amd-rocm

录音转写与会议纪要 · 自部署 Web 版 · GPU 加速 · 隐私优先 · Recording transcription & meeting notes self-hosted web app

MinerU on AMD ROCm/Windows-HIP — evaluation-backed port. OmniDocBench v1.6: Linux W7900 pipeline 86.48 + VLM 95.56. Windows Strix Halo: pipeline 86.59 verified + VLM llama.cpp HIP 95.46 verified.

Autonomous global cargo pathfinder routing agent.

A local LLM inference framework, hand-tuned for the Ryzen AI Max APU. 3467 tok/s prefill, 76.7 tok/s decode on Qwen3.6-35B-A3B-FP8.

One-command OmniDocBench v1.6 evaluation setup on AMD Windows (ROCm/HIP). Model-agnostic: swap any document parsing model via adapters. PaddleOCR-VL-1.6 as reference. CDM formula scoring included.

Private local AI Photographer Agent on AMD Radeon and ROCm

QLoRA fine-tuning of Qwen2.5-1.5B on telecom support conversations using AMD MI300X. Submission for TCS × AMD AI Hackathon 2026 (FINETUNING_002).

Autonomous AI agent platform connecting private tutors and students via WhatsApp — powered by AMD ROCm + LangGraph

📖 中文小说转有声读物 CLI | AMD 消费级显卡本地运行 Qwen3-TTS + Edge TTS | 可能是 GitHub 上唯一针对 AMD 显卡的中文小说 TTS 完整方案

A neural network that grows its own brain while it trains. It uses an on-device LLM to dynamically write, validate, and hot-swap new PyTorch expert modules mid-training.

Agent-orchestrated, deterministic batch image generation on a local ComfyUI server. stdlib-only Python, resumable batches, reproducible seeds, built-in latency benchmark. Proven on an 81-card generative art production run (SDXL, AMD ROCm).

Themis (Θέμις): Autonomous AI legal & risk analysis platform powered by AMD Radeon PRO W7900D (ROCm 7.2.1, vLLM speculative decoding) and multi-agent LangGraph.

AMD GPU assistant that turns ROCm profiler output into bottleneck diagnoses and actionable code fixes.

Autonomous AI agent for volcanic emergency broadcast — instantly notifies residents via WhatsApp when Mount Semeru erupts. Powered by AMD ROCm + LangGraph

GemmaRoute: A 4-layer hybrid AI router that cuts LLM API costs by 80%, local Gemma 2B on AMD ROCm for free tiers, cloud escalation via Fireworks AI only when a quality judge demands it.