NVIDIA/warp v1.10.1

v1.10.1

Repository: NVIDIA/warp

Tag: v1.10.1

Published: 2025-12-01T13:13:27Z

Prerelease: no

Release notes:

Warp v1.10.1

Warp v1.10.1 is a bugfix release following v1.10.0. For a complete list of changes, see the changelog.

Highlights

This is primarily a bugfix release with no major new features. Key fixes include:

• Module reuse with `module="unique"`: Fixed kernels using @wp.kernel(module="unique") to properly reuse existing module objects when the kernel is defined multiple times, avoiding unnecessary module creation overhead.
• Kernel-local arrays: Fixed several issues with arrays created using wp.zeros() inside kernels, including .ptr access, indexing for subarrays, and accepting single integers for the shape parameter.
• Custom gradients: Fixed a code-generation ordering bug that could prevent custom gradient functions (@wp.func_grad) from compiling when used with nested function calls.
• FEM improvements: Fixed invalid reads when using wp.fem.TemporaryStore during tape capture and resolved reference cycles in wp.fem.Temporary and wp.fem.ShapeBasisSpace.

Announcements

Upcoming removals

The following feature is deprecated and will be removed in v1.11 (planned for January 2026):

• `graph_compatible` parameter in `jax_callable()`: The boolean graph_compatible flag has been deprecated in favor of the new graph_mode parameter which accepts GraphMode enum values. Use GraphMode.JAX, GraphMode.WARP, or GraphMode.NONE instead.

# Deprecated (v1.10.1, will be removed in v1.11)
callable = wp.jax_experimental.jax_callable(func, graph_compatible=True)

# Use instead
from warp.jax_experimental import GraphMode
callable = wp.jax_experimental.jax_callable(func, graph_mode=GraphMode.JAX)

Platform support

• Python 3.8: We plan to drop support for Python 3.8 (end-of-life since October 2024) starting with v1.11.
• CUDA Toolkit: Starting with v1.11, the default pre-built wheels published on PyPI will be built with CUDA Toolkit 12.9 instead of 12.8. This does not change driver requirements but enables new compiler options to control the tradeoff between kernel compilation speed and runtime performance. We plan a second transition to CUDA Toolkit 13.x in mid-2026.

Acknowledgments

We thank the following contributors:

• @mehdiataei for fixing loop unrolling with wp.static() expressions that prevented certain code patterns from compiling correctly.