| | --- |
| | library_name: pytorch |
| | license: other |
| | tags: |
| | - llm |
| | - generative_ai |
| | - android |
| | pipeline_tag: text-generation |
| |
|
| | --- |
| | |
| |  |
| |
|
| | # Llama-v3.2-3B-Instruct: Optimized for Qualcomm Devices |
| |
|
| | Llama 3 is a family of LLMs. The model is quantized to w4a16 (4-bit weights and 16-bit activations) and part of the model is quantized to w8a16 (8-bit weights and 16-bit activations) making it suitable for on-device deployment. |
| |
|
| | This is based on the implementation of Llama-v3.2-3B-Instruct found [here](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct/). |
| | This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the [Qualcomm® AI Hub Models](https://github.com/quic/ai-hub-models/blob/main/qai_hub_models/models/llama_v3_2_3b_instruct) library to export with custom configurations. More details on model performance across various devices, can be found [here](#performance-summary). |
| |
|
| | Qualcomm AI Hub Models uses [Qualcomm AI Hub Workbench](https://workbench.aihub.qualcomm.com) to compile, profile, and evaluate this model. [Sign up](https://myaccount.qualcomm.com/signup) to run these models on a hosted Qualcomm® device. |
| |
|
| | ## Deploying Llama 3.2 3B on-device |
| |
|
| | Please follow the [LLM on-device deployment](https://github.com/quic/ai-hub-apps/tree/main/tutorials/llm_on_genie) tutorial. |
| |
|
| | ## Getting Started |
| | Due to licensing restrictions, we cannot distribute pre-exported model assets for this model. |
| | Use the [Qualcomm® AI Hub Models](https://github.com/quic/ai-hub-models/blob/main/qai_hub_models/models/llama_v3_2_3b_instruct) Python library to compile and export the model with your own: |
| | - Custom weights (e.g., fine-tuned checkpoints) |
| | - Custom input shapes |
| | - Target device and runtime configurations |
| |
|
| | See our repository for [Llama-v3.2-3B-Instruct on GitHub](https://github.com/quic/ai-hub-models/blob/main/qai_hub_models/models/llama_v3_2_3b_instruct) for usage instructions. |
| |
|
| |
|
| | ## Model Details |
| |
|
| | **Model Type:** Model_use_case.text_generation |
| | |
| | **Model Stats:** |
| | - Input sequence length for Prompt Processor: 128 |
| | - Maximum context length: 4096 |
| | - Quantization Type: w4 + w8 (few layers) with fp16 activations and w4a16 + w8a16 (few layers) are supported |
| | - Supported languages: English. |
| | - TTFT: Time To First Token is the time it takes to generate the first response token. This is expressed as a range because it varies based on the length of the prompt. The lower bound is for a short prompt (up to 128 tokens, i.e., one iteration of the prompt processor) and the upper bound is for a prompt using the full context length (4096 tokens). |
| | - Response Rate: Rate of response generation after the first response token. |
| | |
| | ## Performance Summary |
| | | Model | Runtime | Precision | Chipset | Context Length | Response Rate (tokens per second) | Time To First Token (range, seconds) |
| | |---|---|---|---|---|---|--- |
| | | Llama-v3.2-3B-Instruct | GENIE | w4 | Snapdragon® 8 Elite Mobile | 4096 | 13.83 | 0.088195 - 2.82225 |
| | | Llama-v3.2-3B-Instruct | GENIE | w4 | Qualcomm® SA8295P | 1024 | 3.523 | 0.37311700000000003 - 2.9849360000000003 |
| | | Llama-v3.2-3B-Instruct | GENIE | w4 | Snapdragon® 8 Elite Gen 5 Mobile | 4096 | 18.00883 | 0.131546 - 4.209475 |
| | | Llama-v3.2-3B-Instruct | GENIE | w4a16 | Snapdragon® 8 Elite Mobile | 4096 | 28.03 | 0.08204900000000001 - 2.6255680000000003 |
| | | Llama-v3.2-3B-Instruct | GENIE | w4a16 | Snapdragon® X Elite | 4096 | 11.87 | 0.116884 - 3.740288 |
| | | Llama-v3.2-3B-Instruct | GENIE | w4a16 | Qualcomm® SA8775P | 4096 | 17.47 | 0.109614 - 3.507648 |
| | | Llama-v3.2-3B-Instruct | GENIE | w4a16 | Snapdragon® 8 Elite Gen 5 Mobile | 4096 | 32.65 | 0.06895399999999999 - 2.2065279999999996 |
| | | Llama-v3.2-3B-Instruct | GENIE | w4a16 | Snapdragon® X2 Elite | 4096 | 42.77 | 0.075045 - 2.40144 |
| | |
| | ## License |
| | * The license for the original implementation of Llama-v3.2-3B-Instruct can be found |
| | [here](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct/blob/main/LICENSE.txt). |
| | |
| | ## References |
| | * [LLaMA: Open and Efficient Foundation Language Models](https://www.llama.com/docs/model-cards-and-prompt-formats/llama3_2/) |
| | * [Source Model Implementation](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct/) |
| | |
| | ## Community |
| | * Join [our AI Hub Slack community](https://aihub.qualcomm.com/community/slack) to collaborate, post questions and learn more about on-device AI. |
| | * For questions or feedback please [reach out to us](mailto:ai-hub-support@qti.qualcomm.com). |
| | |
| | ## Usage and Limitations |
| | |
| | This model may not be used for or in connection with any of the following applications: |
| | |
| | - Accessing essential private and public services and benefits; |
| | - Administration of justice and democratic processes; |
| | - Assessing or recognizing the emotional state of a person; |
| | - Biometric and biometrics-based systems, including categorization of persons based on sensitive characteristics; |
| | - Education and vocational training; |
| | - Employment and workers management; |
| | - Exploitation of the vulnerabilities of persons resulting in harmful behavior; |
| | - General purpose social scoring; |
| | - Law enforcement; |
| | - Management and operation of critical infrastructure; |
| | - Migration, asylum and border control management; |
| | - Predictive policing; |
| | - Real-time remote biometric identification in public spaces; |
| | - Recommender systems of social media platforms; |
| | - Scraping of facial images (from the internet or otherwise); and/or |
| | - Subliminal manipulation |
| | |
