Trending Model:#1Inklingthinkingmachines⬇16kTrending Model:#2Ternary-Bonsai-27B-ggufprism-ml⬇432kTrending Model:#3Unlimited-OCRbaidu⬇2237kTrending Model:#4Bonsai-27B-ggufprism-ml⬇1405kTrending Model:#5GLM-5.2zai-org⬇545kTrending Model:#6Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUFDavidAU⬇63kTrending Model:#7Qwen3.6-35B-A3B-Uncensored-HauhauCS-AggressiveHauhauCS⬇1998kTrending Model:#8Qwythos-9B-Claude-Mythos-5-1M-GGUFempero-ai⬇2133kTrending Model:#9krea2-identity-editconradlocke⬇0kTrending Model:#10Laguna-S-2.1poolside⬇3kTrending Model:#1Inklingthinkingmachines⬇16kTrending Model:#2Ternary-Bonsai-27B-ggufprism-ml⬇432kTrending Model:#3Unlimited-OCRbaidu⬇2237kTrending Model:#4Bonsai-27B-ggufprism-ml⬇1405kTrending Model:#5GLM-5.2zai-org⬇545kTrending Model:#6Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUFDavidAU⬇63kTrending Model:#7Qwen3.6-35B-A3B-Uncensored-HauhauCS-AggressiveHauhauCS⬇1998kTrending Model:#8Qwythos-9B-Claude-Mythos-5-1M-GGUFempero-ai⬇2133kTrending Model:#9krea2-identity-editconradlocke⬇0kTrending Model:#10Laguna-S-2.1poolside⬇3k

New Tool ComfyUI-ROCm-Windows-Native Breathes Life Into Old AMD GPUs

By vramkickedin | | 2 min read
A glowing red AMD Radeon GPU chip made up of a luminous crystalline circuitry.

ComfyUI-ROCm-Windows-Native is a new open-source release that automates setting up ComfyUI to run directly on AMD GPUs in Windows using the native ROCm/HIP backend. It completely sidesteps Microsoft’s DirectML translation layer, letting AI image generation talk straight to the silicon. This approach avoids well-known DirectML headaches like the OpaqueTensorImpl error and missing FP8 data type support.

Developer Pedrodenovo built the project after growing tired of slow 3.5-second iterations and out-of-memory crashes on a modest RX 5500 XT. The developer packaged an automated .bat script that builds a safe virtual environment and pulls nightly multi-architecture PyTorch packages matched to a user’s exact GPU. The result opens up modern memory-saving features to AMD card owners who were previously locked out.

Bypassing DirectML for native ROCm speed

Key Features
  • Slash iteration time to 1.6s/it on older cards.
  • Cut VRAM consumption from 8GB to 4GB.
  • Unlock CPU GGUF decoding for memory savings.
  • Add native FP8 support to avoid OOM errors.
  • Simple plug-and-play .bat installation script.

This tool is for AMD GPU owners who have been stuck with slow or broken DirectML-based ComfyUI setups on Windows. Hobbyists and small studios can get usable generation speeds from older Radeon cards like the RX 5500 XT all the way up to the latest RX 9000 series. Professionals who need a fully local, privacy-first workflow avoid cloud dependencies while finally running FP8 and GGUF models natively.

Developer observations and real-world gains

The developer points out that DirectML on Windows imposes severe physical limits, blocking GGUF files and crashing on 8GB cards due to missing Float8_e4m3fn support. The script requires you to pass your GPU’s GFX target code so it downloads the correct kernel binaries from AMD’s experimental multi-arch builds. In testing, an RX 5500 XT dropped its raw SDXL iteration time to 1.6 seconds per step and halved VRAM usage, moving from a nearly useless state to practical image generation.

"This repository provides an automated installation workflow to run ComfyUI on Windows using AMD's native ROCm/HIP backend, completely bypassing the DirectML translation layer." — Source: GitHub

Project Links