PyCosmoMMF 1.0.3__tar.gz → 2.0.0__tar.gz

{pycosmommf-1.0.3 → pycosmommf-2.0.0}/.github/workflows/cd.yml

@@ -29,7 +29,7 @@ jobs:
         with:
           fetch-depth: 0
 
-      - uses: hynek/build-and-inspect-python-package@v2
+      - uses: hynek/build-and-inspect-python-package@v3.0.2
 
   publish:
     needs: [dist]

{pycosmommf-1.0.3 → pycosmommf-2.0.0}/.github/workflows/ci.yml

@@ -38,8 +38,8 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python-version: ["3.8", "3.12"]
-        runs-on: [ubuntu-latest, windows-latest, macos-latest, macos-13]
+        python-version: ["3.9", "3.12"]
+        runs-on: [ubuntu-latest, windows-latest, macos-latest]
 
     steps:
       - uses: actions/checkout@v4

{pycosmommf-1.0.3 → pycosmommf-2.0.0}/.gitignore

@@ -156,3 +156,7 @@ Thumbs.db
 # Common editor files
 *~
 *.swp
+
+# AGENTS
+CLAUDE.md
+AGENTS.md

pycosmommf-2.0.0/CHANGELOG.md

@@ -0,0 +1,224 @@
+# Changelog
+
+All notable changes to `pycosmommf` are documented here. Version numbers come
+from git tags (`hatch-vcs`); releases are published on
+[PyPI](https://pypi.org/project/pycosmommf/).
+
+## Version 2.0.0
+
+Changes since [1.0.3](https://pypi.org/project/pycosmommf/1.0.3/), the latest
+release on PyPI. This release changes the return value of
+`calc_structure_bools()` (see Breaking changes below).
+
+### Highlights
+
+- **GPU backend.** `maximum_signature()` and `calc_structure_bools()` take a new
+  `backend` argument. `backend="gpu"` runs the signatures with JAX and the
+  structure tagging with CuPy. Install it with `pip install 'pycosmommf[gpu]'`.
+  The GPU backend is capable of computing the cosmic web structure tags on a
+  1024^3 grid in just a couple minutes.
+- **Faster CPU signatures.** The Hessian eigenvalues are now computed with an
+  analytic closed-form solver instead of a per-voxel `np.linalg.eigvals` call,
+  and the FFTs run in complex64. `maximum_signature()` on the CPU is about 4×
+  faster at 128³.
+- **One structure filter.** `calc_structure_bools()` returns a single `uint8`
+  array that labels every voxel, instead of four boolean arrays.
+- **Tagging bug fixes.** A race condition in the cluster finder dropped voxels
+  from large clumps, and a threshold with no clumps could be selected as the
+  cluster threshold. Both are fixed, so cluster filters can differ from 1.0.3.
+
+### Breaking changes
+
+- `calc_structure_bools()` returns one `uint8` array, `structure_filter`, with
+  the same shape as the density field. Each voxel holds its structure type: `1`
+  = cluster, `2` = filament, `3` = wall, `4` = void. Every voxel has exactly one
+  type, so the array has no zeros. With `verbose_flag=True` the function returns
+  `(structure_filter, summary_data)`.
+
+  To migrate code written for 1.0.3:
+
+  ```python
+  # 1.0.3
+  clusbool, filbool, wallbool, voidbool = calc_structure_bools(rho, sigs, False)
+
+  # now
+  structure_filter = calc_structure_bools(rho, sigs, False)
+  clusbool = structure_filter == 1
+  filbool = structure_filter == 2
+  wallbool = structure_filter == 3
+  voidbool = structure_filter == 4
+  ```
+
+  The `uint8` filter uses one byte per voxel. The four boolean arrays used four.
+
+- `maximum_signature(..., algorithm="NEXUSPLUS")` raises `ValueError` if the
+  density field has a negative voxel. NEXUS+ smooths `log10(density)`, so a
+  negative input (usually an overdensity δ passed instead of δ+1 = ρ/⟨ρ⟩) used
+  to fill the whole output with NaN without an error. The error message tells
+  you to pass `density + 1`. `algorithm="NEXUS"` still accepts any values.
+
+- All computation and all returned numbers are float32 on both backends. Inputs
+  are converted to float32. The thresholds, virialized fractions, mass curves,
+  and volume and mass fractions in `summary_data` are float32. In 1.0.3 some
+  were float64.
+
+### Added
+
+#### GPU backend
+
+- New `backend` argument for `maximum_signature()` and `calc_structure_bools()`.
+  It takes `"cpu"` (default) or `"gpu"`. Any other value raises `ValueError`.
+  - `backend="cpu"` uses only NumPy, SciPy, Numba and scikit-image. It never
+    imports JAX or CuPy.
+  - `backend="gpu"` uses JAX for `maximum_signature()` and CuPy for
+    `calc_structure_bools()`.
+- New install extra `gpu`: `pip install 'pycosmommf[gpu]'` installs
+  `jax[cuda12]>=0.4.30` and `cupy-cuda12x`. JAX and CuPy are imported only when
+  `backend="gpu"` is used. If they are missing, you get an `ImportError` that
+  gives the install command.
+- New `maximum_signature(..., backend="gpu", return_numpy=False)` keeps the
+  signatures on the GPU as a JAX array. Pass that array directly to
+  `calc_structure_bools(..., backend="gpu")`: CuPy reads it through DLPack
+  without a copy. With the default `return_numpy=True`, the result is a NumPy
+  array, the same as the CPU backend.
+- `calc_structure_bools(..., backend="gpu")` accepts NumPy, CuPy or JAX GPU
+  arrays for `density_cube`, `max_sigs` and `clusbool`. It always returns NumPy
+  arrays.
+
+Typical GPU use:
+
+```python
+import pycosmommf
+
+sigs = pycosmommf.maximum_signature(Rs, rho, backend="gpu", return_numpy=False)
+structure_filter = pycosmommf.calc_structure_bools(rho, sigs, False, backend="gpu")
+```
+
+How the GPU signatures (`_jax_backend.py`) work:
+
+- The whole per-scale step (smoothing, Hessian, eigenvalues, signatures and
+  running maximum) is one `jax.jit` graph. The running maximum buffer is donated
+  (`donate_argnums`), so XLA updates it in place.
+- The eigenvalues use the same analytic solver as the CPU backend (see
+  Performance below). There is no `(N, N, N, 3, 3)` tensor and no call to
+  `jnp.linalg.eigh`.
+- The six Hessian components stay as six separate arrays, and the signature
+  stage is tiled along the first axis for grids larger than 128³.
+- Peak GPU memory is about 9–10× the size of one float32 field. A 512³ NEXUS+
+  run needs 11.8 GiB and takes 5.4 s on a V100S; a naive JAX port ran out of
+  memory on a 32 GB GPU. A 1024³ run, the largest supported grid, needs about 65
+  GiB, so it needs an 80 GB GPU. A 40 GB GPU is not enough at 1024³. The backend
+  runs on a single GPU.
+- On Apple `jax-metal`, which has no FFT or complex support, the backend warns
+  and runs JAX on the CPU.
+
+How the GPU tagging (`_cupy_tagging.py`) works:
+
+- It follows the CPU tagging step for step and calls the same helper functions
+  for the threshold choices, so both backends make the same decisions.
+- Clumps are labelled with `cupyx.scipy.ndimage.label` and 26-connectivity, the
+  same connectivity as the CPU backend. Per-clump sums use GPU atomics over the
+  foreground voxels only.
+- Each point of a mass change curve is one fused float32 reduction kernel, with
+  no masks or compacted copies.
+- CuPy allocates from CUDA's stream-ordered memory pool while it tags. This lets
+  it reuse memory that JAX freed in the same process. Peak GPU memory is about
+  7× one float32 field (about 29 GiB at 1024³).
+- At 1024³, tagging takes 3.8 s on an A100, against 334 s on the CPU.
+- On the 128³, 256³ and 1024³ validation cubes, GPU and CPU tagging gave the
+  same thresholds and the same structure filters. In general the two backends
+  agree up to float32 summation order, not bit for bit.
+
+Recommended settings for large grids: set `XLA_PYTHON_CLIENT_PREALLOCATE=false`
+and `XLA_PYTHON_CLIENT_ALLOCATOR=cuda_async` before you import JAX. The default
+JAX allocator fragments memory and runs out of memory at 1024³.
+
+GPU and CPU signatures differ by about 10⁻⁶ (relative) after smoothing and by up
+to a few 10⁻³ (absolute) in the final signatures, because XLA and SciPy use
+different float32 FFTs. These differences do not change the structure tagging.
+
+#### Other additions
+
+- If no clump is virialized at any cluster-signature threshold, the cluster
+  filter is empty and a warning says that the grid resolution is too low to
+  resolve virialized clusters. `make_the_clusbool()` then returns `None` as the
+  threshold. In 1.0.3 the function picked a threshold anyway: the lowest one, or
+  one with no clumps (see Fixed below).
+
+### Performance
+
+- **Analytic eigenvalues on the CPU.** `signatures_from_hessian()` used to build
+  a 3×3 array for every voxel and call `np.linalg.eigvals` and `np.sort` on it.
+  It now calls `_eigvalsh_3x3_sym`, a scalar Numba function that uses the
+  closed-form trigonometric solution for real symmetric 3×3 matrices (Smith
+  1961, _Communications of the ACM_ 4, 168). The function allocates nothing per
+  voxel and returns the eigenvalues already sorted. It rescales each matrix by
+  its largest absolute entry before it applies the formula, so it stays accurate
+  for near-zero Hessians in voids. The CPU and GPU backends use the same solver.
+- **complex64 FFTs on the CPU.** Smoothing and the Hessian use `scipy.fft`
+  instead of `np.fft`. `scipy.fft` keeps float32 input in complex64. With numpy
+  < 2, `np.fft` always computes in complex128.
+- **Fewer float64 promotions.** Numba kernels use float32 constants, so they no
+  longer promote float32 arithmetic to float64.
+
+CPU timings, 1.0.3 against the current code, on a synthetic log-normal 128³
+field with NEXUS+, on one node with 128 CPU cores. Each step was timed on its
+second call, so the Numba compilation of module-level functions is excluded:
+
+| Step                                    | 1.0.3  | Now    | Speedup |
+| --------------------------------------- | ------ | ------ | ------- |
+| `smooth_loggauss()`, 1 scale            | 1.00 s | 0.23 s | 4.3×    |
+| `fast_hessian_from_smoothed()`, 1 scale | 4.68 s | 1.02 s | 4.6×    |
+| `signatures_from_hessian()`, 1 scale    | 0.49 s | 0.17 s | 2.9×    |
+| `maximum_signature()`, 1 scale          | 6.47 s | 1.60 s | 4.0×    |
+| `maximum_signature()`, 10 scales        | 58.9 s | 13.8 s | 4.3×    |
+
+At 256³ the current code takes 53 s for 10 scales. The 1.0.3 run at 256³ did not
+finish (the process was killed). On an A100 GPU, the same 256³ field takes 1.7 s
+for 10 scales with `backend="gpu"` (9.7 s on the first call, which includes JIT
+compilation), and the GPU tagging takes 0.08 s against 5.9 s on the CPU, with an
+identical structure filter.
+
+### Fixed
+
+- **Race condition in the cluster finder.** The per-clump overdensity sums ran
+  in a Numba `prange` loop that added into shared arrays. Threads overwrote each
+  other's additions and dropped voxels from large clumps, so the mean
+  overdensity of a clump, and therefore the cluster threshold, could be wrong
+  and change from run to run. The sums are now a serial float32 loop
+  (`_clump_sums`).
+- **Thresholds without clumps.** If a cluster-signature threshold had no clumps,
+  its virialized fraction was 0/0 = NaN. `np.argmin` returns the first NaN, so
+  that empty threshold could be selected as the cluster threshold. These
+  fractions are now NaN on purpose, and `np.nanargmin` skips them.
+- **Threshold comparisons.** The thresholds were float64 and the signatures were
+  float32. Thresholds are now float32, so the comparisons are exact and do not
+  depend on NumPy's scalar promotion rules.
+
+### Dependencies and packaging
+
+- `scipy` is now a declared dependency. The CPU FFTs use `scipy.fft`.
+- New optional extra `gpu` (`jax[cuda12]>=0.4.30`, `cupy-cuda12x`).
+- The `requires-python` upper bound (`<3.13`) is removed. The requirement is now
+  `>=3.8`.
+
+### Internal changes
+
+- `calc_structure_bools()` separates the computation from the printed report.
+  The CPU and GPU backends share the report and the `summary_data` code.
+- `wavevectors3D()` builds the wavevectors directly in float32. The values are
+  unchanged.
+- `sphere()`, `cylinder()` and `wall()` in `utils.py` return float32 arrays.
+
+### Documentation and tests
+
+- `docs/basic_usage.md` describes the new `structure_filter` return value and
+  the GPU backend.
+- New `tests/test_tagging.py` (CPU tagging), `tests/test_jax_backend.py` (JAX vs
+  CPU agreement for each stage) and `tests/test_cupy_tagging.py` (CuPy vs CPU
+  tagging with NumPy, CuPy and JAX inputs). The GPU tests skip when JAX, CuPy or
+  a CUDA GPU is not available, so CI runs only the CPU tests.
+
+## 1.0.3 and earlier
+
+See the [GitHub releases](https://github.com/James11222/pycosmommf/releases).

{pycosmommf-1.0.3 → pycosmommf-2.0.0}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.4
+Metadata-Version: 2.5
 Name: pycosmommf
-Version: 1.0.3
+Version: 2.0.0
 Summary: A package for identifying structures in the cosmic web.
 Project-URL: Homepage, https://github.com/James11222/pycosmommf
 Project-URL: Bug Tracker, https://github.com/James11222/pycosmommf/issues
@@ -27,7 +27,7 @@ License: Copyright 2024 James Sunseri
         OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
         SOFTWARE.
 License-File: LICENSE
-Classifier: Development Status :: 1 - Planning
+Classifier: Development Status :: 5 - Production/Stable
 Classifier: Intended Audience :: Developers
 Classifier: Intended Audience :: Science/Research
 Classifier: License :: OSI Approved :: MIT License
@@ -35,17 +35,17 @@ Classifier: Operating System :: OS Independent
 Classifier: Programming Language :: Python
 Classifier: Programming Language :: Python :: 3
 Classifier: Programming Language :: Python :: 3 :: Only
-Classifier: Programming Language :: Python :: 3.8
 Classifier: Programming Language :: Python :: 3.9
 Classifier: Programming Language :: Python :: 3.10
 Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3.12
 Classifier: Topic :: Scientific/Engineering
 Classifier: Typing :: Typed
-Requires-Python: <3.13,>=3.8
+Requires-Python: >=3.9
 Requires-Dist: numba
 Requires-Dist: numpy>=1.22
 Requires-Dist: scikit-image
+Requires-Dist: scipy
 Provides-Extra: dev
 Requires-Dist: pytest-cov>=3; extra == 'dev'
 Requires-Dist: pytest>=6; extra == 'dev'
@@ -55,6 +55,9 @@ Requires-Dist: myst-parser>=0.13; extra == 'docs'
 Requires-Dist: sphinx-autodoc-typehints; extra == 'docs'
 Requires-Dist: sphinx-copybutton; extra == 'docs'
 Requires-Dist: sphinx>=7.0; extra == 'docs'
+Provides-Extra: gpu
+Requires-Dist: cupy-cuda12x; extra == 'gpu'
+Requires-Dist: jax[cuda12]>=0.4.30; extra == 'gpu'
 Provides-Extra: test
 Requires-Dist: pytest-cov>=3; extra == 'test'
 Requires-Dist: pytest>=6; extra == 'test'

{pycosmommf-1.0.3 → pycosmommf-2.0.0}/docs/basic_usage.md

@@ -13,16 +13,16 @@ to analyze the effects of baryonic matter on the Cosmic Web.
 The NEXUS+ algorithm contains several steps as described in our paper
 ([Sunseri et al. 2022](https://ui.adsabs.harvard.edu/abs/2023PhRvD.107b3514S/abstract)).
 In general, we start with a density field (note: we more specifically mean a 1 +
-δ field), smooth it with a logarithmic Gaussian smoothing filter, then compute
-the hessian of the smoothed density field, use the eigenvalues of the hessian
-matrix to calculate the structure type signatures, find the maximum signatures
-over a range of smoothing scales, and apply physically based threshold criterion
-to categorize structures within the Cosmic Web. The entire package is
-implemented in `python` and all of these steps are summarized inside of two
-functions. The first function `maximum_signature()` does the first several steps
-of the NEXUS+ algorithm to compute the maximum structure signatures, the second
-function is `calc_structure_bools()` which uses physical criteria to tag
-structures into 4 categories: clusters, filaments, walls, and voids.
+δ field), smooth it with a log-Gaussian smoothing filter, then compute the
+hessian of the smoothed density field, use the eigenvalues of the hessian matrix
+to calculate the structure type signatures, find the maximum signatures over a
+range of smoothing scales, and apply physically based threshold criterion to
+categorize structures within the Cosmic Web. The entire package is implemented
+in `python` and all of these steps are summarized inside of two functions. The
+first function `maximum_signature()` does the first several steps of the NEXUS+
+algorithm to compute the maximum structure signatures, the second function is
+`calc_structure_bools()` which uses physical criteria to tag structures into 4
+categories: clusters, filaments, walls, and voids.
 
 We also make the data products from our paper
 ([Sunseri et al. 2022](https://ui.adsabs.harvard.edu/abs/2023PhRvD.107b3514S/abstract))
@@ -81,7 +81,7 @@ API Docs tab for a detailed summary of what is inside `summary_data`.
 verbose_flag = True  # or False
 
 
-clusbool, filbool, wallbool, voidbool, summary_data = PyCosmoMMF.calc_structure_bools(
+structure_filter, summary_data = PyCosmoMMF.calc_structure_bools(
     density_field, max_signatures, verbose_flag
 )  # tag structures
 ```
@@ -92,11 +92,23 @@ clusbool, filbool, wallbool, voidbool, summary_data = PyCosmoMMF.calc_structure_
 verbose_flag = False
 
 
-clusbool, filbool, wallbool, voidbool = PyCosmoMMF.calc_structure_bools(
+structure_filter = PyCosmoMMF.calc_structure_bools(
     density_field, max_signatures, verbose_flag
 )  # tag structures
 ```
 
+`structure_filter` is a `uint8` array with the same shape as the density field
+that gives the structure type of every voxel: `1` = cluster, `2` = filament, `3`
+= wall and `4` = void. Every voxel has exactly one type. Boolean filters for a
+single structure type are one comparison away:
+
+```python
+clusbool = structure_filter == 1
+filbool = structure_filter == 2
+wallbool = structure_filter == 3
+voidbool = structure_filter == 4
+```
+
 We also note in the `calc_structure_bools()` function, one can use their own
 cluster boolean filter instead of the one generated by the NEXUS+ formalism
 (using virialization of clusters as a tool for determining spurious detections).
@@ -114,7 +126,7 @@ clusbool_ext = np.load(
 verbose_flag = False
 
 
-clusbool, filbool, wallbool, voidbool = PyCosmoMMF.calc_structure_bools(
+structure_filter = PyCosmoMMF.calc_structure_bools(
     density_field, max_signatures, verbose_flag, clusbool_ext
 )  # tag structures
 ```
@@ -128,9 +140,8 @@ but other values can be 200 or 500 corresponding to `R_200` or `R_500`.
 average overdensity greater than this value, they are thought to be
 virialized/collapsed.
 
-The boolean filters for each structure type produced by `calc_structure_bools()`
-can be used to tag structures within a density field, the results of this can be
-seen below
+The structure filter produced by `calc_structure_bools()` can be used to tag
+structures within a density field, the results of this can be seen below
 
 ```{image} images/final_tagging_figure.png
 :class: only-light
@@ -141,6 +152,45 @@ seen below
 :class: only-dark
 ```
 
+#### Running on a GPU
+
+Both `maximum_signature()` and `calc_structure_bools()` take a `backend`
+argument. The default, `backend="cpu"`, uses NumPy, SciPy, Numba and
+scikit-image. `backend="gpu"` computes the signatures with JAX and tags the
+structures with CuPy. It needs a CUDA GPU and the `gpu` extra:
+
+```bash
+pip install 'pycosmommf[gpu]'
+```
+
+To keep the signatures on the GPU between the two steps, pass
+`return_numpy=False` to `maximum_signature()`. It then returns a JAX array, and
+`calc_structure_bools()` reads that array without copying it. The returned
+`structure_filter` is always a NumPy array.
+
+```python
+import os
+
+# set these before JAX is imported; they are needed for large grids
+os.environ["XLA_PYTHON_CLIENT_PREALLOCATE"] = "false"
+os.environ["XLA_PYTHON_CLIENT_ALLOCATOR"] = "cuda_async"
+
+import PyCosmoMMF
+
+max_signatures = PyCosmoMMF.maximum_signature(
+    Rs, density_field, algorithm="NEXUSPLUS", backend="gpu", return_numpy=False
+)
+structure_filter = PyCosmoMMF.calc_structure_bools(
+    density_field, max_signatures, verbose_flag, backend="gpu"
+)
+```
+
+The GPU backend runs on one GPU. The GPU signatures need about 9–10× the memory
+of one float32 density field, and the GPU tagging needs about 7×. A 1024³ grid,
+the largest grid PyCosmoMMF supports, needs about 65 GiB, so it needs a GPU with
+80 GB of memory. A 40 GB GPU is not enough at 1024³. The CPU and GPU backends
+agree to float32 precision and give the same structure tagging.
+
 #### Additional Code Information
 
 - Note: The NEXUS+ implementation of tagging clusters is highly dependent on the

pycosmommf-2.0.0/docs/changelog.md

@@ -0,0 +1,3 @@
+```{include} ../CHANGELOG.md
+
+```

{pycosmommf-1.0.3 → pycosmommf-2.0.0}/docs/index.md

@@ -4,6 +4,7 @@
 
 basic_usage
 api_tab
+changelog
 
 ```
 
@@ -24,7 +25,8 @@ api_tab
 # PyCosmoMMF
 
 PyCosmoMMF is a package for identifying structures in the Cosmic Web from a
-given overdensity field.
+given overdensity field. Version 2.0 includes a gpu backend for running the code
+with NVIDIA GPUs.
 
 ### Authors
 

{pycosmommf-1.0.3 → pycosmommf-2.0.0}/pyproject.toml

@@ -11,9 +11,9 @@ authors = [
 description = "A package for identifying structures in the cosmic web."
 readme = "README.md"
 license.file = "LICENSE"
-requires-python = ">=3.8, <3.13"
+requires-python = ">=3.9"
 classifiers = [
-  "Development Status :: 1 - Planning",
+  "Development Status :: 5 - Production/Stable",
   "Intended Audience :: Science/Research",
   "Intended Audience :: Developers",
   "License :: OSI Approved :: MIT License",
@@ -21,7 +21,6 @@ classifiers = [
   "Programming Language :: Python",
   "Programming Language :: Python :: 3",
   "Programming Language :: Python :: 3 :: Only",
-  "Programming Language :: Python :: 3.8",
   "Programming Language :: Python :: 3.9",
   "Programming Language :: Python :: 3.10",
   "Programming Language :: Python :: 3.11",
@@ -33,7 +32,8 @@ dynamic = ["version"]
 dependencies = [
 "numba",
 "numpy>=1.22",
-"scikit-image"
+"scikit-image",
+"scipy",
 ]
 
 [project.optional-dependencies]
@@ -52,6 +52,10 @@ docs = [
   "sphinx_autodoc_typehints",
   "furo>=2023.08.17",
 ]
+gpu = [
+  "jax[cuda12]>=0.4.30",
+  "cupy-cuda12x",
+]
 
 [project.urls]
 Homepage = "https://github.com/James11222/pycosmommf"
@@ -75,6 +79,7 @@ addopts = ["-ra", "--showlocals", "--strict-markers", "--strict-config"]
 xfail_strict = true
 filterwarnings = [
   "error",
+  "ignore:jax-metal does not support FFT:UserWarning",
 ]
 log_cli_level = "INFO"
 testpaths = [
@@ -84,9 +89,16 @@ testpaths = [
 
 [tool.coverage]
 run.source = ["pycosmommf"]
+# GitHub Actions has no free GPUs, so the GPU backend is left out of coverage:
+# its two modules and the branches that dispatch to them.
+run.omit = [
+  "*/pycosmommf/_jax_backend.py",
+  "*/pycosmommf/_cupy_tagging.py",
+]
 report.exclude_also = [
   '\.\.\.',
   'if typing.TYPE_CHECKING:',
+  'if backend == "gpu":',
 ]
 
 [tool.ruff]
@@ -113,7 +125,6 @@ extend-select = [
   "YTT",      # flake8-2020
   "EXE",      # flake8-executable
   "NPY",      # NumPy specific rules
-  "PD",       # pandas-vet
 ]
 ignore = [
   "F403",     # star imports