Automated Model Fine-Tuning and Deployment with TAO Agent Skills

7/23/2026

This post details the integration of Prime Intellect Lab with NVIDIA Nemotron 3 Nano to demonstrate a streamlined workflow for customizing foundation models. It showcases how to set up hosted reinforcement learning, evaluate a baseline model on a Python Math environment, and produce a downloadable LoRA adapter in a short timeframe. The post emphasizes the accessibility of customization for open models and highlights the use of RLVR for improving model performance on specific tasks.

7/14/2026

This post details lessons learned from a large-scale Kaggle competition focused on improving AI reasoning with NVIDIA's Nemotron models. It highlights practical engineering workflows for enhancing model reasoning, including making chain-of-thought data verifiable, designing reasoning to fit token budgets, and separating reusable knowledge from problem-solving. These insights contribute to the broader theme of automating and optimizing the fine-tuning and deployment of foundation models by providing concrete strategies for improving their practical reasoning capabilities.

7/14/2026

This post details the application of NVIDIA TAO agent skills and LoRA for post-training the NVIDIA Cosmos 3 Nano model for video question answering. It demonstrates how these tools can automate dataset handling, baseline evaluation, LoRA configuration, and hyperparameter optimization via TAO AutoML, achieving over 90% accuracy in under a day. It also highlights the deployment of post-trained models via Cosmos 3 Reasoner NIM as OpenAI-compatible endpoints.

7/10/2026

This post details the acceleration of end-to-end co-folding performance using the NVIDIA BioNeMo Agent Toolkit. It highlights specific optimizations for each stage of the biomolecular structure prediction pipeline: MMseqs2-GPU for MSA generation (up to 177x faster than CPU JackHMMER), cuEquivariance for co-folding inference (up to 3x faster and extending sequence length to ~5.9k tokens), and further optimizations via the OpenFold3 NIM (up to 6,400 tokens on a single B300). It also introduces Fold-CP for context-parallel inference, enabling the modeling of large assemblies (up to 32,000 tokens on 64 B300 GPUs). These components are orchestrated by the BioNeMo Agent Toolkit, demonstrating an end-to-end accelerated workflow for drug discovery and protein design.

6/15/2026

This post details the application of Low-Rank Adaptation (LoRA) for parameter-efficient fine-tuning of large biological foundation models (ESM2-3B and Evo2-1B) using NVIDIA BioNeMo Recipes. It demonstrates achieving state-of-the-art performance on protein secondary structure prediction and DNA splice-site classification with significantly reduced trainable parameters (~1%) and computational resources, specifically on a single NVIDIA RTX 6000 Blackwell Workstation Edition GPU. The post highlights the integration of Transformer Engine (TE) and sequence packing for throughput optimization, and provides code snippets for loading TE-compatible models with LoRA adapters.

6/12/2026

This post details the deployment and optimization of DiffusionGemma, a diffusion-based text generation model, on NVIDIA platforms. It highlights the model's parallel token generation capability for high throughput, achieving up to 1000 tokens/sec on an H100 GPU. The post provides practical guidance on accessing and deploying DiffusionGemma through Hugging Face Transformers, NVIDIA NIM (with containerized inference and OpenAI-compatible API), and NVIDIA NeMo AutoModel for fine-tuning. It also specifies supported hardware (H100, DGX Spark, DGX Station, RTX/RTX PRO) and precision formats (BF16, NVFP4), and includes code examples for NIM deployment and testing.

6/8/2026

This post introduces the NVFP4 training recipe for JAX, implemented in MaxText, which enables high-throughput, 4-bit mixed-precision pre-training on NVIDIA Blackwell and Rubin platforms. It details the five core techniques used to preserve convergence with negligible accuracy loss: micro block scaling, E4M3 block scale factors, selective Random Hadamard Transform for WGRAD inputs, 2D FP8 scaling per 16x16 weight block, and stochastic rounding for unbiased quantization. The post also provides guidance on enabling NVFP4 in MaxText and presents performance results showing up to 1.73x speedup over FP8 baselines on GB300 hardware.

5/27/2026

This post introduces multilingual models (Qwen 3.5 4B, Riva Parakeet TDT 600M ASR, Chatterbox Multilingual 500M) and the NVIDIA In-Game Inferencing (NVIGI) SDK 1.6 update, enabling multilingual, autonomous game characters. It also details the DLSS 4.5 Unreal Engine plugin with Dynamic Multi Frame Generation, a 6x mode, and the second-generation transformer for Super Resolution. Updates to the NVIDIA RTX Branch of Unreal Engine (NvRTX) 5.7.4 are also covered, focusing on stability and compatibility improvements.