ComfyUI: GPU (gfx1151 ROCm) build via container

Replace the CPU stock image with a container built from
Containerfile.comfyui on AMD's gfx1151-tagged vLLM base — the only torch
build verified to report torch.cuda.is_available()=True on the Strix Halo
APU (Radeon 8060S, capability (11,5)).

- Containerfile bakes the ComfyUI code in and installs its deps, but skips
  the torch family (the base image already ships the gfx1151 ROCm
  torch/torchvision/torchaudio; an unpinned 'torch' in requirements.txt would
  clobber the gfx1151 kernels).
- compose comfyui service: build + image comfyui-rocm:gfx1151, passes
  /dev/kfd + /dev/dri, adds video+render groups, drops --cpu. Models/data
  persist under /opt/comfyui.
- Un-skip the four compose-wiring acceptance tests (the gate for this work).

dankclaw/Containerfile.comfyui

@@ -0,0 +1,24 @@
+# ComfyUI on the gfx1151 (AMD Strix Halo / Radeon 8060S) ROCm torch build.
+#
+# Base image: AMD's gfx1151-tagged vLLM image. It is the only verified build
+# where torch.cuda.is_available() is True on this APU (device capability
+# (11,5) = gfx1151). It already ships the matching ROCm torch family
+# (torch 2.9.1+rocm, torchvision 0.24.0+rocm, torchaudio 2.9.0+rocm), so we
+# must NOT let pip reinstall them: ComfyUI's requirements.txt lists an unpinned
+# `torch`, which would otherwise clobber the gfx1151 kernels with a default build.
+FROM docker.io/rocm/vllm:rocm7.12.0_gfx1151_ubuntu24.04_py3.12_pytorch_2.9.1_vllm_0.16.0
+
+# Bake the ComfyUI code into the image (rebuild the image to update the code).
+RUN git clone --depth 1 https://github.com/comfyanonymous/ComfyUI /opt/ComfyUI
+WORKDIR /opt/ComfyUI
+
+# Pre-create the user-data dirs. The host mounts its collections over them
+# (see the comfyui service in compose.yaml).
+RUN mkdir -p /opt/ComfyUI/models /opt/ComfyUI/input /opt/ComfyUI/output /opt/ComfyUI/custom_nodes
+
+# Install ComfyUI's dependencies, but skip the torch family (already present
+# above as the gfx1151 ROCm builds). torchsde and the rest are kept.
+RUN grep -viE '^(torch|torchvision|torchaudio)([^a-z]|$)' requirements.txt > /tmp/req-notorch.txt \
+    && pip install --no-cache-dir -r /tmp/req-notorch.txt
+
+EXPOSE 8188

dankclaw/compose.yaml

@@ -11,21 +11,37 @@ services:
   # Deploy with: scripts/deploy-searxng-quadlet.sh dankclaw
   # See ../docs/Dankclaw/SearXNG.md, Gluetun.md, Valkey.md.
 
-  # --- AI image generation (CPU backend; ROCm does not support gfx1151) ----
+  # --- AI image generation (ROCm GPU backend, gfx1151) ---------------------
+  # Built from Containerfile.comfyui on AMD's gfx1151 torch base. The GPU is
+  # passed through (/dev/kfd, /dev/dri) with the video+render groups; no --cpu.
+  # Models + data persist under /opt/comfyui. See ../docs/Dankclaw/ComfyUI.md.
   comfyui:
-    image: ghcr.io/comfyanonymous/comfyui:latest
+    build:
+      context: .
+      dockerfile: Containerfile.comfyui
+    image: comfyui-rocm:gfx1151
     container_name: comfyui
     restart: unless-stopped
-    command: ["python", "main.py", "--cpu", "--listen", "0.0.0.0", "--port", "8188"]
+    command: ["python", "main.py", "--listen", "0.0.0.0", "--port", "8188"]
     ports:
       - "8188:8188"
+    devices:
+      - /dev/kfd:/dev/kfd
+      - /dev/dri:/dev/dri
+    group_add:
+      - video
+      - render
     volumes:
-      - /opt/comfyui:/workspace
+      - /opt/comfyui/models:/opt/ComfyUI/models
+      - /opt/comfyui/input:/opt/ComfyUI/input
+      - /opt/comfyui/output:/opt/ComfyUI/output
+      - /opt/comfyui/custom_nodes:/opt/ComfyUI/custom_nodes
     healthcheck:
       test: ["CMD", "python3", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:8188/')"]
       interval: 30s
       timeout: 10s
-      retries: 3
+      retries: 5
+      start_period: 180s
 
   # --- Chat UI + agent -----------------------------------------------------
   openwebui:

tests/comfyui_rocm.bats

@@ -7,16 +7,15 @@
 #   - NOT pass --cpu
 #   - add the video and render groups
 #
-# STATUS (2026-08-18): the GPU migration is NOT yet applied. The compose still
-# runs ComfyUI with --cpu on the stock image (the gfx1151 torch image is not
-# built yet). The four compose-wiring tests below are skipped until the ROCm
-# build lands. The Containerfile test stays active: the file exists and is ready.
-# See docs/Research/ComfyUI ROCm and Vulkan.md. Re-enable the skips then.
+# STATUS (2026-08-19): the GPU migration is applied. The compose builds
+# comfyui-rocm:gfx1151 from Containerfile.comfyui (AMD's gfx1151 torch base),
+# passes /dev/kfd + /dev/dri, adds the video+render groups, and drops --cpu.
+# The four compose-wiring tests below are active (un-skipped). See
+# docs/Research/ComfyUI ROCm and Vulkan.md.
 
 load helpers
 
 @test "comfyui builds from the gfx1151 ROCm Containerfile" {
-    skip "GPU ComfyUI migration pending (CPU path in use)"
     local compose="$REPO_ROOT/dankclaw/compose.yaml"
     python3 - "$compose" <<'PYEOF'
 import sys, yaml
@@ -39,7 +38,6 @@ PYEOF
 }
 
 @test "comfyui passes /dev/kfd and /dev/dri to the container" {
-    skip "GPU ComfyUI migration pending (CPU path in use)"
     python3 - "$REPO_ROOT/dankclaw/compose.yaml" <<'PYEOF'
 import sys, yaml
 with open(sys.argv[1]) as fh:
@@ -53,7 +51,6 @@ PYEOF
 }
 
 @test "comfyui does not run with --cpu" {
-    skip "GPU ComfyUI migration pending (CPU path in use)"
     python3 - "$REPO_ROOT/dankclaw/compose.yaml" <<'PYEOF'
 import sys, yaml
 with open(sys.argv[1]) as fh:
@@ -67,7 +64,6 @@ PYEOF
 }
 
 @test "comfyui adds the video and render groups" {
-    skip "GPU ComfyUI migration pending (CPU path in use)"
     python3 - "$REPO_ROOT/dankclaw/compose.yaml" <<'PYEOF'
 import sys, yaml
 with open(sys.argv[1]) as fh: