- Home
- Skills
- Workflow Automation
- ComfyUI on AMD RDNA3 (ROCm, Windows)
More screenshots
Works with the AI Agents you already use
ComfyUI on AMD RDNA3 (ROCm, Windows)
Make ComfyUI fast and correct on RX 7000-series AMD GPUs on Windows: a tested SageAttention patch (~1.65x), the Triton header fix, and the silent failures that waste days.
$9
ComfyUI on AMD RDNA3 (ROCm, Windows)
Example session with this skill installed
I patched SageAttention's BLOCK_M/BLOCK_N to 64/32 on my 7900 XTX to fix the shared-memory error. It runs, but every Wan frame is noise.
- Read your context and instructions
- Compiled the comfyui on amd
The quantizer block size is coupled to the kernel's. Reinstall sageattention 1.0.6, run patch_sageattention_rdna3.py (patches all 9 files, including BLKQ/BLKK 64/32), clear the Triton cache, then compare a fixed-seed render with check_render.py --baseline.
Connects securely to your tools. The creator never sees your data.
About this skill
RDNA3 + ROCm on Windows works well, but its worst failures are silent: wrong output, no error, no NaN. The core rule: never trust a speed-up or a "fix" until you have rendered at a fixed seed and LOOKED at the result.
Measured on: RX 7900 XTX 24 GB (gfx1100), torch 2.9.1 + ROCm 7.2.1, ComfyUI portable, embedded Python 3.12. Other RX 7000 cards (gfx1101/1102) share the architecture and the 64 KB LDS limit, so the patch should apply, but they are untested. Speeds will differ. This skill assumes ROCm torch is already installed and ComfyUI already runs.
Run everything in cmd.exe from the ComfyUI portable root (the folder that holds
python_embeded\). <SKILL_DIR> means this skill's folder. Always use the embedded Python, never
a system Python.
| Want | Do | Measured |
|---|---|---|
| Faster attention | SageAttention 1.0.6 + the RDNA3 patch, launch with --use-sage-attention | 72 → 38.7 s/it (Wan 14B), ~1.65x |
| Triton | triton-windows + the embedded Python's C headers | — |
| Less VRAM for diffusion models | GGUF (Q6_K / Q8_0). Never fp8 diffusion checkpoints. | — |
| Faster video decode | Purge models right before VAEDecode (section 5) | 244 → 12.6 s |
| Tips that DON'T work here | fp8, torch.compile, --fast fp16_accumulation, --highvram | reference/dead-ends.md |
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Workflow Automation.
Measured on an RX 7900 XTX (gfx1100) only; other RX 7000 cards are untested. The patch script supports SageAttention 1.0.6 only. Does not install ROCm torch or ComfyUI itself. Not for NVIDIA or Linux.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 4 days ago
- Accounts
Nuget
- Software
- AMD RX 7000-series GPUComfyUI with ROCm torch
Python (ComfyUI embedded)
Permissions
Allowed Hosts
File Scopes
The patch script edits installed SageAttention files and keeps .orig backups. The header script only adds missing files and never overwrites.