build: refresh dependencies and release v0.4.0

CHANGELOG.md

@@ -9,6 +9,32 @@ releases.
 
 ## [Unreleased]
 
+## [v0.4.0] - 2026-08-20 UTC
+
+### Changed
+
+- Update the PyTorch stack from PyTorch 2.8.0, torchvision 0.23.0, and
+  xformers 0.0.32.post2 on CUDA 12.9 to PyTorch 2.13.0, torchvision 0.28.0,
+  and xformers 0.0.35 on CUDA 13.0.
+- Update ONNX from 1.21.0 to 1.22.0.
+- Update the runtime CA certificate bundle to the current Ubuntu 24.04
+  security-update version.
+- Update the locked `charset-normalizer` dependency from 3.4.9 to 3.5.0.
+- Document the NVIDIA CUDA container license included in the published image.
+
+### Security
+
+- Update ONNX and PyTorch to versions containing the available fixes for the
+  repository's current Dependabot advisories.
+
+### Notes
+
+- Continue to bundle sd-scripts v0.11.1 with Python 3.10. Retain the CUDA
+  12.9.1 base image for ONNX Runtime GPU compatibility while PyTorch and
+  xformers use their CUDA 13.0 wheel builds.
+- GitHub Actions, uv, Python, CUDA, hadolint, and APM remain at their latest
+  reviewed versions that completed the seven-day cooldown.
+
 ## [v0.3.1] - 2026-07-26 UTC
 
 ### Fixed
@@ -238,7 +264,9 @@ releases.
 - Add the first GitHub Actions build workflow and disk cleanup step.
 - Add initial ignore rules for local work directories.
 
-[Unreleased]: https://github.com/aoirint/sd_scripts_docker/compare/v0.3.0...HEAD
+[Unreleased]: https://github.com/aoirint/sd_scripts_docker/compare/v0.4.0...HEAD
+[v0.4.0]: https://github.com/aoirint/sd_scripts_docker/compare/v0.3.1...v0.4.0
+[v0.3.1]: https://github.com/aoirint/sd_scripts_docker/compare/v0.3.0...v0.3.1
 [v0.3.0]: https://github.com/aoirint/sd_scripts_docker/compare/v0.2.0...v0.3.0
 [v0.2.0]: https://github.com/aoirint/sd_scripts_docker/compare/v0.1.0...v0.2.0
 [v0.1.0]: https://github.com/aoirint/sd_scripts_docker/releases/tag/v0.1.0

Dockerfile

@@ -30,7 +30,7 @@ RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
     apt-get update
 
     apt-get install -y --no-install-recommends \
-        ca-certificates=20240203 \
+        ca-certificates=20260601~24.04.1 \
         git=1:2.43.0-1ubuntu7.3
 SH
 
@@ -74,7 +74,7 @@ RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \
     apt-get update
 
     apt-get install -y --no-install-recommends \
-        ca-certificates=20240203 \
+        ca-certificates=20260601~24.04.1 \
         libgl1=1.7.0-1build1 \
         libglib2.0-0t64=2.80.0-6ubuntu3.8 \
         tk=8.6.14build1

README.md

@@ -177,3 +177,11 @@ specified by `SD_SCRIPTS_VERSION` in the Dockerfile. sd-scripts is primarily
 licensed under the Apache License 2.0, with some portions under separate license
 terms. See [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md) and the upstream
 license information for details.
+
+The image also uses the
+[NVIDIA CUDA container image](https://hub.docker.com/r/nvidia/cuda) specified by
+`CUDA_RUNTIME_IMAGE` in the Dockerfile. Its container license is included at
+`/NGC-DL-CONTAINER-LICENSE` in the published image. See
+[THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md) and the
+NVIDIA Deep Learning Container License:
+<https://developer.download.nvidia.com/licenses/NVIDIA_Deep_Learning_Container_License.pdf>.

THIRD_PARTY_NOTICES.md

@@ -2,6 +2,15 @@
 
 This project builds Docker images that include third-party software.
 
+## NVIDIA CUDA container image
+
+- Source: <https://hub.docker.com/r/nvidia/cuda>
+- Base image: see `CUDA_RUNTIME_IMAGE` in `Dockerfile`
+- Bundled location in image: `/usr/local/cuda-12.9`
+- License: NVIDIA Deep Learning Container License
+  <https://developer.download.nvidia.com/licenses/NVIDIA_Deep_Learning_Container_License.pdf>
+- License file in image: `/NGC-DL-CONTAINER-LICENSE`
+
 ## kohya-ss/sd-scripts
 
 - Source: <https://github.com/kohya-ss/sd-scripts>

VERSION

@@ -1 +1 @@
-0.3.1
+0.4.0

docs/update-sd-scripts.md

@@ -131,44 +131,52 @@ wheel depends on the CUDA target. Read the selected upstream README and choose
 the PyTorch stack that matches this image's supported GPU generation.
 
 For RTX 50 series support, use the upstream-recommended PyTorch and CUDA line
-when available. For example, PyTorch 2.8.0 with CUDA 12.9 uses:
+when available. For example, PyTorch 2.13.0 with CUDA 13.0 uses:
 
 ```toml
-"torch==2.8.0+cu129",
-"torchvision==0.23.0+cu129",
-"xformers==0.0.32.post2",
+"torch==2.13.0+cu130",
+"torchvision==0.28.0+cu130",
+"xformers==0.0.35",
 ```
 
 and:
 
 ```toml
 [[tool.uv.index]]
-name = "pytorch-cu129"
-url = "https://download.pytorch.org/whl/cu129"
+name = "pytorch-cu130"
+url = "https://download.pytorch.org/whl/cu130"
 explicit = true
 
 [tool.uv.sources]
-torch = { index = "pytorch-cu129" }
-torchvision = { index = "pytorch-cu129" }
-xformers = { index = "pytorch-cu129" }
+torch = { index = "pytorch-cu130" }
+torchvision = { index = "pytorch-cu130" }
+xformers = { index = "pytorch-cu130" }
 ```
 
 Before locking, confirm that the selected wheels exist for Python 3.10 and
 Linux x86_64:
 
 ```shell
-curl -fsSL https://download.pytorch.org/whl/cu129/torch/ | \
-  rg 'torch-2\.8\.0\+cu129-cp310-cp310-.*x86_64'
+curl -fsSL https://download.pytorch.org/whl/cu130/torch/ | \
+  rg 'torch-2\.13\.0\+cu130-cp310-cp310-.*x86_64'
 
-curl -fsSL https://download.pytorch.org/whl/cu129/torchvision/ | \
-  rg 'torchvision-0\.23\.0\+cu129-cp310-cp310-.*x86_64'
+curl -fsSL https://download.pytorch.org/whl/cu130/torchvision/ | \
+  rg 'torchvision-0\.28\.0\+cu130-cp310-cp310-.*x86_64'
 
-curl -fsSL https://download.pytorch.org/whl/cu129/xformers/ | \
-  rg 'xformers-0\.0\.32\.post2-.*x86_64'
+curl -fsSL https://download.pytorch.org/whl/cu130/xformers/ | \
+  rg 'xformers-0\.0\.35-.*x86_64'
 ```
 
-If the upstream README recommends a different CUDA line, replace `cu129`,
-package versions, and the `pytorch-cu129` index name consistently.
+If the upstream README recommends a different CUDA line, replace `cu130`,
+package versions, and the `pytorch-cu130` index name consistently.
+
+Do not assume that the CUDA runtime base image must use the same CUDA release
+as the PyTorch wheel index. PyTorch wheels bundle their CUDA user-space
+libraries, while the PyPI ONNX Runtime GPU wheel can require an older CUDA
+major version from the base image. Verify both stacks on a GPU before changing
+the runtime base. A mixed setup is acceptable when PyTorch, xformers, and ONNX
+Runtime all execute successfully and the reason is recorded in the changelog
+and pull request.
 
 ## 5. Update The Dockerfile
 
@@ -317,6 +325,24 @@ docker run --rm --gpus all --entrypoint python sd-scripts:update-test -c \
   'import torch; print(torch.cuda.is_available()); print(torch.cuda.get_device_name(0) if torch.cuda.is_available() else "no cuda")'
 ```
 
+For PyTorch-stack updates, exercise CUDA kernels instead of relying on imports
+or `torch.cuda.is_available()` alone. At minimum, run a PyTorch tensor
+operation, xformers memory-efficient attention, and a bitsandbytes operation.
+For ONNX Runtime updates, create an inference session with
+`CUDAExecutionProvider`, run a small model, and confirm that the session did
+not fall back to `CPUExecutionProvider`.
+
+Also run the WD14 ONNX captioning path with an actual image and model. This
+loads the full accelerate, timm, ONNX, and ONNX Runtime path that previously
+passed import-only checks but failed at startup:
+
+```shell
+docker run --rm --gpus all -v /path/to/smoke-data:/work \
+  sd-scripts:update-test \
+  finetune/tag_images_by_wd14_tagger.py --onnx --batch_size 1 \
+  --model_dir /work/model /work/images
+```
+
 For a release that changes PyTorch, CUDA, xformers, or sd-scripts training
 behavior, run at least one small project-specific training or captioning
 workflow before merging when practical.
@@ -383,6 +409,8 @@ Confirm:
 - `uv.lock` resolves the selected PyTorch CUDA index.
 - Local extras are intentionally kept, updated, or removed.
 - Upstream sd-scripts pytest release tests passed in the built image.
+- PyTorch, xformers, bitsandbytes, and ONNX Runtime executed GPU operations,
+  and the WD14 ONNX captioning path completed without provider fallback.
 - Every explicit pytest path in `scripts/run-sd-scripts-release-tests.sh`
   exists in the selected upstream checkout.
 - Upstream inpainting shell tests were run, or GPU/checkpoint coverage was

pyproject.toml

@@ -31,7 +31,7 @@ dependencies = [
   "timm==0.6.12",
   "fairscale==0.4.13",
   # for WD14 captioning (onnx)
-  "onnx==1.21.0",
+  "onnx==1.22.0",
   "onnxruntime-gpu==1.23.2",
   # open clip for SDXL
   "open-clip-torch==2.26.1",
@@ -40,9 +40,9 @@ dependencies = [
   # for T5XXL tokenizer (SD3/FLUX)
   "sentencepiece==0.2.1",
   # PyTorch
-  "torch==2.8.0+cu129",
-  "torchvision==0.23.0+cu129",
-  "xformers==0.0.32.post2",
+  "torch==2.13.0+cu130",
+  "torchvision==0.28.0+cu130",
+  "xformers==0.0.35",
   # Additional bundle
   # https://github.com/KohakuBlueleaf/LyCORIS
   "lycoris-lora==3.1.1.post1",
@@ -53,14 +53,14 @@ package = false
 exclude-newer = "P7D"
 
 [[tool.uv.index]]
-name = "pytorch-cu129"
-url = "https://download.pytorch.org/whl/cu129"
+name = "pytorch-cu130"
+url = "https://download.pytorch.org/whl/cu130"
 explicit = true
 
 [tool.uv.sources]
-torch = { index = "pytorch-cu129" }
-torchvision = { index = "pytorch-cu129" }
-xformers = { index = "pytorch-cu129" }
+torch = { index = "pytorch-cu130" }
+torchvision = { index = "pytorch-cu130" }
+xformers = { index = "pytorch-cu130" }
 
 [dependency-groups]
 dev = [