vlm-toolkit 0.1.0__tar.gz

vlm_toolkit-0.1.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Ben Lambert
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

vlm_toolkit-0.1.0/PKG-INFO

@@ -0,0 +1,264 @@
+Metadata-Version: 2.4
+Name: vlm-toolkit
+Version: 0.1.0
+Summary: Batch image processing with vision language models: captioning, logo and watermark detection, and logo removal
+Keywords: vlm,vision-language-model,captioning,image-processing,qwen-vl
+Author: Ben Lambert
+Author-email: Ben Lambert <blambert@gmail.com>
+License-Expression: MIT
+License-File: LICENSE
+Classifier: Development Status :: 3 - Alpha
+Classifier: Environment :: Console
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Programming Language :: Python :: 3.14
+Classifier: Topic :: Multimedia :: Graphics
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Dist: click>=8.1.0
+Requires-Dist: httpx>=0.28.1
+Requires-Dist: pillow>=12.0.0
+Requires-Dist: tqdm>=4.67.1
+Requires-Dist: typer>=0.21.0
+Requires-Dist: vlm-toolkit[local,logo,hub] ; extra == 'all'
+Requires-Dist: vlm-toolkit[local,logo,hub] ; extra == 'dev'
+Requires-Dist: black>=24.10.0 ; extra == 'dev'
+Requires-Dist: isort>=5.13.2 ; extra == 'dev'
+Requires-Dist: mypy>=1.14.0 ; extra == 'dev'
+Requires-Dist: pytest>=8.3.4 ; extra == 'dev'
+Requires-Dist: pytest-cov>=7.1.0 ; extra == 'dev'
+Requires-Dist: datasets>=3.2.0 ; extra == 'hub'
+Requires-Dist: accelerate>=1.12.0 ; extra == 'local'
+Requires-Dist: kernels>=0.12.3 ; extra == 'local'
+Requires-Dist: xgrammar>=0.2.6rc1 ; extra == 'local'
+Requires-Dist: torch>=2.9.1 ; extra == 'local'
+Requires-Dist: torchvision>=0.25.0 ; extra == 'local'
+Requires-Dist: transformers>=5.5.0 ; extra == 'local'
+Requires-Dist: numpy>=2.2.1 ; extra == 'logo'
+Requires-Dist: opencv-python>=4.10.0 ; extra == 'logo'
+Requires-Python: >=3.11
+Project-URL: Homepage, https://github.com/belambert/vlm-toolkit
+Project-URL: Repository, https://github.com/belambert/vlm-toolkit
+Project-URL: Issues, https://github.com/belambert/vlm-toolkit/issues
+Provides-Extra: all
+Provides-Extra: dev
+Provides-Extra: hub
+Provides-Extra: local
+Provides-Extra: logo
+Description-Content-Type: text/markdown
+
+# vlm-toolkit
+
+[![Lint](https://github.com/belambert/vlm-toolkit/actions/workflows/lint.yml/badge.svg?branch=main)](https://github.com/belambert/vlm-toolkit/actions/workflows/lint.yml)
+[![Test](https://github.com/belambert/vlm-toolkit/actions/workflows/test.yml/badge.svg?branch=main)](https://github.com/belambert/vlm-toolkit/actions/workflows/test.yml)
+[![codecov](https://codecov.io/gh/belambert/vlm-toolkit/graph/badge.svg)](https://codecov.io/gh/belambert/vlm-toolkit)
+
+Batch image processing with vision language models: captioning, logo and
+watermark detection, and logo removal.
+
+## Installation
+
+    pip install vlm-toolkit
+
+The base install is small: it covers `vlm server`, `vlm view-output`, and the
+shared plumbing. The heavier pieces are optional extras, so you only pay for
+what you use:
+
+| Extra   | Pulls in                      | Needed by                                  |
+| ------- | ----------------------------- | ------------------------------------------ |
+| `local` | torch, transformers, xgrammar | `process`, `caption-images`, `detect-logo` |
+| `logo`  | opencv, numpy                 | `remove-logo`                              |
+| `hub`   | datasets                      | `upload-dataset`                           |
+| `all`   | all of the above              | everything                                 |
+
+    pip install 'vlm-toolkit[local]'
+    pip install 'vlm-toolkit[all]'
+
+Running a command whose extra is missing tells you which one to install rather
+than raising an ImportError.
+
+For a source checkout:
+
+    uv sync --extra dev
+
+## Commands
+
+Everything is a subcommand of `vlm`; run `vlm --help` for the list. From a
+source checkout, prefix with `uv run`.
+
+### process
+
+Run a VLM over a folder of images locally.
+
+    vlm process <image-dir>
+    vlm process <image-dir> --batch-size 8 --max-dim 1024
+    vlm process <image-dir> --prompt "Describe this image in one sentence"
+    vlm process <image-dir> --prompt-file my_prompt.txt
+    vlm process <image-dir> --schema schema.json
+
+`--schema` takes a JSON schema file and constrains decoding to match it, via
+xgrammar. `--prompt-file` overrides `--prompt`. Output is JSONL, written to
+`--output`.
+
+### server
+
+Same idea, but against a running OpenAI-compatible vision endpoint instead of a
+local model.
+
+    vlm server <image-dir> --base-url http://localhost:8000/v1
+    vlm server <image-dir> --concurrency 8 --max-tokens 512
+
+### caption-images
+
+`vlm process` preset that captions with `prompts/caption.txt`, writing
+`captions.json` into the image folder by default.
+
+    vlm caption-images <image-dir>
+
+### detect-logo
+
+`vlm process` preset that finds logos and watermarks using
+`prompts/detect_logo.txt`, writing `logo_bbox_output.json` into the image folder
+by default.
+
+    vlm detect-logo <image-dir>
+
+### remove-logo
+
+Erases the boxes found by `detect-logo`. Takes that command's JSON output, not
+an image folder. Two methods: `gray` replaces each box with a gray rectangle,
+`opencv` inpaints it (the default).
+
+    vlm remove-logo <image-dir>/logo_bbox_output.json <output-dir>
+    vlm remove-logo <image-dir>/logo_bbox_output.json <output-dir> --method gray
+
+### upload-dataset
+
+Publishes a `vlm process` run to the Hugging Face Hub as an image dataset.
+
+    vlm upload-dataset <output.jsonl> <user>/<dataset>
+    vlm upload-dataset <output.jsonl> <user>/<dataset> --no-private --split test
+    vlm upload-dataset <output.jsonl> <user>/<dataset> --card CARD.md
+
+The images are embedded in the dataset rather than referenced by path, so the
+result is self-contained and the Hub's dataset viewer works. Columns are
+`image`, `output`, and the original `file_name`. Rows whose image is missing
+from disk are reported and skipped.
+
+Repos are created private by default; pass `--no-private` for a public one.
+Authentication comes from `--token`, else `HF_TOKEN`, else a cached
+`huggingface-cli login`.
+
+`--card` attaches a markdown file as the dataset's README. The upload generates
+a `dataset_info` block that the Hub viewer depends on, so the card is merged
+rather than overwritten: your prose becomes the body, and any YAML frontmatter
+in your file (`license`, `task_categories`, …) is layered on top of the
+generated keys. The card path is checked before the upload starts, so a typo
+fails immediately instead of after transferring the images.
+
+### view-output
+
+Renders a JSONL output file as an HTML page and opens it in a browser.
+
+    vlm view-output <output.jsonl>
+    vlm view-output <output.jsonl> --serve --port 8000
+    vlm view-output <output.jsonl> --serve --host 0.0.0.0
+
+By default it writes `<output.jsonl>.html` next to the input and opens it over
+`file://`, with absolute paths to the images. `--serve` skips the file and hosts
+the page instead, which is what you want when the images sit on a remote
+machine.
+
+The server roots itself at the closest directory containing both the JSONL and
+every image it references, so `file_name` entries that point outside the JSON's
+own directory still resolve.
+
+`--host` controls the bind address. It defaults to `127.0.0.1`, so only the
+local machine can connect; `--host 0.0.0.0` accepts external connections and
+prints the LAN URL to open from another machine. There is no authentication and
+the whole server root is readable, so on an untrusted network prefer the default
+and forward the port over SSH instead:
+
+    ssh -L 8000:localhost:8000 <remote>
+
+## Prompts
+
+The presets read their prompts from `src/vlm_toolkit/prompts/`, which ships as
+package data, so they work from an installed wheel as well as a source checkout.
+To use your own prompt, pass `vlm process --prompt-file <path>`.
+
+The viewer page is packaged the same way. Its markup and CSS live in
+`src/vlm_toolkit/templates/` as `viewer.html` (the shell, with an `$items`
+placeholder) and `item.html` (one image plus its output), so restyling the
+viewer means editing HTML rather than a Python string.
+
+## How Batching Works
+
+`--batch-size` sets how many images go through the model at once. Images are
+never resized to a common shape first, so a batch can mix sizes freely. Two
+different mechanisms handle the resulting variation.
+
+**Images are concatenated, not padded.** Qwen-VL processors resize each image on
+its own, preserving aspect ratio and rounding to the processor's patch multiple,
+then flatten it into a sequence of patches and concatenate every image's patches
+into one 2D tensor. `pixel_values` has shape `(total_patches, feature_dim)` —
+there is no per-image batch dimension, so there is nothing to pad. A companion
+`image_grid_thw` records each image's `(t, h, w)` grid so the model can split
+them apart again. A batch of three images with grids `[1,18,18]`, `[1,38,28]`
+and `[1,22,14]` produces 1696 patch rows:
+
+```
+18×18 =  324
+38×28 = 1064
+22×14 =  308
+        ----
+        1696
+```
+
+Bigger images simply cost more tokens; none are spent on padding. Use
+`--max-dim` to cap the longest edge before the processor sees the image, which
+is the practical lever on both memory and token count.
+
+**Text is left padded.** Each image expands into placeholder tokens
+proportional to its patch grid, so differently sized images yield different
+prompt lengths, and `padding=True` pads them to a common length.
+`load_model` sets `padding_side = "left"` for two reasons: decoder-only
+generation needs the last real token flush against the end of the sequence, and
+`_run_inference` trims prompts with a single `out_ids[len(in_ids):]` offset
+applied to every row. That slice is only correct when the padding sits on the
+left — with right padding it would cut at the wrong point and leak pad and
+prompt tokens into the decoded text.
+
+**Batches are prefetched.** `vlm process` prepares the next batch on the CPU in
+a background thread while the current one runs inference. Only the main thread
+moves tensors onto the device.
+
+The concatenated-patch layout above is specific to the Qwen-VL family. Models
+that resize to a fixed square instead, such as the gemma entries in the
+suggested list, produce a conventional stacked 4D `pixel_values`. Both work,
+because images are handed to the processor one per prompt and each processor
+applies its own preprocessing.
+
+## Environment Variables
+
+- `HF_TOKEN` - Hugging Face token, for gated models (optional)
+
+## Notes
+
+Detect logo took 1 hour to do 5000 imgs on an L4 GPU at ~$1/hour.
+
+## Releasing
+
+Bump `version` in `pyproject.toml`, then push a matching tag:
+
+    git tag v0.1.0 && git push origin v0.1.0
+
+`publish.yml` checks the tag against the project version, builds, and publishes
+to PyPI via trusted publishing.
+
+## License
+
+MIT - see [LICENSE](LICENSE).

vlm_toolkit-0.1.0/README.md

@@ -0,0 +1,212 @@
+# vlm-toolkit
+
+[![Lint](https://github.com/belambert/vlm-toolkit/actions/workflows/lint.yml/badge.svg?branch=main)](https://github.com/belambert/vlm-toolkit/actions/workflows/lint.yml)
+[![Test](https://github.com/belambert/vlm-toolkit/actions/workflows/test.yml/badge.svg?branch=main)](https://github.com/belambert/vlm-toolkit/actions/workflows/test.yml)
+[![codecov](https://codecov.io/gh/belambert/vlm-toolkit/graph/badge.svg)](https://codecov.io/gh/belambert/vlm-toolkit)
+
+Batch image processing with vision language models: captioning, logo and
+watermark detection, and logo removal.
+
+## Installation
+
+    pip install vlm-toolkit
+
+The base install is small: it covers `vlm server`, `vlm view-output`, and the
+shared plumbing. The heavier pieces are optional extras, so you only pay for
+what you use:
+
+| Extra   | Pulls in                      | Needed by                                  |
+| ------- | ----------------------------- | ------------------------------------------ |
+| `local` | torch, transformers, xgrammar | `process`, `caption-images`, `detect-logo` |
+| `logo`  | opencv, numpy                 | `remove-logo`                              |
+| `hub`   | datasets                      | `upload-dataset`                           |
+| `all`   | all of the above              | everything                                 |
+
+    pip install 'vlm-toolkit[local]'
+    pip install 'vlm-toolkit[all]'
+
+Running a command whose extra is missing tells you which one to install rather
+than raising an ImportError.
+
+For a source checkout:
+
+    uv sync --extra dev
+
+## Commands
+
+Everything is a subcommand of `vlm`; run `vlm --help` for the list. From a
+source checkout, prefix with `uv run`.
+
+### process
+
+Run a VLM over a folder of images locally.
+
+    vlm process <image-dir>
+    vlm process <image-dir> --batch-size 8 --max-dim 1024
+    vlm process <image-dir> --prompt "Describe this image in one sentence"
+    vlm process <image-dir> --prompt-file my_prompt.txt
+    vlm process <image-dir> --schema schema.json
+
+`--schema` takes a JSON schema file and constrains decoding to match it, via
+xgrammar. `--prompt-file` overrides `--prompt`. Output is JSONL, written to
+`--output`.
+
+### server
+
+Same idea, but against a running OpenAI-compatible vision endpoint instead of a
+local model.
+
+    vlm server <image-dir> --base-url http://localhost:8000/v1
+    vlm server <image-dir> --concurrency 8 --max-tokens 512
+
+### caption-images
+
+`vlm process` preset that captions with `prompts/caption.txt`, writing
+`captions.json` into the image folder by default.
+
+    vlm caption-images <image-dir>
+
+### detect-logo
+
+`vlm process` preset that finds logos and watermarks using
+`prompts/detect_logo.txt`, writing `logo_bbox_output.json` into the image folder
+by default.
+
+    vlm detect-logo <image-dir>
+
+### remove-logo
+
+Erases the boxes found by `detect-logo`. Takes that command's JSON output, not
+an image folder. Two methods: `gray` replaces each box with a gray rectangle,
+`opencv` inpaints it (the default).
+
+    vlm remove-logo <image-dir>/logo_bbox_output.json <output-dir>
+    vlm remove-logo <image-dir>/logo_bbox_output.json <output-dir> --method gray
+
+### upload-dataset
+
+Publishes a `vlm process` run to the Hugging Face Hub as an image dataset.
+
+    vlm upload-dataset <output.jsonl> <user>/<dataset>
+    vlm upload-dataset <output.jsonl> <user>/<dataset> --no-private --split test
+    vlm upload-dataset <output.jsonl> <user>/<dataset> --card CARD.md
+
+The images are embedded in the dataset rather than referenced by path, so the
+result is self-contained and the Hub's dataset viewer works. Columns are
+`image`, `output`, and the original `file_name`. Rows whose image is missing
+from disk are reported and skipped.
+
+Repos are created private by default; pass `--no-private` for a public one.
+Authentication comes from `--token`, else `HF_TOKEN`, else a cached
+`huggingface-cli login`.
+
+`--card` attaches a markdown file as the dataset's README. The upload generates
+a `dataset_info` block that the Hub viewer depends on, so the card is merged
+rather than overwritten: your prose becomes the body, and any YAML frontmatter
+in your file (`license`, `task_categories`, …) is layered on top of the
+generated keys. The card path is checked before the upload starts, so a typo
+fails immediately instead of after transferring the images.
+
+### view-output
+
+Renders a JSONL output file as an HTML page and opens it in a browser.
+
+    vlm view-output <output.jsonl>
+    vlm view-output <output.jsonl> --serve --port 8000
+    vlm view-output <output.jsonl> --serve --host 0.0.0.0
+
+By default it writes `<output.jsonl>.html` next to the input and opens it over
+`file://`, with absolute paths to the images. `--serve` skips the file and hosts
+the page instead, which is what you want when the images sit on a remote
+machine.
+
+The server roots itself at the closest directory containing both the JSONL and
+every image it references, so `file_name` entries that point outside the JSON's
+own directory still resolve.
+
+`--host` controls the bind address. It defaults to `127.0.0.1`, so only the
+local machine can connect; `--host 0.0.0.0` accepts external connections and
+prints the LAN URL to open from another machine. There is no authentication and
+the whole server root is readable, so on an untrusted network prefer the default
+and forward the port over SSH instead:
+
+    ssh -L 8000:localhost:8000 <remote>
+
+## Prompts
+
+The presets read their prompts from `src/vlm_toolkit/prompts/`, which ships as
+package data, so they work from an installed wheel as well as a source checkout.
+To use your own prompt, pass `vlm process --prompt-file <path>`.
+
+The viewer page is packaged the same way. Its markup and CSS live in
+`src/vlm_toolkit/templates/` as `viewer.html` (the shell, with an `$items`
+placeholder) and `item.html` (one image plus its output), so restyling the
+viewer means editing HTML rather than a Python string.
+
+## How Batching Works
+
+`--batch-size` sets how many images go through the model at once. Images are
+never resized to a common shape first, so a batch can mix sizes freely. Two
+different mechanisms handle the resulting variation.
+
+**Images are concatenated, not padded.** Qwen-VL processors resize each image on
+its own, preserving aspect ratio and rounding to the processor's patch multiple,
+then flatten it into a sequence of patches and concatenate every image's patches
+into one 2D tensor. `pixel_values` has shape `(total_patches, feature_dim)` —
+there is no per-image batch dimension, so there is nothing to pad. A companion
+`image_grid_thw` records each image's `(t, h, w)` grid so the model can split
+them apart again. A batch of three images with grids `[1,18,18]`, `[1,38,28]`
+and `[1,22,14]` produces 1696 patch rows:
+
+```
+18×18 =  324
+38×28 = 1064
+22×14 =  308
+        ----
+        1696
+```
+
+Bigger images simply cost more tokens; none are spent on padding. Use
+`--max-dim` to cap the longest edge before the processor sees the image, which
+is the practical lever on both memory and token count.
+
+**Text is left padded.** Each image expands into placeholder tokens
+proportional to its patch grid, so differently sized images yield different
+prompt lengths, and `padding=True` pads them to a common length.
+`load_model` sets `padding_side = "left"` for two reasons: decoder-only
+generation needs the last real token flush against the end of the sequence, and
+`_run_inference` trims prompts with a single `out_ids[len(in_ids):]` offset
+applied to every row. That slice is only correct when the padding sits on the
+left — with right padding it would cut at the wrong point and leak pad and
+prompt tokens into the decoded text.
+
+**Batches are prefetched.** `vlm process` prepares the next batch on the CPU in
+a background thread while the current one runs inference. Only the main thread
+moves tensors onto the device.
+
+The concatenated-patch layout above is specific to the Qwen-VL family. Models
+that resize to a fixed square instead, such as the gemma entries in the
+suggested list, produce a conventional stacked 4D `pixel_values`. Both work,
+because images are handed to the processor one per prompt and each processor
+applies its own preprocessing.
+
+## Environment Variables
+
+- `HF_TOKEN` - Hugging Face token, for gated models (optional)
+
+## Notes
+
+Detect logo took 1 hour to do 5000 imgs on an L4 GPU at ~$1/hour.
+
+## Releasing
+
+Bump `version` in `pyproject.toml`, then push a matching tag:
+
+    git tag v0.1.0 && git push origin v0.1.0
+
+`publish.yml` checks the tag against the project version, builds, and publishes
+to PyPI via trusted publishing.
+
+## License
+
+MIT - see [LICENSE](LICENSE).

vlm_toolkit-0.1.0/pyproject.toml

@@ -0,0 +1,114 @@
+[project]
+name = "vlm-toolkit"
+version = "0.1.0"
+description = "Batch image processing with vision language models: captioning, logo and watermark detection, and logo removal"
+readme = "README.md"
+license = "MIT"
+license-files = ["LICENSE"]
+keywords = [
+    "vlm",
+    "vision-language-model",
+    "captioning",
+    "image-processing",
+    "qwen-vl",
+]
+classifiers = [
+    "Development Status :: 3 - Alpha",
+    "Environment :: Console",
+    "Intended Audience :: Developers",
+    "Intended Audience :: Science/Research",
+    "Programming Language :: Python :: 3",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
+    "Programming Language :: Python :: 3.13",
+    "Programming Language :: Python :: 3.14",
+    "Topic :: Multimedia :: Graphics",
+    "Topic :: Scientific/Engineering :: Artificial Intelligence",
+]
+requires-python = ">=3.11"
+dependencies = [
+    "click>=8.1.0",
+    "httpx>=0.28.1",
+    "pillow>=12.0.0",
+    "tqdm>=4.67.1",
+    "typer>=0.21.0",
+]
+
+[[project.authors]]
+name = "Ben Lambert"
+email = "blambert@gmail.com"
+
+[project.urls]
+Homepage = "https://github.com/belambert/vlm-toolkit"
+Repository = "https://github.com/belambert/vlm-toolkit"
+Issues = "https://github.com/belambert/vlm-toolkit/issues"
+
+[project.optional-dependencies]
+local = [
+    "accelerate>=1.12.0",
+    "kernels>=0.12.3",
+    "xgrammar>=0.2.6rc1",
+    "torch>=2.9.1",
+    "torchvision>=0.25.0",
+    "transformers>=5.5.0",
+]
+logo = [
+    "numpy>=2.2.1",
+    "opencv-python>=4.10.0",
+]
+hub = ["datasets>=3.2.0"]
+all = ["vlm-toolkit[local,logo,hub]"]
+dev = [
+    "vlm-toolkit[local,logo,hub]",
+    "black>=24.10.0",
+    "isort>=5.13.2",
+    "mypy>=1.14.0",
+    "pytest>=8.3.4",
+    "pytest-cov>=7.1.0",
+]
+
+[project.scripts]
+vlm = "vlm_toolkit.cmd.main:cli"
+
+[[tool.uv.sources.torch]]
+index = "pytorch-cpu"
+marker = "sys_platform != 'linux'"
+
+[[tool.uv.sources.torch]]
+index = "pytorch-cuda"
+marker = "sys_platform == 'linux'"
+
+[[tool.uv.index]]
+name = "pytorch-cpu"
+url = "https://download.pytorch.org/whl/cpu"
+explicit = true
+
+[[tool.uv.index]]
+name = "pytorch-cuda"
+url = "https://download.pytorch.org/whl/cu130"
+explicit = true
+
+[tool.black]
+line-length = 88
+target-version = ["py311"]
+include = '\.pyi?$'
+
+[tool.isort]
+profile = "black"
+line_length = 88
+known_third_party = ["wandb"]
+
+[tool.mypy]
+python_version = "3.13"
+files = ["src"]
+ignore_missing_imports = true
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]
+python_files = ["test_*.py"]
+python_classes = ["Test*"]
+python_functions = ["test_*"]
+
+[build-system]
+requires = ["uv_build>=0.9.3,<0.10.0"]
+build-backend = "uv_build"