Skip to content

v0.7.1

Choose a tag to compare

@frlai frlai released this 15 Sep 22:06
· 9 commits to main since this release
272ff17

A bugfix release that tightens Warp runtime setup, preserves shared-buffer input aliases across repeated tracing passes, and avoids normalizing opaque Warp tuple descriptors.

Fixed:

  • Enable Warp patching only when Warp is initialized and the required CUPTI and runtime artifacts are available. Installation guidance now points to the CUDA-specific extras and runtime build command.
  • Preserve coalesced input aliases when a compiled node is entered again, preventing shared allocations from being reported as new I/O.
  • Skip tuple subclasses such as Warp BuiltinCallDesc during argument normalization so incidental calls such as wp.rand_init do not fail.