ulpwise 0.3.3__tar.gz → 0.3.5__tar.gz

{ulpwise-0.3.3 → ulpwise-0.3.5}/CHANGELOG.md

@@ -1,5 +1,38 @@
 # Changelog
 
+## 0.3.5 (2026-10-28)
+
+- kornia 0.9.0 (released 2026-10-23) contains every kornia fix in the corpus and torchrl 0.14.1 (2026-10-10) the four
+  pytorch/rl ones, so `fixed_in_release` is set on those 36 cases and they are plain passes instead of expected
+  failures. The weekly strict run had been failing on them as unexpected passes since 2026-10-12, the four torchrl
+  cases first and all 36 on 2026-10-26.
+- Two more kornia cases, both merged upstream on 2026-10-09 and in 0.9.0. #5653: `gftt_response` took the smaller
+  eigenvalue as `0.5 * (trace - sqrt((trace**2 - 4 * det).abs()))`, and on a flat 9 x 9 footprint the blurred moments
+  are exactly zero, so the sqrt backward is `inf * 0 = nan` and the whole flat region gets a nan gradient; the case
+  takes the gradient of the summed response of an image with black top and bottom rows (2048 nan pixels on 0.8.3).
+  #5655: `equalize_clahe` wrote the clipping residual into the first `residual` histogram bins instead of spreading
+  it over every `max(256 // residual, 1)`-th bin as OpenCV does, which lifted the whole lookup table of a small tile
+  (up to +38 levels against `cv2.createCLAHE`); the case maps a 2 x 4 tile holding levels 0 and 100 with clip 40,
+  where level 100 came out as 255 instead of 159.
+
+## 0.3.4 (2026-10-09)
+
+- Regression corpus case for pytorch/rl #4504, merged upstream: `MultiCategorical.to_one_hot` iterated over `self.nvec`,
+  a 2-D tensor for a spec with a batch shape, so `one_hot` got a tensor as `num_classes` and raised `TypeError`, and the
+  0-D sample of a shape `[1]` spec raised `IndexError` on `val[..., i]`. The case one-hot encodes a sample of
+  `MultiCategorical([3, 2], shape=(4, 2))` and the scalar sample of `MultiCategorical([5])`.
+- `ulpwise scan`: new `dropout-never-applied` rule (medium), the first that reads a whole file rather than one
+  function: a `Dropout`, `DropPath` or a `ModuleDict` or `ModuleList` under a dropout name is assigned to `self`
+  and then never called, never passed on, never returned or iterated and its rate never read, anywhere in the file.
+  The option is accepted and does nothing, so the model trains without the regularisation it reports. peft #3830 is
+  the example: `OFTLayer` built `module_dropout` into `self.oft_dropout`, filled it in `update_layer` and never
+  called it. Reading the rate counts as use (`scaled_dot_product_attention(dropout_p=self.dropout.p)`), as does
+  passing the module to another one; filling it does not, and subclasses of `Sequential`, which run every
+  attribute, are not read. On kornia, ultralytics, torchvision, diffusers, torchrl, vllm and detectron2 main the
+  rule reports nothing; on peft main it reports the #3830 line, on timm main one `attn_drop` in `coat.py` that a
+  comment already calls unused, and on transformers main nine lines, four of them modular files whose `forward`
+  lives in another file.
+
 ## 0.3.3 (2026-10-08)
 
 - Two regression corpus cases for peft, both open upstream with a fix PR: peft #3769, `add_weighted_adapter` with

{ulpwise-0.3.3 → ulpwise-0.3.5}/Cargo.lock

@@ -37,9 +37,9 @@ dependencies = [
 
 [[package]]
 name = "pyo3"
-version = "0.29.3"
+version = "0.30.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "700d18fa267b73b9b521fd7e13580e2f446916f176cacee1ab63fcc8191f1655"
+checksum = "229ca13b8a15bb0d1487b81ab975bfcd97513416499a8312011b6036e42e0de4"
 dependencies = [
  "libc",
  "once_cell",
@@ -51,18 +51,18 @@ dependencies = [
 
 [[package]]
 name = "pyo3-build-config"
-version = "0.29.3"
+version = "0.30.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7b3fc0c4d08f6bb10e71fe39dfb9e2f59c6eb6854e22ec8092f50c69a4499adb"
+checksum = "70130e2b6dd1ee589a8788026a429561ca5c40fa7a2fa7914c2282eddae24700"
 dependencies = [
  "target-lexicon",
 ]
 
 [[package]]
 name = "pyo3-ffi"
-version = "0.29.3"
+version = "0.30.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dfc0b8e19df29aad7086cf977bb0c2a2f143e30567eb113e9cf72b62ca698330"
+checksum = "5e7ff73c555c286d212d4ce7e5fbda845a973c4f61bd5f160f5f2c3c2a04d1c4"
 dependencies = [
  "libc",
  "pyo3-build-config",
@@ -70,9 +70,9 @@ dependencies = [
 
 [[package]]
 name = "pyo3-macros"
-version = "0.29.3"
+version = "0.30.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6100e8a4b5eba53afaa5ed078364851a0b2499c553a44c31026b929049b49dc6"
+checksum = "c46d2e5862c7bb8207e3b639f713dda3c499933444784154c11d013c5a691b60"
 dependencies = [
  "proc-macro2",
  "pyo3-macros-backend",
@@ -82,9 +82,9 @@ dependencies = [
 
 [[package]]
 name = "pyo3-macros-backend"
-version = "0.29.3"
+version = "0.30.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6143877a16e82b5a727b7127ff4cd86858a24a28d745d72f43e6f227c7b1bdb3"
+checksum = "38b89f066f2ae72a2aef52d1537ca2fcd104440ebfadee06e97aa90738a25b9b"
 dependencies = [
  "heck",
  "proc-macro2",
@@ -103,9 +103,9 @@ dependencies = [
 
 [[package]]
 name = "syn"
-version = "2.0.119"
+version = "3.0.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "872831b642d1a07999a962a351ed35b955ea2cfc8f3862091e2a240a84f17297"
+checksum = "d78c8dee4c7bf0e14673097256fed6142ce9d3b85a408189d07482442145823b"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -120,7 +120,7 @@ checksum = "adb6935a6f5c20170eeceb1a3835a49e12e19d792f6dd344ccc76a985ca5a6ca"
 
 [[package]]
 name = "ulpwise"
-version = "0.3.3"
+version = "0.3.5"
 dependencies = [
  "pyo3",
 ]

{ulpwise-0.3.3 → ulpwise-0.3.5}/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "ulpwise"
-version = "0.3.3"
+version = "0.3.5"
 edition = "2021"
 rust-version = "1.86"
 license = "MIT OR Apache-2.0"
@@ -15,7 +15,7 @@ name = "ulpwise"
 crate-type = ["cdylib", "rlib"]
 
 [dependencies]
-pyo3 = { version = "0.29", features = ["extension-module", "abi3-py39"], optional = true }
+pyo3 = { version = "0.30", features = ["extension-module", "abi3-py39"], optional = true }
 
 [features]
 default = []

{ulpwise-0.3.3 → ulpwise-0.3.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ulpwise
-Version: 0.3.3
+Version: 0.3.5
 Classifier: Development Status :: 3 - Alpha
 Classifier: Framework :: Pytest
 Classifier: Intended Audience :: Developers
@@ -173,7 +173,7 @@ environment, side-install it and put it first on the path:
 
 ```sh
 pip install --no-deps --target /tmp/kornia-0.8.3 kornia==0.8.3
-PYTHONPATH=/tmp/kornia-0.8.3 ulpwise corpus --repo kornia     # 31 present, 0 fixed on kornia 0.8.3
+PYTHONPATH=/tmp/kornia-0.8.3 ulpwise corpus --repo kornia     # 34 present, 0 fixed on kornia 0.8.3; 0 present, 34 fixed on 0.9.0
 ```
 
 | case | kind | merged |
@@ -185,6 +185,7 @@ PYTHONPATH=/tmp/kornia-0.8.3 ulpwise corpus --repo kornia     # 31 present, 0 fi
 | pytorch/rl #4443 `arange(0, 1, 1/n)` gives `n + 1` positions for 140 values of `n` below 2000 | rounding | 2026-09-20 |
 | pytorch/rl #4444 `min_value or -inf` drops `min_value=0` | falsy zero | 2026-09-20 |
 | pytorch/rl #4445 scheduler `state_dict()` contained a module object | crash | 2026-09-20 |
+| pytorch/rl #4504 `MultiCategorical.to_one_hot` crashed for a batched spec and a 0-D sample | crash | 2026-10-07 |
 | timm #2786 Mars kept a reference to `p.grad` as the previous gradient | aliasing | 2026-09-17 |
 | timm #2790 AdafactorBigVision clipped updates in the wrong direction | direction | 2026-09-18 |
 | timm #2791 AdaMuon conv LR scale computed from the wrong dims | scale | 2026-09-18 |
@@ -229,13 +230,15 @@ PYTHONPATH=/tmp/kornia-0.8.3 ulpwise corpus --repo kornia     # 31 present, 0 fi
 | pytorch #199850 `torch.erf` in bfloat16 and float16 on CPU returns 0 at and below 1.8e-7 and loses relative accuracy below 1e-3 (13404 bfloat16 ulps, 5 float16 ulps) | cancellation | open |
 | pytorch #199867 `torch.special.logit` in float16 and bfloat16 on CPU rounds `1 - x` and `x / (1 - x)` to the input dtype before the log, logit(0.499756) is -0.000488 for -0.000977 (512 float16 ulps, 64 bfloat16 ulps) | rounding | open |
 | kornia #5505 `angle_error_mat` and `angle_error_vec` returned exactly 0 for every rotation below 0.03 degrees in float32 (`acos` next to 1, issue #5500) | cancellation | 2026-10-06 |
+| kornia #5653 `gftt_response` had a nan gradient on every flat patch, `sqrt` of a zero discriminant (issue #5652) | gradient | 2026-10-09 |
+| kornia #5655 `equalize_clahe` wrote the clipping residual into the first histogram bins, small tiles tens of levels brighter than OpenCV (issue #5654) | redistribution | 2026-10-09 |
 | peft #3769 `add_weighted_adapter` with `combination_type='svd'` crashed for Conv1d and Conv3d LoRA layers | crash | open |
 | peft #3830 OFT `module_dropout` was never applied, every training forward used all rotation blocks | no-op | open |
 
 Cases marked `open` have an issue with the complete patch attached and no merged fix yet. Every case
 is an expected failure until the installed release contains the fix (`fixed_in_release` in
-`cases.json`; the timm and ultralytics fixes above are released, the kornia, pytorch/rl, peft and
-pytorch ones are merged but not in a release yet), and the `max_ulp` check type measures the digits
+`cases.json`; the timm, ultralytics, kornia (0.9.0) and pytorch/rl (0.14.1) fixes above are released,
+the peft and pytorch ones are merged but not in a release yet), and the `max_ulp` check type measures the digits
 directly. The ultralytics dataset and label cases build their one-image
 dataset in a temporary directory and need no weights; two more ultralytics fixes (#26240, #26246) are not in
 the corpus yet because their repros need model weights or the COCO evaluator.
@@ -249,9 +252,9 @@ ulpwise scan path/to/repo --rules one-minus-cos,small-angle-division --top 40
 ```
 
 The scan is static and needs nothing installed: it parses each file with `ast`, walks every
-function and matches fifteen patterns, each with a severity, the reason it loses digits,
-overflows or loses its gradient, the usual replacement and, where one exists, the upstream bug it
-comes from.
+function and matches sixteen patterns, each with a severity, the reason it loses digits,
+overflows or loses its gradient, or in one case does nothing at all, the usual replacement and,
+where one exists, the upstream bug it comes from.
 
 | rule | severity | pattern |
 |---|---|---|
@@ -269,6 +272,7 @@ comes from.
 | `eps-floor` | medium | `x * (1 - eps) + eps`: every value moves, 0 becomes `eps`, so a one-hot's entries sum to `1 + (C - 1) eps` and a perfect prediction scores a loss that grows with the image (kornia #5538, found by the kornia conventions audit) |
 | `where-nan-gradient` | medium | `where(d > eps, f(d), other)` with `f` a division by `d` or a `sqrt`, `log`, `acos` or `asin` of it: `where` evaluates both branches and hands the discarded one a zero gradient, and the backward of `f` at the singularity turns that zero into `0 / 0 = nan`, so the guard protects the value and not the gradient; quiet for `numpy.where`, for a comparison against a number above 1, for a floor named `eps`, `tol` or `floor` on the other side, and once `d` is re-bound to a `where`, `clamp` or `maximum` of itself (kornia #5579, found by the kornia conventions audit) |
 | `clamp-at-singularity` | medium | `clamp(x, min=0).sqrt()`, `sqrt(clamp(x, min=0))`, `clamp(c, -1, 1).acos()`: the bound is the point where the next function has an infinite derivative, and clamp's derivative at its own bound is not the same across torch versions, 1 on 2.5.1 and 2.9.1 and 0 on 2.14, so the gradient there is `inf` or `nan` on the older half of a supported range; a bound strictly inside the domain, `min=1e-8`, is a floor and is not reported, and `numpy` is quiet (kornia #4229, found by the kornia conventions audit; kornia #5500) |
+| `dropout-never-applied` | medium | a `Dropout`, `DropPath` or a `ModuleDict` of them assigned to `self` and then never called, never passed on and its rate never read, anywhere in the file: the option is accepted and does nothing, so the model trains without the regularisation it reports (peft #3830) |
 
 On kornia `main` at `e05b0ee` the scan takes 4 s for 506 files and reports 35 findings. The
 `one-minus-cos` and `small-angle-division` findings are the four lines of `So3.right_jacobian` and
@@ -284,6 +288,19 @@ report seven lines: three `clamp` bounds in `_solve_cubic_real`, and four `where
 boxes and polygons. On ultralytics `main` at `8df3534` they report three lines, all in metrics or
 inference code that is never differentiated, and nothing in torchvision.
 
+`dropout-never-applied` reads the whole file rather than one function, since the module is built
+in `__init__` and used, if at all, in `forward` or in a sibling class. Calling it, calling it
+through a subscript, passing it to another module, returning it, iterating over it and reading
+its rate (`scaled_dot_product_attention(dropout_p=self.dropout.p)`) all count as use; filling it
+(`self.oft_dropout.update(...)`) does not, and a subclass of `Sequential` runs every attribute and
+is not read. On kornia `main` at `6d579a72`, ultralytics `6d51b7e`, torchvision `9a8d545`,
+diffusers `d961a38`, torchrl `648f50c`, vllm `73c742b` and detectron2 `1e3e13b` it reports
+nothing. On peft `main` at `f6d8480` it reports `OFTLayer.oft_dropout`, which is #3830; on timm
+`83e6eb5` it reports `FactorAttnConvRelPosEnc.attn_drop` in `coat.py`, which a comment on that
+line already calls unused; on transformers `9167f73` it reports nine lines, four of them in
+modular files whose `forward` is inherited from another file, which the rule cannot see, and
+five attention, embedding and head dropouts built in `__init__` and used nowhere in their file.
+
 `--run` adds the dynamic half. The module level functions with the most elementary math are
 imported and called with the same 91 point grid (both signs of `1e-8` to `1e3`, and zero) for every
 required argument, in float32 and in float64, torch first and numpy second, and the float32 result

{ulpwise-0.3.3 → ulpwise-0.3.5}/README.md

@@ -144,7 +144,7 @@ environment, side-install it and put it first on the path:
 
 ```sh
 pip install --no-deps --target /tmp/kornia-0.8.3 kornia==0.8.3
-PYTHONPATH=/tmp/kornia-0.8.3 ulpwise corpus --repo kornia     # 31 present, 0 fixed on kornia 0.8.3
+PYTHONPATH=/tmp/kornia-0.8.3 ulpwise corpus --repo kornia     # 34 present, 0 fixed on kornia 0.8.3; 0 present, 34 fixed on 0.9.0
 ```
 
 | case | kind | merged |
@@ -156,6 +156,7 @@ PYTHONPATH=/tmp/kornia-0.8.3 ulpwise corpus --repo kornia     # 31 present, 0 fi
 | pytorch/rl #4443 `arange(0, 1, 1/n)` gives `n + 1` positions for 140 values of `n` below 2000 | rounding | 2026-09-20 |
 | pytorch/rl #4444 `min_value or -inf` drops `min_value=0` | falsy zero | 2026-09-20 |
 | pytorch/rl #4445 scheduler `state_dict()` contained a module object | crash | 2026-09-20 |
+| pytorch/rl #4504 `MultiCategorical.to_one_hot` crashed for a batched spec and a 0-D sample | crash | 2026-10-07 |
 | timm #2786 Mars kept a reference to `p.grad` as the previous gradient | aliasing | 2026-09-17 |
 | timm #2790 AdafactorBigVision clipped updates in the wrong direction | direction | 2026-09-18 |
 | timm #2791 AdaMuon conv LR scale computed from the wrong dims | scale | 2026-09-18 |
@@ -200,13 +201,15 @@ PYTHONPATH=/tmp/kornia-0.8.3 ulpwise corpus --repo kornia     # 31 present, 0 fi
 | pytorch #199850 `torch.erf` in bfloat16 and float16 on CPU returns 0 at and below 1.8e-7 and loses relative accuracy below 1e-3 (13404 bfloat16 ulps, 5 float16 ulps) | cancellation | open |
 | pytorch #199867 `torch.special.logit` in float16 and bfloat16 on CPU rounds `1 - x` and `x / (1 - x)` to the input dtype before the log, logit(0.499756) is -0.000488 for -0.000977 (512 float16 ulps, 64 bfloat16 ulps) | rounding | open |
 | kornia #5505 `angle_error_mat` and `angle_error_vec` returned exactly 0 for every rotation below 0.03 degrees in float32 (`acos` next to 1, issue #5500) | cancellation | 2026-10-06 |
+| kornia #5653 `gftt_response` had a nan gradient on every flat patch, `sqrt` of a zero discriminant (issue #5652) | gradient | 2026-10-09 |
+| kornia #5655 `equalize_clahe` wrote the clipping residual into the first histogram bins, small tiles tens of levels brighter than OpenCV (issue #5654) | redistribution | 2026-10-09 |
 | peft #3769 `add_weighted_adapter` with `combination_type='svd'` crashed for Conv1d and Conv3d LoRA layers | crash | open |
 | peft #3830 OFT `module_dropout` was never applied, every training forward used all rotation blocks | no-op | open |
 
 Cases marked `open` have an issue with the complete patch attached and no merged fix yet. Every case
 is an expected failure until the installed release contains the fix (`fixed_in_release` in
-`cases.json`; the timm and ultralytics fixes above are released, the kornia, pytorch/rl, peft and
-pytorch ones are merged but not in a release yet), and the `max_ulp` check type measures the digits
+`cases.json`; the timm, ultralytics, kornia (0.9.0) and pytorch/rl (0.14.1) fixes above are released,
+the peft and pytorch ones are merged but not in a release yet), and the `max_ulp` check type measures the digits
 directly. The ultralytics dataset and label cases build their one-image
 dataset in a temporary directory and need no weights; two more ultralytics fixes (#26240, #26246) are not in
 the corpus yet because their repros need model weights or the COCO evaluator.
@@ -220,9 +223,9 @@ ulpwise scan path/to/repo --rules one-minus-cos,small-angle-division --top 40
 ```
 
 The scan is static and needs nothing installed: it parses each file with `ast`, walks every
-function and matches fifteen patterns, each with a severity, the reason it loses digits,
-overflows or loses its gradient, the usual replacement and, where one exists, the upstream bug it
-comes from.
+function and matches sixteen patterns, each with a severity, the reason it loses digits,
+overflows or loses its gradient, or in one case does nothing at all, the usual replacement and,
+where one exists, the upstream bug it comes from.
 
 | rule | severity | pattern |
 |---|---|---|
@@ -240,6 +243,7 @@ comes from.
 | `eps-floor` | medium | `x * (1 - eps) + eps`: every value moves, 0 becomes `eps`, so a one-hot's entries sum to `1 + (C - 1) eps` and a perfect prediction scores a loss that grows with the image (kornia #5538, found by the kornia conventions audit) |
 | `where-nan-gradient` | medium | `where(d > eps, f(d), other)` with `f` a division by `d` or a `sqrt`, `log`, `acos` or `asin` of it: `where` evaluates both branches and hands the discarded one a zero gradient, and the backward of `f` at the singularity turns that zero into `0 / 0 = nan`, so the guard protects the value and not the gradient; quiet for `numpy.where`, for a comparison against a number above 1, for a floor named `eps`, `tol` or `floor` on the other side, and once `d` is re-bound to a `where`, `clamp` or `maximum` of itself (kornia #5579, found by the kornia conventions audit) |
 | `clamp-at-singularity` | medium | `clamp(x, min=0).sqrt()`, `sqrt(clamp(x, min=0))`, `clamp(c, -1, 1).acos()`: the bound is the point where the next function has an infinite derivative, and clamp's derivative at its own bound is not the same across torch versions, 1 on 2.5.1 and 2.9.1 and 0 on 2.14, so the gradient there is `inf` or `nan` on the older half of a supported range; a bound strictly inside the domain, `min=1e-8`, is a floor and is not reported, and `numpy` is quiet (kornia #4229, found by the kornia conventions audit; kornia #5500) |
+| `dropout-never-applied` | medium | a `Dropout`, `DropPath` or a `ModuleDict` of them assigned to `self` and then never called, never passed on and its rate never read, anywhere in the file: the option is accepted and does nothing, so the model trains without the regularisation it reports (peft #3830) |
 
 On kornia `main` at `e05b0ee` the scan takes 4 s for 506 files and reports 35 findings. The
 `one-minus-cos` and `small-angle-division` findings are the four lines of `So3.right_jacobian` and
@@ -255,6 +259,19 @@ report seven lines: three `clamp` bounds in `_solve_cubic_real`, and four `where
 boxes and polygons. On ultralytics `main` at `8df3534` they report three lines, all in metrics or
 inference code that is never differentiated, and nothing in torchvision.
 
+`dropout-never-applied` reads the whole file rather than one function, since the module is built
+in `__init__` and used, if at all, in `forward` or in a sibling class. Calling it, calling it
+through a subscript, passing it to another module, returning it, iterating over it and reading
+its rate (`scaled_dot_product_attention(dropout_p=self.dropout.p)`) all count as use; filling it
+(`self.oft_dropout.update(...)`) does not, and a subclass of `Sequential` runs every attribute and
+is not read. On kornia `main` at `6d579a72`, ultralytics `6d51b7e`, torchvision `9a8d545`,
+diffusers `d961a38`, torchrl `648f50c`, vllm `73c742b` and detectron2 `1e3e13b` it reports
+nothing. On peft `main` at `f6d8480` it reports `OFTLayer.oft_dropout`, which is #3830; on timm
+`83e6eb5` it reports `FactorAttnConvRelPosEnc.attn_drop` in `coat.py`, which a comment on that
+line already calls unused; on transformers `9167f73` it reports nine lines, four of them in
+modular files whose `forward` is inherited from another file, which the rule cannot see, and
+five attention, embedding and head dropouts built in `__init__` and used nowhere in their file.
+
 `--run` adds the dynamic half. The module level functions with the most elementary math are
 imported and called with the same 91 point grid (both signs of `1e-8` to `1e3`, and zero) for every
 required argument, in float32 and in float64, torch first and numpy second, and the float32 result

{ulpwise-0.3.3 → ulpwise-0.3.5}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "maturin"
 
 [project]
 name = "ulpwise"
-version = "0.3.3"
+version = "0.3.5"
 description = "Numerical conformance testing for ML code: float edge values, exact rounding oracles, knife-edge finders and a pytest plugin"
 readme = "README.md"
 license = { text = "MIT OR Apache-2.0" }

{ulpwise-0.3.3 → ulpwise-0.3.5}/python/ulpwise/__init__.py

@@ -30,7 +30,7 @@ from ._core import special as _special_core
 from ._core import ulp_distance as _ulp_distance_core
 from ._core import ulp_distances as _ulp_distances_core
 
-__version__ = "0.3.3"
+__version__ = "0.3.5"
 
 __all__ = [
     "UNARY_OPS",

{ulpwise-0.3.3 → ulpwise-0.3.5}/python/ulpwise/corpus/cases.json

@@ -14,7 +14,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nimport kornia.filters as KF\nxs = torch.arange(-4, 5, dtype=torch.float64)\ny, x = torch.meshgrid(xs, xs, indexing='ij')\ndxx = KF.spatial_gradient((x * x)[None, None], mode='sobel', order=2)[0, 0, 0, 4, 4]\ndxy = KF.spatial_gradient((x * y)[None, None], mode='sobel', order=2)[0, 0, 1, 4, 4]\ndxy_diff = KF.spatial_gradient((x * y)[None, None], mode='diff', order=2)[0, 0, 1, 4, 4]\nactual = [dxx.item(), dxy.item(), dxy_diff.item()]\nexpected = [2.0, 1.0, 1.0]\n",
    "check": {
     "type": "sign"
@@ -33,7 +33,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nimport kornia.feature as KFe\nxs = torch.arange(-4, 5, dtype=torch.float64)\ny, x = torch.meshgrid(xs, xs, indexing='ij')\ndet_sobel_xy = KFe.hessian_response((x * y)[None, None], grads_mode='sobel')[0, 0, 4, 4]\ndet_diff_quad = KFe.hessian_response((x * x + y * y)[None, None], grads_mode='diff')[0, 0, 4, 4]\ndet_sobel_degenerate = KFe.hessian_response(((x + y) ** 2 / 2)[None, None], grads_mode='sobel')[0, 0, 4, 4]\nactual = [det_sobel_xy.item(), det_diff_quad.item(), det_sobel_degenerate.item()]\nexpected = [-1.0, 4.0, 0.0]\n",
    "check": {
     "type": "allclose",
@@ -54,7 +54,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nimport kornia.feature as KF\n# Oxford ellipses [x, y, a, b, c]; the LAF's A must satisfy A @ A.T == inverse([[a, b], [b, c]])\nells = torch.tensor([[[0.0, 0.0, 1.0, 0.5, 1.0], [3.0, -2.0, 2.0, -0.9, 0.5]]], dtype=torch.float64)\nA = KF.ellipse_to_laf(ells)[..., :2]\nregion = torch.linalg.inv(A @ A.transpose(-1, -2))[0]\nactual = region.reshape(-1).tolist()\nexpected = [1.0, 0.5, 0.5, 1.0, 2.0, -0.9, -0.9, 0.5]\n",
    "check": {
     "type": "allclose",
@@ -96,7 +96,7 @@
     "torch",
     "torchrl"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.14.1",
    "repro": "import torch\nfrom torchrl.envs.transforms import ActionDiscretizer\nt = ActionDiscretizer(num_intervals=49)\nt.dtype = torch.float32\nactual = [t.custom_arange(n, torch.device('cpu')).numel() for n in (49, 98, 103, 107, 10)]\nexpected = [49, 98, 103, 107, 10]\n",
    "check": {
     "type": "equal"
@@ -115,7 +115,7 @@
     "torch",
     "torchrl"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.14.1",
    "repro": "from torchrl.data.replay_buffers import ListStorage, PrioritizedSampler, ReplayBuffer\nfrom torchrl.data.replay_buffers.scheduler import StepScheduler\nrb = ReplayBuffer(storage=ListStorage(16), sampler=PrioritizedSampler(16, alpha=0.5, beta=0.5))\nsched = StepScheduler(rb, 'alpha', gamma=-0.2, n_steps=1, mode='additive', min_value=0)\nseen = []\nfor _ in range(4):\n    sched.step()\n    seen.append(float(rb.sampler.alpha))\nactual = seen\nexpected = [0.3, 0.1, 0.0, 0.0]\n",
    "check": {
     "type": "allclose",
@@ -136,13 +136,32 @@
     "torch",
     "torchrl"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.14.1",
    "repro": "import io\nimport torch\nfrom torchrl.data.replay_buffers import ListStorage, PrioritizedSampler, ReplayBuffer\nfrom torchrl.data.replay_buffers.scheduler import StepScheduler\nrb = ReplayBuffer(storage=ListStorage(16), sampler=PrioritizedSampler(16, alpha=0.5, beta=0.5))\nsched = StepScheduler(rb, 'alpha', gamma=0.5, n_steps=1)\nsched.step()\nsched.step()\nbuf = io.BytesIO()\ntorch.save(sched.state_dict(), buf)\nbuf.seek(0)\nstate = torch.load(buf, weights_only=False)\nactual = [state['_step_cnt']]\nexpected = [2]\n",
    "check": {
     "type": "equal"
    },
    "before_fix": "TypeError: cannot pickle 'module' object"
   },
+  {
+   "id": "torchrl-4504-multicategorical-to-one-hot-batched-spec",
+   "repo": "pytorch/rl",
+   "pr": 4504,
+   "merged_at": "2026-10-07T10:30:35Z",
+   "kind": "crash",
+   "title": "MultiCategorical.to_one_hot crashed for a spec with a batch shape and for the scalar sample of a shape [1] spec",
+   "symptom": "to_one_hot iterated over self.nvec, which is a 2-D tensor for a spec with a batch shape, so one_hot received a tensor as num_classes and raised TypeError; a shape [1] spec returns 0-D samples and val[..., i] on them raised IndexError.",
+   "requires": [
+    "torch",
+    "torchrl"
+   ],
+   "fixed_in_release": "0.14.1",
+   "repro": "import torch\nfrom torchrl.data import MultiCategorical\nspec = MultiCategorical([3, 2], shape=(4, 2))\nhot = spec.to_one_hot(spec.rand())\nscalar = MultiCategorical([5]).to_one_hot(torch.tensor(3))\nactual = [list(hot.shape), hot.dtype == torch.bool, int(hot.sum()), list(scalar.shape), int(scalar.sum()), bool(scalar[3])]\nexpected = [[4, 5], True, 8, [5], 1, True]\n",
+   "check": {
+    "type": "equal"
+   },
+   "before_fix": "TypeError: one_hot(): argument 'num_classes' (position 2) must be int, not Tensor for the batched spec; IndexError: too many indices for tensor of dimension 0 for the scalar sample"
+  },
   {
    "id": "timm-2786-mars-last-grad-aliasing",
    "repo": "huggingface/pytorch-image-models",
@@ -299,7 +318,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nimport kornia.geometry.conversions as C\naxis = torch.tensor([1.0, 2.0, 3.0], dtype=torch.float64) / 14.0**0.5\nv = (9.99e-4 * axis)[None]\nactual = C.axis_angle_to_rotation_matrix(v).flatten().tolist()\nexpected = C.quaternion_to_rotation_matrix(C.axis_angle_to_quaternion(v)).flatten().tolist()\n",
    "check": {
     "type": "allclose",
@@ -321,7 +340,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nfrom kornia.geometry.liegroup import Se3, So3\nrot = So3.exp(torch.tensor([[1e-4, 0.0, 0.0]], dtype=torch.float32))\nse = Se3.exp(torch.tensor([[1.0, 0.0, 0.0, 0.0, 0.0, 1e-4]], dtype=torch.float32))\nactual = [rot.log()[0, 0].item(), se.t[0, 1].item()]\nexpected = [1e-4, 5e-5]\n",
    "check": {
     "type": "allclose",
@@ -445,7 +464,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nfrom kornia.geometry.liegroup import Se2\nv = torch.tensor([[1.0, 2.0, 1e-8]], dtype=torch.float64)  # translation (1, 2) and a turn of 1e-8 rad\ng = Se2.exp(v)\nactual = g.matrix()[0, :2, 2].tolist() + g.log().flatten().tolist()\nb = 0.5e-8  # (1 - cos(theta)) / theta = theta / 2 - theta^3 / 24; sin(theta) / theta = 1 - theta^2 / 6 rounds to 1\nexpected = [1.0 - 2.0 * b, 2.0 + b, 1.0, 2.0, 1e-8]\n",
    "check": {
     "type": "allclose",
@@ -466,7 +485,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nfrom kornia.geometry.linalg import point_line_distance\nline = torch.tensor([[0.0, 1.0, -1.0]], dtype=torch.float64)  # the line y = 1\npoint = torch.tensor([[[3.0, 5.0, 2.0]]], dtype=torch.float64)  # the homogeneous point (1.5, 2.5): w = 2\nactual = point_line_distance(point, line).flatten().tolist()\nexpected = [1.5]\n",
    "check": {
     "type": "allclose",
@@ -487,7 +506,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nfrom kornia.geometry.quaternion import Quaternion\nq = Quaternion(torch.tensor([[-1.0, 0.0, 0.0, 0.0]], dtype=torch.float64))  # a half turn: angle pi, axis undefined\nroot = q ** 0.5\nactual = root.data.flatten().tolist() + (root * root).data.flatten().tolist()\nexpected = [0.0, 1.0, 0.0, 0.0, -1.0, 0.0, 0.0, 0.0]\n",
    "check": {
     "type": "allclose",
@@ -508,7 +527,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nfrom kornia.geometry.liegroup import So2\nangle = torch.tensor([[0.0], [1.5707963267948966]], dtype=torch.float64)  # (B, 1): one angle per sample, as a column\npoints = torch.tensor([[1.0, 0.0], [1.0, 0.0]], dtype=torch.float64)\nrotated = So2.exp(angle) * points\nactual = [list(rotated.shape), rotated.flatten().tolist()]\nexpected = [[2, 2], [1.0, 0.0, 0.0, 1.0]]\n",
    "check": {
     "type": "allclose",
@@ -529,7 +548,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nfrom kornia.geometry.solvers import solve_cubic\ncoeffs = torch.tensor([[1e-30, 0.0, 2.0, -6.0]], dtype=torch.float64)  # 1e-30 x^3 + 2 x - 6: the real root is 3\nactual = solve_cubic(coeffs).flatten().tolist()\nexpected = [3.0, 0.0, 0.0]\n",
    "check": {
     "type": "allclose",
@@ -550,7 +569,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nfrom kornia.color import RgbToGrayscale, rgb_to_grayscale\nimage = torch.full((1, 3, 2, 2), 200, dtype=torch.uint8)\nactual = [int(RgbToGrayscale()(image).max()), int(RgbToGrayscale()(image).min())]\nexpected = [int(rgb_to_grayscale(image).max()), int(rgb_to_grayscale(image).min())]\n",
    "check": {
     "type": "equal"
@@ -569,7 +588,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nfrom kornia.geometry.subpix import conv_soft_argmax2d\nquiet = torch.zeros(1, 1, 7, 7)\nquiet[0, 0, 6, 6] = 1.0  # a weak bump in the bottom right corner\nloud = quiet.clone()\nloud[0, 0, 0, 0] = 200.0  # a peak far away, 200 temperatures above the bump\nkwargs = dict(kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), temperature=torch.tensor(1.0), normalized_coordinates=True)\nactual = conv_soft_argmax2d(loud, **kwargs)[0, 0, :, 6, 6].tolist()\nexpected = conv_soft_argmax2d(quiet, **kwargs)[0, 0, :, 6, 6].tolist()\n",
    "check": {
     "type": "allclose",
@@ -590,7 +609,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nfrom kornia.filters import get_gaussian_discrete_kernel1d\none = get_gaussian_discrete_kernel1d(1, 1.0)\nfive = get_gaussian_discrete_kernel1d(5, 1.0)\nactual = [one.numel(), five.numel(), one.flatten().tolist()]\nexpected = [1, 5, [1.0]]\n",
    "check": {
     "type": "equal"
@@ -609,7 +628,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nfrom kornia.geometry.liegroup import Se3, So3\nv = torch.tensor([[1.0, 2.0, 3.0, 0.0, 0.0, 0.0]], dtype=torch.float64)  # upsilon = (1, 2, 3), omega = 0\njac = torch.autograd.functional.jacobian(lambda x: Se3.exp(x).matrix()[:, :3, 3], v)[0, :, 0, :]\nactual = jac[:, 3:].flatten().tolist()\nexpected = (-0.5 * So3.hat(v[0, :3])).flatten().tolist()\n",
    "check": {
     "type": "allclose",
@@ -630,7 +649,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nfrom kornia.geometry.quaternion import Quaternion\ndata = torch.tensor([[1.0, 0.0, 0.0, 0.0]], dtype=torch.float64, requires_grad=True)\nq = Quaternion(data)\nq.polar_angle.sum().backward()\ngrad = data.grad if data.grad is not None else q.data.grad\nactual = grad.flatten().tolist()\nexpected = [0.0, 0.0, 0.0, 0.0]\n",
    "check": {
     "type": "equal"
@@ -649,7 +668,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nfrom kornia.geometry.epipolar import decompose_essential_matrix\nE = torch.tensor([[0.0, -1.0, 0.0], [1.0, 0.0, 0.0], [0.0, 0.0, 0.0]])  # [t]_x with t = e_z and R = I\nR1, R2, t = decompose_essential_matrix(E)\nactual = [list(R1.shape), list(R2.shape), list(t.shape)]\nexpected = [[3, 3], [3, 3], [3, 1]]\n",
    "check": {
     "type": "equal"
@@ -668,7 +687,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nfrom kornia.geometry.plane import Hyperplane\npoints = [torch.tensor([p], dtype=torch.float64, requires_grad=True) for p in ([0.0, 0.0, 1.0], [1.0, 0.0, 1.0], [0.0, 1.0, 1.0])]\nplane = Hyperplane.through(*points)\nplane.normal.data.sum().backward()\nactual = [bool(torch.isfinite(p.grad).all()) for p in points] + plane.normal.data.flatten().abs().tolist()\nexpected = [True, True, True, 0.0, 0.0, 1.0]\n",
    "check": {
     "type": "equal"
@@ -687,7 +706,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nfrom kornia.geometry.vector import Vector3\nactual = Vector3(torch.zeros(1, 3, dtype=torch.float16)).normalized().data.flatten().tolist()\nexpected = [0.0, 0.0, 0.0]\n",
    "check": {
     "type": "equal"
@@ -706,7 +725,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nfrom kornia.filters import otsu_threshold\nx = torch.tensor([[[[-4.0, -3.0], [-2.0, -1.0]]]])\nmask, threshold = otsu_threshold(x, return_mask=True)\nactual = mask.flatten().tolist()\nexpected = [False, False, True, True]\n",
    "check": {
     "type": "equal"
@@ -725,7 +744,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nfrom kornia.losses import MS_SSIMLoss\ntorch.manual_seed(0)\nimg1 = torch.rand(1, 3, 16, 20)\nimg2 = torch.rand(1, 3, 16, 20)\nactual = list(MS_SSIMLoss(sigmas=(0.5, 1.3), reduction=\"none\")(img1, img2).shape)\nexpected = [1, 16, 20]\n",
    "check": {
     "type": "equal"
@@ -744,7 +763,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nfrom kornia.geometry.epipolar import sampson_epipolar_distance\nF = torch.tensor([[[0.0, 0.0, 0.0], [0.0, 0.0, -1.0], [0.0, 1.0, 0.0]]])  # a pure horizontal shift: the epiline of (x, y) is y' = y\npts1 = torch.zeros(1, 1, 2)\npts2 = torch.tensor([[[2.0, 0.0]]])  # on its epiline\nactual = sampson_epipolar_distance(pts1, pts2, F, squared=False).item()\nexpected = 0.0\n",
    "check": {
     "type": "equal"
@@ -763,7 +782,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nfrom kornia.geometry.epipolar import sampson_epipolar_distance\nF = torch.tensor([[[0.0, 0.0, 0.0], [0.0, 0.0, -1.0], [0.0, 1.0, 0.0]]])  # a pure horizontal shift: the epiline of (x, y) is y' = y\npts1 = torch.zeros(1, 1, 2)\npts2 = torch.tensor([[[2.0, 1.0]]])  # one pixel off its epiline\nactual = (sampson_epipolar_distance(pts1, pts2, 1e-4 * F) / sampson_epipolar_distance(pts1, pts2, F)).item()\nexpected = 1.0\n",
    "check": {
     "type": "allclose",
@@ -784,7 +803,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nimport kornia\nfrom kornia.augmentation import RandomHue\ntorch.manual_seed(0)\nimage = torch.rand(1, 3, 4, 4, dtype=torch.float64)\nhue = kornia.color.rgb_to_hsv(image)[:, 0]\nshifted = kornia.color.rgb_to_hsv(RandomHue(hue=(0.5, 0.5), p=1.0)(image))[:, 0]\nactual = (torch.remainder(shifted - hue, 2 * torch.pi) - torch.pi).abs().max().item()  # circular distance from a half turn\nexpected = 0.0\n",
    "check": {
     "type": "allclose",
@@ -805,7 +824,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nfrom kornia.geometry.liegroup import So3\nzero = torch.zeros(3, dtype=torch.float64)\nactual = torch.autograd.functional.hessian(lambda v: So3.exp(v).q.data[0], zero).flatten().tolist()\nexpected = (-0.25 * torch.eye(3, dtype=torch.float64)).flatten().tolist()\n",
    "check": {
     "type": "allclose",
@@ -826,7 +845,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nfrom kornia.losses import MS_SSIMLoss\ntorch.manual_seed(0)\nimg1 = torch.randint(0, 256, (1, 1, 48, 40), dtype=torch.uint8)\nimg2 = torch.randint(0, 256, (1, 1, 48, 40), dtype=torch.uint8)\nactual = MS_SSIMLoss(data_range=255.0)(img1, img2).item()\nexpected = MS_SSIMLoss()(img1.float() / 255, img2.float() / 255).item()\n",
    "check": {
     "type": "allclose",
@@ -847,7 +866,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nfrom kornia.geometry.liegroup import So3\nomega = torch.tensor([[0.48, 0.6, 0.64]]) * 45.0  # 45 rad about a unit axis\nactual = So3.right_jacobian(omega.half()).flatten().tolist()\nexpected = So3.right_jacobian(omega.double()).flatten().tolist()\n",
    "check": {
     "type": "allclose",
@@ -868,7 +887,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import math\nimport torch\nfrom kornia.geometry.quaternion import Quaternion, average_quaternions\nq1 = torch.tensor([1.0, 0.0, 0.0, 0.0])\nq2 = torch.tensor([1.0, 1.0, 0.0, 0.0]) / math.sqrt(2.0)  # a quarter turn about x\nmean = average_quaternions(Quaternion(torch.stack([q1, 3.0 * q2]))).data.flatten()\nactual = (mean * torch.sign(mean[0])).tolist()  # the sign of a quaternion is free\nexpected = [math.cos(math.pi / 8), math.sin(math.pi / 8), 0.0, 0.0]  # the bisector, an eighth of a turn\n",
    "check": {
     "type": "allclose",
@@ -889,7 +908,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nfrom kornia.geometry.plane import Hyperplane\nfrom kornia.geometry.vector import Vector3\np0 = Vector3(torch.tensor([0.0, 0.0, 0.0], dtype=torch.float16))\np1 = Vector3(torch.tensor([300.0, 0.0, 0.0], dtype=torch.float16))\np2 = Vector3(torch.tensor([0.0, 300.0, 0.0], dtype=torch.float16))\nactual = Hyperplane.through(p0, p1, p2).normal.data.flatten().tolist()\nexpected = [0.0, 0.0, -1.0]  # (p2 - p0) x (p1 - p0)\n",
    "check": {
     "type": "allclose",
@@ -910,7 +929,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nfrom kornia.filters import filter2d\ntorch.manual_seed(0)\nimage = torch.rand(2, 3, 6, 6)\nkernel = torch.rand(2, 3, 3)  # one kernel per sample\nactual = (filter2d(image.to(memory_format=torch.channels_last), kernel) - filter2d(image, kernel)).abs().max().item()\nexpected = 0.0\n",
    "check": {
     "type": "equal"
@@ -929,7 +948,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nfrom kornia.losses import lovasz_hinge_loss\ntorch.manual_seed(0)\npred = torch.randn(2, 1, 4, 4)\ntarget = torch.randint(0, 2, (2, 4, 4))\nactual = [str(lovasz_hinge_loss(pred.half(), target).dtype), str(lovasz_hinge_loss(pred.double(), target).dtype)]\nexpected = [\"torch.float16\", \"torch.float64\"]\n",
    "check": {
     "type": "equal"
@@ -948,7 +967,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nfrom kornia.filters import get_box_kernel1d\nkernel = get_box_kernel1d(3)\nkernel[0, 0] = 0.0\nactual = kernel.flatten().tolist()\nexpected = [0.0, 1.0 / 3.0, 1.0 / 3.0]\n",
    "check": {
     "type": "allclose",
@@ -1117,7 +1136,7 @@
     "torch",
     "kornia"
    ],
-   "fixed_in_release": null,
+   "fixed_in_release": "0.9.0",
    "repro": "import torch\nimport kornia\nR1 = kornia.geometry.axis_angle_to_rotation_matrix(torch.tensor([[0.3, -0.5, 0.8]], dtype=torch.float64))\nR2 = R1 @ kornia.geometry.axis_angle_to_rotation_matrix(torch.deg2rad(torch.tensor([[0.0, 0.0, 0.01]], dtype=torch.float64)))\nv1 = torch.tensor([[0.0, 0.6, 0.8]], dtype=torch.float64)\nv2 = (kornia.geometry.axis_angle_to_rotation_matrix(torch.deg2rad(torch.tensor([[0.01, 0.0, 0.0]], dtype=torch.float64))) @ v1[..., None])[..., 0]\nactual = [kornia.metrics.angle_error_mat(R1.float(), R2.float()).item(), kornia.metrics.angle_error_vec(v1.float(), v2.float()).item()]\nexpected = [0.01, 0.01]\n",
    "check": {
     "type": "allclose",
@@ -1167,6 +1186,46 @@
     "type": "sign"
    },
    "before_fix": "[0.0] on peft main 91ca7945; with the fix the ten forwards spread by 0.56"
+  },
+  {
+   "id": "kornia-5652-gftt-response-flat-patch-nan-gradient",
+   "repo": "kornia/kornia",
+   "pr": 5653,
+   "issue": 5652,
+   "merged_at": "2026-10-09T11:16:37Z",
+   "kind": "gradient",
+   "title": "gftt_response had a nan gradient on every flat patch",
+   "symptom": "The smaller eigenvalue of the structure tensor was 0.5 * (trace - sqrt((trace**2 - 4 * det).abs())). Wherever the 9 x 9 footprint is flat (zero padding, a black border, a saturated region) the blurred moments are exactly zero, the sqrt backward is inf * 0 = nan and the whole flat region gets a nan gradient, while harris_response and hessian_response stay finite. The patch writes the discriminant as the sum of squares (dx2 - dy2)**2 + 4 * dxy**2, which is never negative, and guards its sqrt with a where, so the gradient is zero where the discriminant is zero; the forward moves at rounding level only.",
+   "requires": [
+    "torch",
+    "kornia"
+   ],
+   "fixed_in_release": "0.9.0",
+   "repro": "import torch\nimport kornia\ntorch.manual_seed(0)\nimg = torch.zeros(1, 1, 64, 64)\nimg[..., 16:48, :] = torch.rand(1, 1, 32, 64)\nimg.requires_grad_(True)\n(grad,) = torch.autograd.grad(kornia.feature.gftt_response(img).sum(), img)\nactual = [grad[0, 0, 0, 0].item(), grad[0, 0, 63, 63].item(), float(torch.isfinite(grad).all())]\nexpected = [0.0, 0.0, 1.0]\n",
+   "check": {
+    "type": "equal"
+   },
+   "before_fix": "[nan, nan, 0.0] on kornia 0.8.3, 2048 nan pixels in the flat rows"
+  },
+  {
+   "id": "kornia-5654-equalize-clahe-residual-in-first-bins",
+   "repo": "kornia/kornia",
+   "pr": 5655,
+   "issue": 5654,
+   "merged_at": "2026-10-09T10:38:46Z",
+   "kind": "redistribution",
+   "title": "equalize_clahe wrote the clipping residual into the first histogram bins, small tiles came out tens of levels brighter than OpenCV",
+   "symptom": "After clipping a tile histogram at the clip limit, the residual of the clipped pixels was written into the first `residual` bins instead of being spread over every max(256 // residual, 1)-th bin from bin 0 as OpenCV does. The residual piled onto the darkest levels and lifted the whole lookup table, so small tiles came out up to 38 levels brighter than cv2.createCLAHE (mean over a 96 x 96 image with an 8 x 8 grid and clip 2.0). On a 2 x 4 tile holding levels 0 and 100 with clip 40, level 100 mapped to 255 where the spread residual gives 159. The patch spreads the residual in one helper shared by the scalar clip path, the per-sample clip path of RandomClahe and the differentiable path.",
+   "requires": [
+    "torch",
+    "kornia"
+   ],
+   "fixed_in_release": "0.9.0",
+   "repro": "import torch\nimport kornia\nimg = torch.tensor([[0.0] * 4, [100.0] * 4]).div(255.0)[None, None]\nout = (kornia.enhance.equalize_clahe(img, 40.0, (1, 1)) * 255).round()[0, 0]\nactual = [out[1, 0].item()]\nexpected = [159.0]\n",
+   "check": {
+    "type": "equal"
+   },
+   "before_fix": "[255.0] on kornia 0.8.3"
   }
  ]
 }

{ulpwise-0.3.3 → ulpwise-0.3.5}/python/ulpwise/scan.py

@@ -136,6 +136,12 @@ RULES = {
         "take the value from the real argument and the gradient from a safe one, f(where(at_bound, safe, x)) with the value fixed up by where, or atan2 instead of acos for an angle",
         "kornia #4229 (found by the kornia conventions audit), _cdist's clamp(min=0.0).sqrt() gave nan gradients for identical float16 descriptors on torch 2.5.1; kornia #5500, clamp(-1, 1).acos() in angle_error_mat had a -inf backward at exactly 0 and 180 degrees on torch 2.5.1",
     ),
+    "dropout-never-applied": (
+        "medium",
+        "a Dropout, DropPath or a ModuleDict of them is assigned to self and then never called, never passed on and its rate never read, anywhere in the file: the option is accepted and does nothing, so the model trains without the regularisation it reports and every training forward of one input agrees",
+        "apply it in forward, or hand its rate to the function that does (scaled_dot_product_attention(dropout_p=self.dropout.p)), or drop the argument so the caller learns it is not supported",
+        "peft #3830, OFTLayer built module_dropout into self.oft_dropout and never called it, so module_dropout=0.5 trained like 0",
+    ),
 }
 SEVERITY_ORDER = {"high": 0, "medium": 1, "info": 2}
 
@@ -653,6 +659,79 @@ def python_files(root: str, include_tests: bool = False) -> List[str]:
     return out
 
 
+DROPOUT_CLASSES = re.compile(r"^(Dropout\w*|AlphaDropout|FeatureAlphaDropout|DropPath|StochasticDepth|DropBlock\w*)$")
+DROPOUT_NAMES = re.compile(r"drop_?out|drop_?path|stochastic_?depth|drop_?block", re.I)
+DROPOUT_HOLDERS = {"ModuleDict", "ModuleList", "Identity"}
+RATE_ATTRS = {"p", "prob", "rate", "drop_prob", "drop_rate", "dropout_p", "dropout", "drop_path_rate"}
+
+
+def _dropout_like(value: ast.AST, attr: str) -> bool:
+    """Whether ``self.<attr> = value`` builds a dropout: a Dropout-class call, either branch of a conditional
+    expression, or a ModuleDict, ModuleList or Identity under a name that says dropout (filled in later)."""
+    if isinstance(value, ast.IfExp):
+        return _dropout_like(value.body, attr) or _dropout_like(value.orelse, attr)
+    if not isinstance(value, ast.Call):
+        return False
+    name, _ = _callee(value)
+    if name is None:
+        return False
+    return bool(DROPOUT_CLASSES.match(name)) or (name in DROPOUT_HOLDERS and bool(DROPOUT_NAMES.search(attr)))
+
+
+def _applied(node: ast.Attribute, parents: dict) -> bool:
+    """Whether this read of ``self.x`` uses the module: a call, a call through a subscript, an argument, a return
+    or yield, an operand, a loop target, an alias to a local, or a read of its rate. Assigning to it, filling it
+    (``self.x.update(...)``, ``self.x[k] = ...``) and reading other attributes of it are not uses."""
+    parent = parents.get(node)
+    if isinstance(parent, ast.Subscript) and parent.value is node:
+        node, parent = parent, parents.get(parent)
+    if isinstance(parent, ast.Call):
+        return parent.func is node or node in parent.args or any(k.value is node for k in parent.keywords)
+    if isinstance(parent, ast.Attribute):
+        return parent.attr in RATE_ATTRS
+    if isinstance(parent, ast.Assign):
+        return parent.value is node
+    if isinstance(parent, (ast.AnnAssign, ast.AugAssign)):
+        return parent.value is node
+    if isinstance(parent, ast.For):
+        return parent.iter is node
+    return isinstance(parent, (ast.Return, ast.Yield, ast.BinOp, ast.Compare, ast.IfExp, ast.Tuple, ast.List,
+                               ast.Dict, ast.Starred, ast.keyword, ast.FormattedValue, ast.Await, ast.BoolOp))
+
+
+def _dropouts_never_applied(tree: ast.Module, path: str, source_lines: Sequence[str]) -> List[Finding]:
+    """The ``dropout-never-applied`` rule: a file-level pass, since the module is built in one method and used, if
+    at all, in another, or in a sibling class of the same file. Subclasses of ``Sequential`` run every attribute
+    in order and are not read."""
+    parents: dict = {}
+    for node in ast.walk(tree):
+        for child in ast.iter_child_nodes(node):
+            parents[child] = node
+    used: set = set()
+    for node in ast.walk(tree):
+        if isinstance(node, ast.Attribute) and _applied(node, parents):
+            used.add(node.attr)
+    findings: List[Finding] = []
+    for cls in ast.walk(tree):
+        if not isinstance(cls, ast.ClassDef) or any(ast.unparse(b).endswith("Sequential") for b in cls.bases):
+            continue
+        for fn in cls.body:
+            if not isinstance(fn, (ast.FunctionDef, ast.AsyncFunctionDef)):
+                continue
+            for node in _own_nodes(fn):
+                if not (isinstance(node, ast.Assign) and len(node.targets) == 1):
+                    continue
+                target = node.targets[0]
+                if not (isinstance(target, ast.Attribute) and isinstance(target.value, ast.Name)
+                        and target.value.id == "self"):
+                    continue
+                if target.attr not in used and _dropout_like(node.value, target.attr):
+                    qualname = cls.name + "." + fn.name
+                    findings.append(Finding("dropout-never-applied", path, node.lineno, qualname,
+                                            source_lines[node.lineno - 1].strip()[:160]))
+    return findings
+
+
 def scan_source(source: str, path: str = "<string>") -> Tuple[List[Finding], List[HotSpot]]:
     try:
         with warnings.catch_warnings():
@@ -669,6 +748,7 @@ def scan_source(source: str, path: str = "<string>") -> Tuple[List[Finding], Lis
         findings.extend(f)
         if spot.calls:
             spots.append(spot)
+    findings.extend(_dropouts_never_applied(tree, path, lines))
     return findings, spots
 
 

{ulpwise-0.3.3 → ulpwise-0.3.5}/tests/test_scan.py

@@ -165,6 +165,17 @@ SNIPPET = textwrap.dedent(
         e = (-(x.pow(2))).exp()
         f = torch.exp(-torch.square(x))
         return a, b, c, d, e, f
+
+
+    class Unregularised(torch.nn.Module):
+        def __init__(self, dim, p):
+            super().__init__()
+            self.proj = torch.nn.Linear(dim, dim)
+            self.attn_drop = torch.nn.Dropout(p)
+            self.proj_drop = torch.nn.Dropout(p)
+
+        def forward(self, x):
+            return self.proj_drop(self.proj(x))
     '''
 )
 
@@ -182,6 +193,7 @@ EXPECTED = {
     ("acos-for-angle", "angle_between"),
     ("eps-floor", "one_hot_floor"),
     ("eps-floor", "Smoother.forward"),
+    ("dropout-never-applied", "Unregularised.__init__"),
     ("one-minus-cos", "so3_jacobian"),
     ("small-angle-division", "so3_jacobian"),
     ("where-nan-gradient", "so3_jacobian_guarded"),
@@ -215,6 +227,88 @@ def test_every_rule_fires_once_on_its_snippet_and_the_guard_silences_it():
     assert busiest.function == "misc" and busiest.calls == 7 and "exp" in busiest.names
 
 
+DROPOUTS = textwrap.dedent(
+    '''
+    import torch
+    from torch import nn
+    import torch.nn.functional as F
+
+
+    class Attention(nn.Module):
+        def __init__(self, dim, p):
+            super().__init__()
+            self.dropout = nn.Dropout(p)  # its rate is read and handed to the kernel
+
+        def forward(self, q, k, v):
+            p = self.dropout.p if self.training else 0.0
+            return F.scaled_dot_product_attention(q, k, v, dropout_p=p)
+
+
+    class Block(nn.Module):
+        def __init__(self, dim, p, drop_path):
+            super().__init__()
+            self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()  # never applied
+            self.drop = nn.Dropout(p) if p > 0.0 else nn.Identity()
+            self.mlp = Mlp(dim, drop=self.drop)  # passed on
+
+        def forward(self, x):
+            return x + self.mlp(x)
+
+
+    class Mlp(nn.Sequential):
+        def __init__(self, dim, drop):
+            super().__init__()
+            self.fc = nn.Linear(dim, dim)
+            self.drop = nn.Dropout(0.5)  # a Sequential runs every attribute
+
+
+    class Adapters(nn.Module):
+        def __init__(self):
+            super().__init__()
+            self.lora_dropout = nn.ModuleDict({})  # filled and called through a subscript
+            self.oft_dropout = nn.ModuleDict({})  # filled and never called
+
+        def update_layer(self, name, p):
+            self.lora_dropout.update(nn.ModuleDict({name: nn.Dropout(p)}))
+            self.oft_dropout.update(nn.ModuleDict({name: nn.Dropout(p)}))
+
+        def forward(self, x, name):
+            return self.lora_dropout[name](x)
+
+
+    class Head(nn.Module):
+        def __init__(self, dim, p):
+            super().__init__()
+            self.dense = nn.Linear(dim, dim)
+            self.dropout = nn.Dropout(p)  # applied by a sibling class of this file
+            self.drops = nn.ModuleList([nn.Dropout(p) for _ in range(2)])  # iterated
+
+        def forward(self, x):
+            for drop in self.drops:
+                x = drop(x)
+            return x
+
+
+    class Wrapped(nn.Module):
+        def forward(self, x):
+            return self.inner.dropout(self.dense(x))
+    '''
+)
+
+
+def test_dropout_never_applied_reads_the_whole_file_and_knows_the_ways_a_module_is_used():
+    findings, _ = scan.scan_source(DROPOUTS, "dropouts.py")
+    assert [(f.function, f.line) for f in findings if f.rule == "dropout-never-applied"] == [
+        ("Block.__init__", 20),
+        ("Adapters.__init__", 39),
+    ]
+    assert findings[0].snippet.startswith("self.drop_path = DropPath(drop_path)")
+    assert findings[1].snippet.startswith("self.oft_dropout = nn.ModuleDict({})")
+    # a float rate is not a module, and a function is not a class
+    quiet = scan.scan_source("class A:\n    def __init__(self, p):\n        self.dropout = p\n\ndef f(self):\n    self.dropout = Dropout(0.1)\n", "q.py")[0]
+    assert not quiet
+
+
 def test_scan_is_quiet_about_the_escape_sequences_of_the_scanned_file():
     source = 'import torch\n\n\ndef f(x):\n    return "\\%" + str(torch.exp(x))\n'
     with warnings.catch_warnings():