local/tmp/genie_bundleSet LD_LIBRARY_PATH to the current directory:
export LD_LIBRARY_PATH=$PWDThen run:
./genie-t2t-run -c genie_config.json -p "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\nWhat is France's capital?<|eot_id|><|start_header_id|>assistant<|end_header_id|>"Using libGenie.so version 1.1.0
[WARN] "Unable to initialize logging in backend extensions."
[INFO] "Using create From Binary List Async"
[INFO] "Allocated total size = 323453440 across 10 buffers"
[PROMPT]: <|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\nWhat is France's capital?<|eot_id|><|start_header_id|>assistant<|end_header_id|>
[BEGIN]: \n\nFrance's capital is Paris.[END]
[KPIS]:
Init Time: 6549034 us
Prompt Processing Time: 196067 us, Prompt Processing Rate : 86.707710 toks/sec
Token Generation Time: 740568 us, Token Generation Rate: 12.152884 toks/sec
We provide a sample C++ app to show how to build an application using the Genie SDK. See CLI Windows ChatApp for more details.
We provide a sample Android (Java and C++ app) to show how to build an application using the Genie SDK for mobile. See Android ChatApp for more details.