Releases: nv-legate/legate-sparse
v25.01.00
This release supports the unification of two different memory pools in Legate - eager and deferred. As a result of this unification of memory pools, the option --eager-alloc-percentage is no longer supported. Misalignment errors for sparse matrix-matrix multiplication have been resolved.
V24.11.00
41ed1ac
shriram-jagan
Shriram Jagannathan
v24.11.00 is an alpha release of Legate Sparse and now uses the C++ Legate core.
Conda packages are available here: https://anaconda.org/legate/legate-sparse. Note that the link to legate sparse homepage is incorrect on the conda page.
Notes
-
The implementation of spGEMM has computational hints that prevent copying over data from GPUs to CPU and vice-versa for partitioning. Multi-GPU runs that use spGEMM may benefit from this improvement
-
Support for two different spGEMM algorithms from cuSparse is now included. A fast but memory hungry version can be turned on by setting the environment variable LEGATE_SPARSE_FAST_SPGEMM to 1
-
Support for all matrix formats other than CSR has been dropped in this release but might be included over the next several releases
