Iris Coleman
Jul 27, 2024 05:35
NVIDIA and Mistral have developed NeMo 12B, a high-performance language model optimized to run on a single GPU, enhancing text-generation applications.
NVIDIA, in collaboration with Mistral, has unveiled the Mistral NeMo 12B, a groundbreaking language model that promises leading performance across various benchmarks. This advanced model is optimized to run on a single GPU, making it a cost-effective and efficient solution for text-generation applications, according to the NVIDIA Technical Blog.
Mistral NeMo 12B
The Mistral NeMo 12B model is a dense transformer model with 12 billion parameters, trained on a vast multilingual vocabulary of 131,000 words. It excels in a wide range of tasks, including common sense reasoning, coding, math, and multilingual chat. The model’s performance on benchmarks such as HellaSwag, Winograd, and TriviaQA highlights its superior capabilities compared to other models like Gemma 2 9B and Llama 3 8B.
| Model | Context Window | HellaSwag (0-shot) | Winograd (0-shot) | NaturalQ (5-shot) | TriviaQA (5-shot) | MMLU (5-shot) | OpenBookQA (0-shot) | CommonSenseQA (0-shot) | TruthfulQA (0-shot) | MBPP (pass@1 3-shots) |
| Mistral NeMo 12B | 128k | 83.5% | 76.8% | 31.2% | 73.8% | 68.0% | 60.6% | 70.4% | 50.3% | 61.8% |
| Gemma 2 9B | 8k | 80.1% | 74.0% | 29.8% | 71.3% | 71.5% | 50.8% | 60.8% | 46.6% | 56.0% |
| Llama 3 8B | 8k | 80.6% | 73.5% | 28.2% | 61.0% | 62.3% | 56.4% | 66.7% | 43.0% | 57.2% |
With a 128K context length, Mistral NeMo can process extensive and complex information, resulting in coherent and contextually relevant outputs. The model is trained on Mistral’s proprietary dataset, which includes a significant amount of multilingual and code data, enhancing feature learning and reducing bias.
Optimized Training and Inference
The training of Mistral NeMo is powered by NVIDIA Megatron-LM, a PyTorch-based library that provides GPU-optimized techniques and system-level innovations. This library includes core components such as attention mechanisms, transformer blocks, and distributed checkpointing, facilitating large-scale model training.
For inference, Mistral NeMo leverages TensorRT-LLM engines, which compile the model layers into optimized CUDA kernels. These engines maximize inference performance through techniques like pattern matching and fusion. The model also supports inference in FP8 precision using NVIDIA TensorRT-Model-Optimizer, making it possible to create smaller models with lower memory footprints without sacrificing accuracy.
The ability to run the Mistral NeMo model on a single GPU improves compute efficiency, reduces costs, and enhances security and privacy. This makes it suitable for various commercial applications, including document summarization, classification, multi-turn conversations, language translation, and code generation.
Deployment with NVIDIA NIM
The Mistral NeMo model is available as an NVIDIA NIM inference microservice, designed to streamline the deployment of generative AI models across NVIDIA’s accelerated infrastructure. NIM supports a wide range of generative AI models, offering high-throughput AI inference that scales with demand. Enterprises can benefit from increased token throughput, which directly translates to higher revenue.
Use Cases and Customization
The Mistral NeMo model is particularly effective as a coding copilot, providing AI-powered code suggestions, documentation, unit tests, and error fixes. The model can be fine-tuned with domain-specific data for higher accuracy, and NVIDIA offers tools for aligning the model to specific use cases.
The instruction-tuned variant of Mistral NeMo demonstrates strong performance across several benchmarks and can be customized using NVIDIA NeMo, an end-to-end platform for developing custom generative AI. NeMo supports various fine-tuning techniques such as parameter-efficient fine-tuning (PEFT), supervised fine-tuning (SFT), and reinforcement learning from human feedback (RLHF).
Getting Started
To explore the capabilities of the Mistral NeMo model, visit the Artificial Intelligence solution page. NVIDIA also offers free cloud credits to test the model at scale and build a proof of concept by connecting to the NVIDIA-hosted API endpoint.
Image source: Shutterstock
