The grid is sized in block pairs, so its last program always covers a
whole group of 8 blocks. The absmax store was unmasked, which wrote up
to 6 entries past the end of the absmax tensor whenever the block count
was not a multiple of 8.

wide_n_simt already routes sm120 with >=48 SMs to the SIMT kernel at M<=6 with wide N. GB10 (sm121, 48 SMs) clears the same SM threshold but was excluded by the minor-version check, so it fell through to MMA.

Measured on GB10: all 8 rerouted cells are faster with SIMT, 1.03x to 1.86x. Enumeration over 78,936 and 177,540 dispatch decisions confirms the change is confined to cc_maj==12 && cc_min==1 at M<=6, False->True only, with no other architecture affected.

…#2036)

Bumps the actions group with 2 updates in the / directory: [Jimver/cuda-toolkit](https://github.com/jimver/cuda-toolkit) and [pypa/gh-action-pypi-publish](https://github.com/pypa/gh-action-pypi-publish).


Updates `Jimver/cuda-toolkit` from 0.2.35 to 0.2.36
- [Commits](Jimver/cuda-toolkit@3d45d15...b8bf9c6)

Updates `pypa/gh-action-pypi-publish` from 1.14.1 to 1.14.2
- [Commits](pypa/gh-action-pypi-publish@ba38be9...dc37677)

---
updated-dependencies:
- dependency-name: Jimver/cuda-toolkit
  dependency-version: 0.2.36
  dependency-type: direct:production
  update-type: version-update:semver-patch
  dependency-group: actions
- dependency-name: pypa/gh-action-pypi-publish
  dependency-version: 1.14.2
  dependency-type: direct:production
  update-type: version-update:semver-patch
  dependency-group: actions
...

Signed-off-by: dependabot[bot] <support@github.com>
Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>

…1888)

* Fix ROCm library version detection

Use ROCm release metadata consistently at build and runtime while keeping packaged-binary fallbacks explicit.

* Clarify native library fallback order

Document the current same-major and ROCm cross-major selection priority more explicitly.

* Add examples to library fallback docs

Illustrate older and newer same-major fallback ordering in the loader docstring.

* Fix ROCm artifact collection

Only pass shared libraries that exist to the packaging copy step.

…ersion (…" (#2060)

This reverts commit 3ac90c4.

* ci: add RDNA1 targets to ROCm 7.14 builds

Include gfx1010-gfx1012 in the ROCm 7.14 artifacts and document the expanded target coverage.

* docs: document RDNA1 support

Include gfx101X in the platform support summary alongside the other supported RDNA families.

* ci: align ROCm architecture coverage

Build the older Linux releases and ROCm 7.14 Windows artifacts for every architecture available from their corresponding ROCm distributions, and document the expanded support.

* build: align default ROCm targets

Include gfx908 and gfx1030 in source builds that do not provide an explicit architecture override.

* CI: add ROCm 10.0 build

* CI: add ROCm 10.0 Windows build

* CI: document ROCm 10.0 support

* CI: flatten ROCm index selection

* CI: default to legacy ROCm package index

* docs: update TheRock package indexes