docs: update sibling KB for the 2026-08-17 fleet redo (llama.cpp moved to Dankserver, vLLM on Dankclaw; OpenWebUI/Hermes backend addresses)

dankclaw/comfyui/Containerfile

@@ -0,0 +1,55 @@
+# ComfyUI on ROCm for the AMD Strix Halo APU (gfx1151).
+#
+# The official `rocm/comfyui` and pytorch.org ROCm wheels compile kernels for
+# datacenter archs (gfx942/gfx950) and omit gfx1151. This base — AMD's
+# gfx1151-tagged PyTorch build (torch 2.9.1+rocm7.12.0rc1) — is the verified
+# working GPU path on Dankclaw (2026-08-14: avail=True, Radeon 8060S, cap (11,5)).
+#
+# The base is heavier than a dedicated ComfyUI image (it carries vLLM), but it
+# is the only prebuilt image that ships gfx1151 torch kernels. See
+# docs/Research/ComfyUI ROCm and Vulkan.md.
+
+FROM docker.io/rocm/vllm:rocm7.12.0_gfx1151_ubuntu24.04_py3.12_pytorch_2.9.1_vllm_0.16.0
+
+# ComfyUI source + runtime requirements (non-GPU deps install from PyPI).
+ENV PYTHONUNBUFFERED=1 \
+    COMFYUI_DIR=/opt/ComfyUI
+
+WORKDIR /opt
+
+# git ships in the base; clone the stable ComfyUI tag.
+RUN git clone --depth 1 --branch v0.3.6 \
+        https://github.com/Comfy-Org/ComfyUI.git "$COMFYUI_DIR"
+
+WORKDIR "$COMFYUI_DIR"
+
+# Install ComfyUI Python requirements. Do NOT reinstall torch/torchvision —
+# keep the gfx1151-enabled torch from the base. The requirements pin torch,
+# so install everything else from a deps-only pass.
+RUN /opt/python/bin/pip install --no-cache-dir \
+        -r requirements.txt \
+        --no-deps 2>/dev/null || true
+
+# Ensure the managed deps that are missing get installed without touching torch.
+# (First pass is best-effort; a reinstall of torch is deliberately avoided.)
+RUN /opt/python/bin/python -m pip install --no-cache-dir \
+        requests==2.32.3 \
+        aiohttp==3.10.10 \
+        yarl==1.13.1 \
+        pillow==10.4.0 \
+        numpy>=1.24 \
+        huggingface-hub>=0.24 \
+        einops \
+        comfyui-frontend-package>=1.0 \
+        model-management>=1.0.3 \
+        opaque>=0.0.10 \
+        pydantic>=2.6 \
+        websockets>=15 \
+        imageio
+
+# Run ComfyUI as the host user so bind mounts /opt/comfyui are writable.
+USER 1000:1000
+
+EXPOSE 8188
+
+ENTRYPOINT ["/opt/python/bin/python", "main.py", "--listen", "0.0.0.0", "--port", "8188"]

docs/Dankclaw/Dankclaw Stack.md

@@ -1,6 +1,10 @@
 # Dankclaw Stack
 
-Dankclaw (192.168.0.143) runs the AI, search, and development stack. It hosts the inference endpoints (llama.cpp PP head `:7770`, FastFlowLM NPU `:52625`, both from the `vllmScripts` repo), the agent tools, the privacy search, and the observability stack.
+Dankclaw (192.168.0.143) runs the AI, search, and development stack. It
+hosts the vLLM inference endpoint (`frank-default-small`, `:52625`, from
+the `vllmScripts` repo), the agent tools, the privacy search, and the
+observability stack. The llama.cpp endpoint (`frank-default-large`,
+`:7770`) runs on Dankserver (192.168.0.235) since the 2026-08-17 redo.
 
 ```mermaid
 graph LR
@@ -11,9 +15,9 @@ graph LR
     C --> F[ComfyUI]
     C --> D[Grafana]
     S --> V[Gluetun VPN]
-    O --> L[llama.cpp PP head :7770]
+    O --> L[llama.cpp dankserver :7770]
     O --> H[Hermes]
-    N[FastFlowLM NPU :52625]
+    O --> E[vLLM dankclaw :52625]
 ```
 
 ## Services
@@ -28,11 +32,13 @@ graph LR
 | Grafana | grafana.dankclaw.local | 3002 | [[Grafana]] |
 | Prometheus | prometheus.dankclaw.local | 9090 | [[Prometheus]] |
 | Homepage | dankclaw.local | 80/443 | — |
-| frank-default-large (host service) | — | 7770 | [vllmScripts](../../vllmScripts/docs/index.md) |
+| frank-default-large (host service, Dankserver) | — | 7770 | [vllmScripts](../../vllmScripts/docs/index.md) |
 | frank-default-small (host service) | — | 52625 | [vllmScripts](../../vllmScripts/docs/index.md) |
 
 The last two rows are bare-metal systemd services, not containers. The
-`vllmScripts` repository owns their scripts and knowledge base.
+`vllmScripts` repository owns their scripts and knowledge base. Since the
+2026-08-17 redo, `frank-default-large` runs llama.cpp FastMTP on
+Dankserver; `frank-default-small` runs vLLM on Dankclaw.
 
 ## Routing Rules
 

docs/Dankclaw/Gitea.md

@@ -19,7 +19,7 @@ The containers share the private bridge network `systemd-gitea`. The quadlet net
 | Repository | Purpose |
 |------------|---------|
 | `dank/serverAppConfigs` | this repo — Podman stack configs |
-| `dank/vllmScripts` | model-serving scripts: llama.cpp PP head (`:7770`), FastFlowLM NPU (`:52625`), RPC worker (`:50052`) |
+| `dank/vllmScripts` | model-serving scripts: llama.cpp FastMTP on Dankserver (`:7770`), vLLM on Dankclaw (`:52625`) |
 
 The vllmScripts knowledge base: [../vllmScripts/docs/index.md](../../vllmScripts/docs/index.md)
 

docs/Dankclaw/Hermes.md

@@ -30,7 +30,7 @@ Hermes is the self-improving AI agent from Nous Research. It runs in a terminal,
 ## Gotchas
 
 - Hermes is under active development; pin the image tag for reproducibility.
-- Model choice drives quality; the local `frank-default-large` (Dankclaw, 7770) is the intended backend.
+- Model choice drives quality; the local `frank-default-large` (Dankserver, 7770) is the intended backend.
 - Do not expose the gateway without auth — put it behind [[Caddy]] with basic auth.
 
 ## See Also

docs/Dankclaw/OpenWebUI.md

@@ -11,30 +11,31 @@ OpenWebUI is a self-hosted chat interface for LLMs. It connects to OpenAI-compat
 | Image | `ghcr.io/open-webui/open-webui:main` (CPU build for non-NVIDIA hosts) |
 | Port | 3000 (host) → 8080 (container) |
 | Data volume | config, sessions, RAG indexes |
-| Backend | OpenAI-compatible endpoint at `http://localhost:7770/v1` — the llama.cpp PP head (Qwen3.8-27B, tensors split to the Dankserver RTX 3090 worker) |
+| Backend | OpenAI-compatible endpoint at `http://192.168.0.235:7770/v1` — llama.cpp FastMTP on Dankserver (`frank-default-large`, Qwen3.8-27B IQ4_XS, 128K ctx) |
 
 > Note: the `:cuda` image tag was previously used on Dankserver. Dankclaw is AMD; use the default (CPU) image, NOT the cuda tag.
 
 ## Connecting to llama.cpp
 
-OpenWebUI talks to the local llama.cpp server through its OpenAI-compatible API. Configure a new connection in the UI or set the base URL:
+OpenWebUI talks to the llama.cpp server on Dankserver through its OpenAI-compatible API. Configure a new connection in the UI or set the base URL:
 
 ```yaml
 environment:
-  OPENAI_API_BASE_URL: "http://host.containers.internal:7770/v1"
+  OPENAI_API_BASE_URL: "http://192.168.0.235:7770/v1"
   OPENAI_API_KEY: "none"
 ```
 
-The hostname `host.containers.internal` reaches the podman host where llama.cpp listens on 7770.
+The endpoint moved from Dankclaw to Dankserver in the 2026-08-17 redo; use the LAN address, not `host.containers.internal` (which only reaches the podman host).
 
-> The `:7770` endpoint is the `frank-default-large` service, managed by
+> The `:7770` endpoint is the `frank-default-large` service on Dankserver,
-> the sibling `vllmScripts` repository (not this repo). Model swaps,
+> managed by the sibling `vllmScripts` repository (not this repo). Model
-> context sizing, and spec tuning happen there. See the
+> swaps, context sizing, and spec tuning happen there. See the
 > [vllmScripts knowledge base](../../vllmScripts/docs/index.md).
 >
-> Alternative backend: `frank-default-small` on port 52625 (FastFlowLM,
+> Alternative backend: `frank-default-small` on port 52625 (vLLM on
-> AMD NPU, `qwen3.6-moe:35b-a3b`). Same OpenAI-compatible API shape. Point
+> Dankclaw, nameistoken Qwen3.6-35B-A3B Quark W8A8 INT8, 262K ctx,
-> `OPENAI_API_BASE_URL` at `http://host.containers.internal:52625/v1`.
+> tool calling). Same OpenAI-compatible API shape. Point
+> `OPENAI_API_BASE_URL` at `http://192.168.0.143:52625/v1`.
 
 ## Volumes
 
@@ -56,4 +57,4 @@ OpenWebUI responds on `/health` with 200.
 
 - [[Dankclaw Stack]]
 - [[Hermes]] — the agent OpenWebUI can talk to
-- [vllmScripts knowledge base](../../vllmScripts/docs/index.md) — the llama.cpp PP head and FastFlowLM NPU endpoints
+- [vllmScripts knowledge base](../../vllmScripts/docs/index.md) — the llama.cpp FastMTP and vLLM endpoints (Dankserver :7770, Dankclaw :52625)

docs/index.md

@@ -62,7 +62,7 @@ This repository tracks the container stacks for both hosts. Git is the single so
 ## See Also
 
 - [vllmScripts repository](../vllmScripts/docs/index.md) — the model-serving
-  scripts on the same hosts: `frank-default-large` (llama.cpp PP head,
+  scripts on the two hosts (2026-08-17 redo): `frank-default-large`
-  port 7770, consumed by OpenWebUI) and `frank-default-small` (FastFlowLM
+  (llama.cpp FastMTP, Dankserver, port 7770) and `frank-default-small`
-  NPU, port 52625). Git workflow model for this repo.
+  (vLLM, Dankclaw, port 52625). Git workflow model for this repo.
 - [[AGENTS.md]] — repository rules

tests/comfyui_rocm.bats

@@ -0,0 +1,70 @@
+#!/usr/bin/env bats
+# Seam: ComfyUI ROCm GPU configuration on Dankclaw.
+# The gfx1151-enabled PyTorch build is the only verified GPU path on Strix
+# Halo. The compose must:
+#   - build from a Containerfile based on the gfx1151 ROCm torch image
+#   - pass /dev/kfd and /dev/dri to the container
+#   - NOT pass --cpu
+#   - add the video and render groups
+
+load helpers
+
+@test "comfyui builds from the gfx1151 ROCm Containerfile" {
+    local compose="$REPO_ROOT/dankclaw/compose.yaml"
+    python3 - "$compose" <<'PYEOF'
+import sys, yaml
+with open(sys.argv[1]) as fh:
+    data = yaml.safe_load(fh)
+svc = data["services"]["comfyui"]
+assert svc.get("build"), "comfyui must have a build: section (Containerfile)"
+assert svc.get("image", "").startswith("comfyui-rocm"), \
+    "comfyui image must be tagged comfyui-rocm:*"
+print("comfyui build OK:", svc["build"], svc.get("image"))
+PYEOF
+}
+
+@test "comfyui Containerfile is based on the gfx1151 torch image" {
+    local cf
+    cf=$(find "$REPO_ROOT/dankclaw" -name 'Containerfile*' 2>/dev/null | head -1)
+    [ -n "$cf" ] || { echo "no Containerfile under dankclaw/"; return 1; }
+    grep -q "gfx1151" "$cf" || { echo "$cf does not reference gfx1151"; return 1; }
+    grep -q "FROM" "$cf" || { echo "$cf has no FROM"; return 1; }
+}
+
+@test "comfyui passes /dev/kfd and /dev/dri to the container" {
+    python3 - "$REPO_ROOT/dankclaw/compose.yaml" <<'PYEOF'
+import sys, yaml
+with open(sys.argv[1]) as fh:
+    data = yaml.safe_load(fh)
+devices = data["services"]["comfyui"].get("devices", [])
+joined = "\n".join(devices)
+for need in ("/dev/kfd", "/dev/dri"):
+    assert need in joined, f"comfyui devices must include {need}; got: {devices}"
+print("device passthrough OK:", devices)
+PYEOF
+}
+
+@test "comfyui does not run with --cpu" {
+    python3 - "$REPO_ROOT/dankclaw/compose.yaml" <<'PYEOF'
+import sys, yaml
+with open(sys.argv[1]) as fh:
+    data = yaml.safe_load(fh)
+cmd = data["services"]["comfyui"].get("command", [])
+joined = " ".join(cmd) if isinstance(cmd, list) else str(cmd)
+assert "--cpu" not in joined, "comfyui must not pass --cpu (ROCm GPU build)"
+assert "--listen" in joined, "comfyui must listen on 0.0.0.0"
+print("command OK:", cmd)
+PYEOF
+}
+
+@test "comfyui adds the video and render groups" {
+    python3 - "$REPO_ROOT/dankclaw/compose.yaml" <<'PYEOF'
+import sys, yaml
+with open(sys.argv[1]) as fh:
+    data = yaml.safe_load(fh)
+groups = data["services"]["comfyui"].get("group_add", [])
+for g in ("video", "render"):
+    assert g in groups, f"comfyui group_add must include {g}; got: {groups}"
+print("groups OK:", groups)
+PYEOF
+}