Top Builders

Explore the top contributors showcasing the highest number of app submissions within our community.

ROCm

ROCm (Radeon Open Compute) is AMD's open-source software platform for GPU-accelerated computing. It is the AMD equivalent of NVIDIA's CUDA and provides a complete stack for running AI, machine learning, and HPC workloads on AMD GPUs. ROCm supports major ML frameworks including PyTorch, TensorFlow, JAX, and ONNX Runtime, and includes the HIP (Heterogeneous-compute Interface for Portability) programming model for writing GPU code that runs on both AMD and NVIDIA hardware.

General
AuthorAMD
TypeOpen-source GPU Computing Platform
DocumentationROCm Docs
Repositorygithub.com/ROCm
InstallationROCm Installation Guide
Current VersionROCm 7
LicenseMIT and Apache 2.0

Start building with ROCm

ROCm gives you a complete software stack to run AI training and inference workloads on AMD GPUs. It integrates directly with PyTorch, TensorFlow, and JAX so most standard pipelines run with minimal changes from a CUDA environment. Hugging Face Optimum-AMD and vLLM both support ROCm, making it straightforward to run transformer inference and fine-tuning jobs on AMD hardware. Check out the community-built AMD Use Cases and Applications to see what developers are running on ROCm today.

ROCm Tutorials

Documentation and Resources

Framework Support

  • PyTorch Full support for training and inference, including integration with Hugging Face Accelerate and PEFT
  • TensorFlow GPU-accelerated training and inference on AMD hardware
  • JAX Supported via the ROCm JAX build
  • ONNX Runtime Cross-framework model deployment on AMD GPUs
  • Hugging Face Optimum-AMD Optimized inference and fine-tuning pipelines for transformer models
  • vLLM High-throughput LLM serving with a ROCm backend

Libraries

  • hipBLAS BLAS implementation for AMD GPUs
  • MIOpen Deep learning primitives library for AMD GPUs
  • rocRAND Random number generation for AMD hardware
  • hipSPARSE Sparse matrix operations on AMD GPUs
  • rocBLAS BLAS implementation optimized for AMD Instinct accelerators
Edit on GitHub

amd AMD ROCm AI technology Hackathon projects

Discover innovative solutions crafted with amd AMD ROCm AI technology, developed by our community members during our engaging hackathons.

AuditAgent: AI Smart-Contract Oracle

AuditAgent: AI Smart-Contract Oracle

AuditAgent is an autonomous AI smart-contract audit oracle running a four-step loop: FETCH, REASON, SCORE, and ACT, recording every verdict permanently on-chain. The agent fetches any Solidity contract source or bytecode, runs 11 static vulnerability rules covering reentrancy, access-control, oracle manipulation, tx.origin misuse, selfdestruct risks, and integer precision, then augments findings with LLM semantic reasoning via Fireworks AI running on AMD Instinct MI300X GPUs in AMD Developer Cloud. The LLM catches semantic vulnerabilities the static pass misses: broken business logic, missing input validation, unsafe upgrade patterns, and economic exploits. The combined risk score from 0 to 100 and all findings are signed by an agent wallet and written to an on-chain AttestationRegistry contract, creating a permanent queryable verifiable record of AI security assessments. AMD integration: all LLM inference runs through Fireworks AI OpenAI-compatible endpoint backed by AMD Instinct MI300X hardware. The model is configurable via FIREWORKS_MODEL. Keyless fallback ensures the agent works without an API key, with static analysis alone proving the on-chain architecture is the core product.

ATLAS AI

ATLAS AI

Atlas AI is a next-generation AI-powered travel planning platform designed to simplify every stage of trip planning. Instead of spending hours comparing websites, searching for attractions, and organizing bookings, users can simply describe their travel preferences, budget, destination, and duration. Atlas AI analyzes this information and generates a fully personalized itinerary within seconds. The platform recommends flights, hotels, restaurants, attractions, transportation options, and local experiences while optimizing the travel plan based on budget, interests, weather conditions, and available time. Atlas AI also provides estimated trip expenses, interactive maps, packing suggestions, visa information, travel tips, and safety recommendations to help travelers make informed decisions. Built using modern AI technologies and intelligent agents, Atlas AI delivers conversational assistance that feels like interacting with a personal travel consultant. Users can modify itineraries through natural language, compare travel options, discover hidden gems, and receive recommendations tailored specifically to their preferences. The project aims to reduce the complexity of travel planning while improving personalization, convenience, and efficiency. Whether planning a solo adventure, family vacation, business trip, or group tour, Atlas AI provides an end-to-end travel experience through a simple, intuitive, and intelligent interface. By combining artificial intelligence with real-time travel information, Atlas AI transforms the traditional planning process into a fast, engaging, and highly personalized experience.

HyperContext – Portable AI Memory Across Every LLM

HyperContext – Portable AI Memory Across Every LLM

**HyperContext** is an open-source, portable AI context layer that lets users continue projects seamlessly across multiple AI assistants without repeatedly explaining the same information. Today, developers, researchers, students, and founders rely on different AI tools for different tasks—ChatGPT for brainstorming, Claude for writing, Gemini for research, Cursor for coding, and Perplexity for search. Since each assistant maintains its own isolated memory, users must repeatedly restate project goals, architecture, constraints, and previous decisions whenever they switch platforms. HyperContext solves this by creating a unified, user-owned context profile that works across AI ecosystems rather than being locked to a single provider. It captures project goals, technical decisions, constraints, uploaded documents, conversations, and research notes, storing them as structured long-term memory. Using semantic embeddings, vector search, and Retrieval-Augmented Generation (RAG), HyperContext retrieves only the most relevant context for each prompt and injects a concise summary into supported AI tools through a browser extension. This reduces repetitive prompting, lowers token usage, and ensures every AI assistant understands the project's history. The platform consists of a Context Capture Engine, Semantic Retrieval Engine, Vector Database, Browser Extension, and Timeline Manager that tracks project evolution. Built with a modular architecture, it is self-hostable, privacy-first, and fully open source, giving users complete ownership of their data. For the AMD Developer Hackathon, HyperContext leverages AMD Ryzen AI and the ROCm ecosystem to accelerate embedding generation, semantic retrieval, and local LLM inference. By combining efficient retrieval with on-device AI, HyperContext enables faster, privacy-preserving, cross-platform AI workflows—allowing multiple AI assistants to work as if they shared a single memory.

amd-hackathon-track1

amd-hackathon-track1

This project is a token-efficient routing agent built for Track 1, modeled on how enterprises control AI spend: cheap local compute handles routine work, and premium API calls are reserved for tasks that genuinely need them. Incoming tasks are classified by a zero-cost router into eight capability categories. Each task then flows through a three-tier pipeline. Tier 0 uses deterministic solvers, including a safe arithmetic evaluator for math and constraint-based solving for logic puzzles, producing exact answers at zero token cost. Tier 1 runs a quantized Gemma model locally inside the container, handling factual questions, sentiment classification, summarization, and entity recognition without any external calls. Tier 2 escalates only the hardest tasks, primarily code debugging, code generation, and complex reasoning, to Fireworks AI models with temperature zero and tightly capped outputs. The result is an agent where most tasks never touch the metered API, keeping accuracy high while driving scored token usage toward zero. Built with Python and llama.cpp, packaged as a single lightweight Docker image.

Metis

Metis

▎ Metis is a general-purpose AI agent built on one thesis: most tasks don't need a frontier model, and the ones that do should be identified before you pay for them — not after. ▎ ▎ Every task first runs on a small quantized model (Qwen2.5-3B) baked into the container, executing locally at zero API cost. Metis samples the local model several times in parallel and measures agreement: consistent answers across independent samples are strong evidence of correctness. That evidence — consensus, token-level confidence, and per-category signals — feeds an isotonic-calibrated confidence gate fitted on real observations. Only when the gate predicts the local answer would miss the accuracy bar does Metis escalate to a paid Fireworks model, chosen per task category with reasoning effort tuned so escalations cost single-digit tokens where possible. ▎ ▎ Known-hard categories (multi-step math, entity extraction) are force-escalated based on measured failure data, math gets a program-aided verification pass first, and an adaptive governor watches the runtime budget and downshifts sampling if the clock runs short. The whole system degrades gracefully: any failure at any layer falls back to the next-best answer rather than crashing. ▎ ▎ In our full 48-task rehearsal, Metis hit 95.8% judged accuracy while escalating only 29% of tasks — spending just ~1,400 total paid tokens, versus tens of thousands for a naive route-everything-to-the-cloud agent at comparable accuracy.

Grafting: Compose Domain-Expert LLMs

Grafting: Compose Domain-Expert LLMs

Fine-tuning a full model for a new domain is expensive and brittle, RAG bolted on top gives you retrieval, not real skill transfer. Grafting trains domain expertise as swappable weight deltas that snap onto a base model, the way you'd snap Lego bricks onto a baseplate. Need a legal assistant with financial reasoning bolted on? Graft the legal module onto the finance module. Need to swap out an underperforming module without retraining the whole stack? Pull it off and graft in the next version. Composed modules can move between models and teams as a portable format, closer to a marketplace of interchangeable domain parts than a retrained monolith. Under the hood, Grafting uses Tri-Partite Hard Slicing (Axis ARW) to train domain-specific additive weight deltas across medical, legal, coding, and finance domains on top of SmolLM3-3B. Our first architecture (rotated orthogonal subspaces) looked clean on paper and collapsed under real stacking, +25.8 perplexity, total logit failure once four domains shared a residual stream. Axis ARW fixes this with hard axis-aligned masks instead of math that assumes non-linearities play fair: stacking four domains now costs as little as +0.10 PPL (finance) up to +2.41 PPL (legal, the densest domain, still the problem child). We built a custom kernel in HIP (grafting.hip) from scratch for AMD MI300X to make that stacking tractable on a single GPU, moving off the original Python baseline's ~130GB footprint. This HIP remake is still in active beta, and VRAM figures are shifting as we correct upstream batching behavior in the TPHS pipeline, so we're not publishing a locked-in number here.