-
1 · Cited, not measured on the local desktop
Ubuntu 24.04+ first — see the GPU before you clone
Bare-metal Ubuntu 24.04+ is the AMDai measured ROCm path. Image-gen follows that OS even though we have no first-party ComfyUI card yet. AMD’s page assumes ROCm-enabled PyTorch on a supported system. Confirm rocminfo prints an HSA GPU agent, and that /dev/kfd and /dev/dri exist and your user is in the right groups. Windows is a cited second path. WSL2 ROCm is a 10.1 technical preview at /setup/wsl — not a measured image-gen path. Fedora and Arch have their own cited first-run pages; we do not invent a ComfyUI tok/s on those.
-
2 · Cited, not measured on the local desktop
Python venv + ROCm PyTorch — cite the wheel, do not invent a pin
ComfyUI runs in its own Python environment. Official docs recommend Python 3.13, with 3.12 as a fallback if a custom-node dep fails. Create a venv: python3 -m venv .venv && source .venv/bin/activate. ComfyUI’s README Linux AMD command is pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm7.2. AMD’s Install PyTorch for ROCm page and PyTorch’s Get started locally matrix (Linux, Python, pip, ROCm) are the official AMD/PyTorch pickers. House /rocm/versions is the 10.1 line for this site’s llama.cpp path — we do not invent a 10.1 ComfyUI wheel. After a ROCm version move, reinstall the matching wheel — /rocm/upgrade. Instinct-oriented Docker tags are not R9700 pins. If PyTorch reports no GPU, stop.
-
3 · Cited, not measured on the local desktop
Clone ComfyUI and install requirements
AMD’s page: git clone https://github.com/comfyanonymous/ComfyUI.git, activate the ROCm PyTorch venv, then pip install -r ComfyUI/requirements.txt. ComfyUI’s README says the same after you are inside the clone: pip install -r requirements.txt. Do not treat community Docker Hub images as official — ComfyUI documents that it has no official Docker image. Desktop AppImage / .deb exists on docs.comfy.org; AMDai’s first walk is the git clone AMD publishes so the venv stays the ROCm wheel you just installed.
-
4 · Cited, not measured on the local desktop
First-run UI at :8188
Start with python ComfyUI/main.py. Official prompt: To see the GUI go to: http://127.0.0.1:8188. Open that URL on the same box. ComfyUI is a node graph on HIP/ROCm PyTorch. It does not speak llama-server or Strata /v1 — do not paste http://127.0.0.1:8188 into a chat harness. llama.cpp stays first for LLMs. Image-in text-out is /vision, not this page.
-
5 · Cited, not measured on the local desktop
Pick a starter workflow — SD 1.5 first
ComfyUI’s first-generation guide: the default text-to-image graph usually loads on launch. The documented starter checkpoint is v1-5-pruned-emaonly-fp16.safetensors. AMD’s page also names Template → Model Filter → SD3.5 → SD3.5 Simple, and Chroma → Chroma1 Radiance, then download missing models. Start smaller (SD 1.5) on 8 / 12 / 16 GB class cards. Size the family on /image-gen. Do not invent seconds-per-image. Queue with Run or Ctrl+Enter as their docs describe.
-
6 · Cited, not measured on the local desktop
Put checkpoints on fast NVMe — models/checkpoints
Official first-generation page: models live under ComfyUI/models/ with subfolders checkpoints, embeddings, vae, lora, and extra_model_paths.yaml for more roots. README: put SD checkpoints in models/checkpoints and VAE in models/vae. Git-clone path is ComfyUI/models/checkpoints. Keep files on fast NVMe (house ~/models or the clone’s models/), not a spinning disk and not a slow bulk mount such as /mnt/bulk. Staging is not RAM offload. Refresh with R after you drop a file. Not measured on the local desktop.
-
7 · Cited, not measured on the local desktop
VRAM fit is an estimate — reuse the /image-gen classes
Publisher cards do not publish a VRAM-gigabyte figure. /image-gen already maps 32 / 24 / 20 / 16 / 12 / 8 GB discrete classes plus Strix Halo shared memory. 32 GB R9700 has the most discrete headroom on that list. Start SD 1.5 on 8 / 12 / 16 GB. Flux-class 12B BF16 is tighter than SDXL. Seconds-per-image stay not yet measured by AMDai. Do not copy a tok/s from the 35B llama.cpp table.
-
8 · Cited, not measured on the local desktop
System RAM — both shopper labels, KV only on the LLM
Minimum system RAM needed is a kit estimate rounded to a real size. For a consumer box that may also run a local LLM, the /image-gen default is a 64 GB kit (8 GB OS/desktop plus 16 GB headroom plus whatever the LLM offload and PyTorch need) — labeled estimate, not a measured ComfyUI peak. RAM needed at full context applies to the llama.cpp KV window, not to a diffusion graph. ComfyUI has no llama.cpp -c. If you are image-gen only, omit the full-context line and keep the min-kit estimate. If an LLM shares the box, show both labels on the LLM. GB here is Linux GiB printed as GB.
-
9 · Cited, not measured on the local desktop
Portable and Desktop notes — do not take the Nvidia zip on Radeon
ComfyUI’s README: Windows portable Nvidia is for Nvidia or CPU-only. The AMD zip is ComfyUI_windows_portable_amd.7z. Desktop exists for Windows, macOS, and Linux (AppImage / .deb) and tracks stable; portable or git gets latest commits. AMDai’s Linux first try stays the AMD-documented git clone so the venv is the ROCm wheel. Community DirectML forks and unofficial wrappers stay unofficial. Cited, not measured on the local desktop.
-
10 · Cited, not measured on the local desktop
Two cards — no invented ComfyUI split
The AMD ComfyUI page and ComfyUI README we cite do not walk a two-GPU ComfyUI --tensor-split. We do not invent one. On a second card, read /multi-gpu for the llama.cpp job. A local LLM and ComfyUI can share a box; they do not share a /v1. On one card, VRAM is one pool — do not stack two full loads and guess. We have not measured image-gen-plus-LLM on the local desktop.
-
11 · Cited, not measured on the local desktop
When the GPU never lights up
ComfyUI is HIP/ROCm PyTorch. If the process never sees /dev/kfd, or PyTorch reports no GPU, you are on CPU. Check groups, device nodes, and the wheel’s gfx list. This is not a llama.cpp Vulkan miss. OOM: lower resolution or batch; SDXL and Flux cards document CPU offload. Wrong gfx / hipErrorNoBinaryForGpu: /rocm-problems/hip-no-binary-for-gpu. HSA_OVERRIDE_GFX_VERSION is unofficial — ComfyUI’s README names it for some 6700 / 7600 cases; gfx1201 on ROCm 10.x should not need it. Smart App Control on Windows is AMD’s cited start-block. Start from /troubleshoot.