build: the agent installs one CUDA-13 stack instead of two, the web image drops ML packages it never imported, and Redis moves to 8 (1451, 1452)

Agent:
- The image ran PyPI's CUDA-13 torch 2.14 and onnxruntime-gpu 1.30 on a
  CUDA 12.9 cudnn-runtime base. requirements.txt had silently replaced the
  Dockerfile's torch 2.6+cu124, because ultralytics pulls torchvision, which
  pulls its own torch. That left ~3 GB of base libraries and a ~3 GB torch
  nothing loaded: 10 GB compressed.
- Now: an nvidia/cuda 13.0.3 `base` image, with torch and torchvision
  installed together from cu130. CUDA and cuDNN come from the nvidia-* pip
  packages; onnxruntime-gpu declares its [cuda,cudnn] extras.
- fc_agent/accel.py preloads those libraries for onnxruntime. It then logs,
  and reports in /status, whether torch and the ONNX CUDA provider actually
  got the GPU, since both fall back to the CPU silently.

Web image:
- Drop opencv-python-headless and onnxruntime, plus the opencv-only apt libs.
  Both have been listed since the scaffold and nothing in backend/ imports
  them.
- torch/torchvision move to 2.14/0.29, and the unexplained caps are lifted
  (rule 154).

Redis: 8-alpine in both compose files and both CI service containers. That
gives an AGPLv3 licence option, where 7.4 was RSAL/SSPL only. The client
moves to >=8.1.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LVjrnpQjRgHdvq95rASoiR

agent/Dockerfile

@@ -1,10 +1,21 @@
 # FabledCurator GPU agent — runs on the desktop with the GPU.
-# CUDA 12.9 + cuDNN 9 runtime so onnxruntime-gpu can use the card (it needs
-# cuDNN 9 — the plain -runtime image lacks it: "libcudnn.so.9: cannot open
-# shared object file"); ffmpeg for video frames. Ubuntu 24.04 → Python 3.12.
-# Stays on the CUDA-12 / cuDNN-9 line the default onnxruntime-gpu + torch are
-# built against (CUDA 13 has only nascent ONNX Runtime support).
-FROM nvidia/cuda:12.9.2-cudnn-runtime-ubuntu24.04
+#
+# The `base` flavour, not `cudnn-runtime`: CUDA and cuDNN arrive as the
+# `nvidia-*` pip packages torch and onnxruntime-gpu depend on, so the base only
+# has to hand the container the driver (it sets NVIDIA_VISIBLE_DEVICES /
+# NVIDIA_DRIVER_CAPABILITIES for the Container Toolkit). Until #1451 this was
+# `12.9.2-cudnn-runtime` under a `torch==2.6.0+cu124` — and requirements.txt then
+# REPLACED that torch with PyPI's CUDA-13 build (ultralytics pulls torchvision,
+# which pulls its matching torch), beside a CUDA-13 onnxruntime-gpu. The image
+# ran CUDA 13 on a CUDA-12 base, carrying ~3 GB of base libraries and a ~3 GB
+# torch nothing loaded: 10 GB compressed.
+#
+# 13.0 because that is the line both wheels are built for (torch's cu130 index,
+# onnxruntime-gpu's `nvidia-cuda-runtime~=13.0`). Needs an NVIDIA driver that
+# supports CUDA 13 (580+); fc_agent/accel.py logs at startup whether torch and
+# onnxruntime actually got the GPU, since both fall back to the CPU silently.
+# ffmpeg for video frames. Ubuntu 24.04 → Python 3.12.
+FROM nvidia/cuda:13.0.3-base-ubuntu24.04
 
 # PIP_BREAK_SYSTEM_PACKAGES: Ubuntu 24.04 marks its system Python as externally
 # managed (PEP 668), so a global `pip install` errors without this. It's a
@@ -16,10 +27,12 @@ RUN apt-get update \
     && rm -rf /var/lib/apt/lists/*
 
 WORKDIR /app
-# torch from the CUDA-12.4 wheel index; its wheels bundle their own CUDA + cuDNN
-# so they run on the 12.9 base and coexist with onnxruntime-gpu. Installed first
-# + separately so the GPU build of torch is deterministic and layer-cached.
-RUN pip3 install --no-cache-dir torch==2.6.0 --index-url https://download.pytorch.org/whl/cu124
+# torch AND torchvision from the cu130 index, together and first. Installing
+# torch alone is what let the next step swap it out: ultralytics needs
+# torchvision, PyPI's torchvision pins its own torch, and pip replaced ours to
+# match. With both present, requirements.txt finds them satisfied.
+RUN pip3 install --no-cache-dir --index-url https://download.pytorch.org/whl/cu130 \
+        torch torchvision
 COPY requirements.txt .
 RUN pip3 install --no-cache-dir -r requirements.txt
 COPY fc_agent ./fc_agent

agent/README.md

@@ -15,7 +15,8 @@ sudo pacman -S nvidia-container-toolkit
 sudo nvidia-ctk runtime configure --runtime=docker
 sudo systemctl restart docker
 # verify:
-docker run --rm --gpus all nvidia/cuda:12.4.1-base-ubuntu22.04 nvidia-smi
+docker run --rm --gpus all nvidia/cuda:13.0.3-base-ubuntu24.04 nvidia-smi
+# the header's CUDA version must be 13.0 or later (driver 580+)
 ```
 
 ## 1. Get a token

agent/fc_agent/accel.py

@@ -0,0 +1,86 @@
+"""Which accelerator each runtime actually got — reported once, at startup.
+
+The agent has two GPU runtimes and both fall back to the CPU without raising:
+torch when the driver is too old for its CUDA build, and onnxruntime (the imgutils
+detector + CCIP models) when its CUDA provider cannot load its libraries. A
+fallback shows up only as slower work, and nothing reported it. On 2026-09-24 the
+image turned out to be running a CUDA-13 torch and onnxruntime on a CUDA-12 base
+(#1451), and whether the ONNX half was on the GPU could not be answered from
+anything the agent had ever logged.
+
+Also the fix for the likeliest way the ONNX half misses: onnxruntime-gpu's CUDA
+provider finds libcudart/cuBLAS/cuDNN only on the loader path, and in this image
+they live in the `nvidia-*` pip packages torch installs. `preload_dlls()` (ORT
+1.21+) loads them from there, so the provider resolves them by soname.
+
+Stdlib-only at import, so the unit suite can import it — torch and onnxruntime
+are imported inside the functions.
+"""
+
+from __future__ import annotations
+
+import ctypes
+import importlib
+import logging
+from pathlib import Path
+
+log = logging.getLogger("fc_agent.accel")
+
+# Filled by report(); /status carries it so the page can show it too.
+LAST: dict = {}
+
+
+def torch_status(imp=importlib.import_module) -> dict:
+    try:
+        torch = imp("torch")
+    except Exception as e:
+        return {"device": "unavailable", "error": str(e)}
+    out = {"version": torch.__version__, "cuda_build": torch.version.cuda}
+    if torch.cuda.is_available():
+        out["device"] = "cuda"
+        out["gpu"] = torch.cuda.get_device_name(0)
+    else:
+        out["device"] = "cpu"
+    return out
+
+
+def onnx_status(imp=importlib.import_module, load=ctypes.CDLL) -> dict:
+    try:
+        ort = imp("onnxruntime")
+    except Exception as e:
+        return {"device": "unavailable", "error": str(e)}
+    out = {"version": ort.__version__, "providers": list(ort.get_available_providers())}
+    if "CUDAExecutionProvider" not in out["providers"]:
+        out["device"] = "cpu"
+        return out
+    preload = getattr(ort, "preload_dlls", None)
+    if preload is not None:
+        try:
+            preload()
+        except Exception as e:
+            out["preload_error"] = str(e)
+    # "Available" only means the build HAS the provider. Loading its library is
+    # what resolves libcudart/cuBLAS/cuDNN — the step that fails when they are
+    # missing, and the one a session would otherwise fail silently on.
+    capi = Path(ort.__file__).parent / "capi"
+    try:
+        load(str(capi / "libonnxruntime_providers_shared.so"), mode=ctypes.RTLD_GLOBAL)
+        load(str(capi / "libonnxruntime_providers_cuda.so"))
+    except OSError as e:
+        out["device"] = "cpu"
+        out["error"] = str(e)
+    else:
+        out["device"] = "cuda"
+    return out
+
+
+def report() -> dict:
+    """Check both runtimes, log the result, and keep it for /status."""
+    LAST.clear()
+    LAST.update(torch=torch_status(), onnx=onnx_status())
+    for name, s in LAST.items():
+        if s.get("device") == "cuda":
+            log.info("accel: %s on GPU (%s)", name, s)
+        else:
+            log.warning("accel: %s is NOT on the GPU — work runs on the CPU (%s)", name, s)
+    return dict(LAST)

agent/fc_agent/app.py

@@ -11,7 +11,7 @@ import logging
 from fastapi import FastAPI, Request
 from fastapi.responses import HTMLResponse, JSONResponse
 
-from . import logbuf
+from . import accel, logbuf
 from .build_info import FC_CHANNEL, FC_REVISION, FC_VERSION, build_id, display_version
 from .config import Config
 from .gpu import read_gpu
@@ -47,6 +47,9 @@ async def _no_store(request, call_next):
 
 @app.on_event("startup")
 def _maybe_autostart() -> None:
+    # Before the worker: the report also preloads the CUDA libraries the ONNX
+    # models need, and it says in the log which runtimes landed on the GPU.
+    accel.report()
     # With AUTO_START set, a container restart (host reboot, or `restart:
     # unless-stopped` after a crash) resumes the worker on its own — the slots
     # then ride out a still-down curator via lease backoff. Lets the agent
@@ -137,6 +140,7 @@ def status():
     s["version"] = FC_VERSION or None
     s["channel"] = FC_CHANNEL or None
     s["revision"] = FC_REVISION or None
+    s["accel"] = accel.LAST or None
     return JSONResponse(s)
 
 

agent/requirements.txt

@@ -1,10 +1,12 @@
 # CCIP + figure detection (ONNX models, auto-downloaded from HuggingFace).
 dghs-imgutils>=0.4
 # GPU inference for the ONNX models. Swap to onnxruntime (CPU) for a slow
-# server-side fallback run.
-onnxruntime-gpu
-# The crop EMBEDDER (concept bag). torch is installed separately in the
-# Dockerfile from the CUDA-12.4 wheel index so the GPU build is deterministic;
+# server-side fallback run. The extras declare the CUDA/cuDNN pip packages its
+# CUDA provider loads (fc_agent/accel.py preloads them) rather than relying on
+# torch happening to install the same ones.
+onnxruntime-gpu[cuda,cudnn]
+# The crop EMBEDDER (concept bag). torch + torchvision are installed separately
+# in the Dockerfile from the cu130 wheel index, so pip never swaps them out;
 # transformers loads whatever SigLIP-family model the server announces.
 transformers>=4.45
 # Crop PROPOSERS — small YOLO detectors (booru_yolo anatomy, COCO person, comic