Now you can speed up your large language model (LLM) fine-tuning process using "Unsloth." Unsloth is a breakthrough library designed to work seamlessly with HuggingFace, enhancing the efficiency of LLM fine-tuning on NVIDIA GPUs. You can leverage Unsloth for fine-tuning different architectures, including "Llama" and "Mistral," using the TRL trainers (SFTTrainer, DPOTrainer, PPOTrainer). I tried to fine-tune a Mistral 7B LLM (4bit) on the IMDB dataset for text generation, all within Google Colab using Unsloth. Here's the link to my hugging face model - https://lnkd.in/gJ6Bxb9X Unsloth github - https://lnkd.in/gSxk5s96
Saurav Mohanty’s Post
More Relevant Posts
-
Curious about a generalizable, simple, small deep-learning model to reconstruct distributed acoustic sensing data? Check out Yiyu Ni's recent presentation at the Seismological Society of America with J. Nathan Kutz, Shaowu Pan, Brad Lipovsky, Qibin Shi, and myself on reconstructing DAS data using sparse sensor network strategies. The model is small, can be trained on the edge with GPUs, deployed on CPUs, and is generalizable in time for a given DAS cable. For efficient streaming, spatially subsample the DAS channel data and reconstruct the full data locally. While the model is not fully multi-scale yet, it reconstructs offshore DAS ocean waves very well, which can be removed from earthquake records and improve seismological studies, locally or on the edge. GitHub repository: https://lnkd.in/gESakh-5 Watch the presentation here: https://lnkd.in/gN8PbYrS #DAS #DeepLearning #EdgeComputing #SeismicNetwork Great work Yiyu Ni!
Yiyu Ni (UW) "Wavefield Reconstruction of Distributed Acoustic Sensing using Machine Learning"
https://www.youtube.com/
To view or add a comment, sign in
-
Our governments collective budget on AI is one tenth of this. The government needs to step up the game on AI studies with a strong focus on fundamental sciences , empirical work, industry and infrastructure ( digital ). https://lnkd.in/g_VF_Mft
Historic Vids (@historyinmemes) on X
twitter.com
To view or add a comment, sign in
-
✨ Announcement: I’m excited to share the release of version 0.1 of my new model, FeynModel! https://lnkd.in/ePJfQ_3S 🔍 FeynModel is a hybrid model that combines vision and scientific reasoning, all while being extremely lightweight. It’s ready to use with just 1GB of RAM on a CPU and two lines of code, yet it competes with the leading large language models! 💡 ⚙️ The key innovation lies in the integration of S4 blocks, enhancing contextual memory and boosting visual reasoning capabilities. For maximum performance, it’s best used on a non-quantized GPU. 💻 I invite you to test it out and share your feedback, whether on usability or technical aspects. For those interested in diving into the technical details, I’m available for discussions! 🔗 Feel free to reach out or comment on this post if you’d like to exchange ideas. I’m looking forward to hearing your thoughts!
Imagroune/feynmodel · Hugging Face
huggingface.co
To view or add a comment, sign in
-
This way Gaussian Splatting can leapfrog - big 🤘 #gaussiansplatting Gaussian Splatting #3DGS #colmap
COLMAP-Free 3D Gaussian Splatting UC San Diego, NVIDIA, UC Berkeley CVPR 2024, Seattle ✨ Highlight ✨ page: https://lnkd.in/eNQMXtnS arxiv: https://lnkd.in/ergUrMdP video: https://lnkd.in/e363juUW code: https://lnkd.in/exDrDdUJ
COLMAP-Free 3D Gaussian Splatting
oasisyang.github.io
To view or add a comment, sign in
-
🚀 Elevate your computer vision game with YOLOv8! Discover the power of the YOLOv8 Object Detection Model for customized training. Unlike previous versions, YOLOv8 revolutionizes training by incorporating negative samples alongside positive ones. Ultralytics innovative approach aims to slash false positives and prevent the model from mistaking similar background objects for the real deal. By directly including negative samples in training data, a blend of positives and negatives enhances the model's accuracy. Positive samples require annotations of at least one object, while negative samples can simply have empty annotation files. No changes are needed in the existing class/es within the data.yaml file; negative samples do not constitute a separate class. With this balanced mix in training, faster inference speeds particularly on newer hardware platforms like NVIDIA GPUs and improved scalability and accuracy, YOLOv8 delivers precise real-time object detection like never before. 🔍💡 For deeper insights, explore https://lnkd.in/gbHP2SDE 📚✨ #LearnMore #Ultralytics #ComputerVision #YOLOv8 #ObjectDetection
Home
docs.ultralytics.com
To view or add a comment, sign in
-
With multi-node GPU computing, inference, high-speed Ethernet networking, and networked storage, check out these RAG workflow examples to harness the potential of retrieval-augmented generation (#RAG) and efficiently and effectively leverage enterprise data and domain expertise. https://nvda.ws/3TpVxHK
Scaling Enterprise RAG with Accelerated Ethernet Networking and Network Storage
share.nvidia.com
To view or add a comment, sign in
-
With multi-node GPU computing, inference, high-speed Ethernet networking, and networked storage, check out these RAG workflow examples to harness the potential of retrieval-augmented generation (#RAG) and efficiently and effectively leverage enterprise data and domain expertise. https://nvda.ws/3TpVxHK
Scaling Enterprise RAG with Accelerated Ethernet Networking and Network Storage
share.nvidia.com
To view or add a comment, sign in
-
With multi-node GPU computing, inference, high-speed Ethernet networking, and networked storage, check out these RAG workflow examples to harness the potential of retrieval-augmented generation (#RAG) and efficiently and effectively leverage enterprise data and domain expertise. https://nvda.ws/3TpVxHK
Scaling Enterprise RAG with Accelerated Ethernet Networking and Network Storage
share.nvidia.com
To view or add a comment, sign in
-
recently been tinkering with running hand rolled neural nets on FPGAs for audio generation. ( it's crazy being able to run inference almost 200,000 inferences / second e.g. https://lnkd.in/gqepf6Cy ) stumbled today onto a cool idea called LUTnet ( https://lnkd.in/gi9GiH2Y ) where binary neural nets can be expressed as 2-LUTs ( generalisable to N-LUTs ). what a great idea! has anyone worked with these kinds of architectures? am particularly interested in how i can use the same ideas with my power-of-two quantisation networks (where "multiplies" are just left/right shifts )
To view or add a comment, sign in