rdna4

Windows-only version of ComfyUI which uses AMD's official ROCm and PyTorch libraries to get better performance with AMD GPUs. [auto-installation and popular performance enhancing packages like triton * sage-attention * flash-attention * bitsandbytes included ]

Quantized attention for AMD RX 9070 / 9070 XT (RDNA4, gfx1201): a hand-written HIP attention kernel behind SageAttention, serving the fused q/k/v views video models send, at 0.87-0.94x PR #368's time over production sizes.

让 MinerU 在 AMD 显卡上以 vLLM + hybrid-auto-engine 满血运行！ 本项目是对官方文档中 AMD/ROCm 生态支持的有效补充。针对 MinerU 3.x 提供了一套完整的 ROCm 7.x + PyTorch 2.11.0 + vLLM 源码编译与部署方案，解析质量与速度对标 NVIDIA 旗舰显卡。

vLLM for AMD RDNA4 (gfx1201): Radeon AI PRO R9700 & RX 9070 XT — native MXFP4 linear kernel + spec-decode verify fix, carried as rebased branches (fork-carry model)

Run Gemma-4-31B at full 256K context on a $1,400 AMD RDNA4 GPU (gfx1201): TurboQuant KV cache + HIP-graph-safe Flash-Attention for llama.cpp, fully measured on real hardware.

Evidence-driven reverse engineering of AMD FSR 4.1: DXIL structure, neural weights, dispatch behavior, reproducible tooling and strict claim verification.

Qwen3.6-27B-FP8 at 256k context on dual AMD RDNA4 (gfx1201) — working vLLM recipe, audited measurement harness, and the full debugging record

AWQ/W4A16 quantization, benchmarking and public release of Polish LLMs (PLLuM, Bielik) on AMD RDNA 4 (ROCm + vLLM-native) — reproducible methodology and hardware-envelope studies.

Reproducible Docker + LiteLLM stack for running Qwen 3.6 27B with MTP on AMD RDNA4 (gfx1201). Verified on Radeon AI PRO R9700 32 GB. Built from llama.cpp PR #22673.

RAM-Backed MCP Memory Architecture for Consumer LLM Inference — 900K token context on 16GB VRAM

Native 3D Gaussian Splatting (gsplat) training on AMD RDNA4 / gfx1201 GPUs (RX 9070, RX 9070 XT, Radeon AI PRO R9700) via ROCm 7.2.1 + PyTorch — wave32-correct kernels, Dockerized, multi-GPU ready.

Deep Vulkan vs HIP/ROCm auto-benchmarker for AMD local LLMs (Ollama + LM Studio). Warm-up, median of 3 runs, VRAM unload, telemetry.

Qwen3.8-27B EXL3 on AMD gfx1201: RX 9070 XT (16 GB, tested) and Radeon AI PRO R9700 (32 GB). vLLM plugin with custom kernels: ~80 tok/s with MTP speculative decoding, int8 KV cache, 32k-64k context on 16 GB (up to 262k on 32 GB), vision and tool calls.

Visualize and monitor ROCm-SMI data locally or remotely via REST API — a more convenient alternative to rocm-smi.

Implicit-GEMM Conv3d for AMD RDNA4 (gfx1201) via Triton WMMA

在 8GB A 卡（RX 9060 XT / RDNA4）上跑 Qwen-Image-2.1：一键安装、完整实测数据、17 个被否掉的提速方向、13 张附提示词的图库。

Field notes from running LLMs on AMD Radeon AI PRO R9700 (RDNA4): engines, backends, multi-GPU, speculative decoding, power, model qualification, and every mistake along the way.

Validated AMD gsplat 1.5.3b2 port for ROCm 7.2, PyTorch 2.13 and RDNA 4 gfx1201, with Wave32 fixes and tested 3DGS/2DGS paths.

Full local AI stack on AMD ROCm — no NVIDIA, no CUDA. Self-hosted LLM chat, image generation, web search & agent tools on one GPU.

AI Compute Performance Analysis & Statistics Suite - RDNA4 AI Compute Toolset for Windows