| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
Add CUDA 13.3 support (#582) * Generated bindings: cuda-13030 * Non-generated support code for cuda-13030 * bindings_generator: support cuda 13.3 NVIDIA renamed the cuda_cccl redist module to cccl in the 13.3 manifest. Match on (major, minor) so 13.3+ uses the new name, and add 13.3.0 to the generated version list. | 3 个月前 | |
regen on linux | 3 个月前 | |
Load symbol on use instead of at initialization (#573) * Lazy load symbols on use instead of at initialization * Tidy * Regenerate bindings * Fix missing braces on extern c functions * Adding rustfmt::skip to bindings * Move rustfmt::skip to mod level | 4 个月前 | |
Add CUDA 13.3 support (#582) * Generated bindings: cuda-13030 * Non-generated support code for cuda-13030 * bindings_generator: support cuda 13.3 NVIDIA renamed the cuda_cccl redist module to cccl in the 13.3 manifest. Match on (major, minor) so 13.3+ uses the new name, and add 13.3.0 to the generated version list. | 3 个月前 | |
Add CUDA 13.3 support (#582) * Generated bindings: cuda-13030 * Non-generated support code for cuda-13030 * bindings_generator: support cuda 13.3 NVIDIA renamed the cuda_cccl redist module to cccl in the 13.3 manifest. Match on (major, minor) so 13.3+ uses the new name, and add 13.3.0 to the generated version list. | 3 个月前 | |
Added some missing cupti headers for new cuda versions (#592) Co-authored-by: Your Name <you@example.com> | 1 个月前 | |
Add CUDA 13.3 support (#582) * Generated bindings: cuda-13030 * Non-generated support code for cuda-13030 * bindings_generator: support cuda 13.3 NVIDIA renamed the cuda_cccl redist module to cccl in the 13.3 manifest. Match on (major, minor) so 13.3+ uses the new name, and add 13.3.0 to the generated version list. | 3 个月前 | |
regen on linux | 3 个月前 | |
Add CUDA 13.3 support (#582) * Generated bindings: cuda-13030 * Non-generated support code for cuda-13030 * bindings_generator: support cuda 13.3 NVIDIA renamed the cuda_cccl redist module to cccl in the 13.3 manifest. Match on (major, minor) so 13.3+ uses the new name, and add 13.3.0 to the generated version list. | 3 个月前 | |
regen on linux | 3 个月前 | |
regen on linux | 3 个月前 | |
feat(driver): allow selecting pinned host allocation flags (#580) * feat(driver): allow selecting pinned host allocation flags Expose alloc_pinned_with_flags so callers can choose default pinned memory when CPU reads matter while retaining the existing write-combined default. * chore: add pinned host CPU read performance test Measure default and write-combined pinned allocations on the CPU and guard against choosing read-hostile flags for read-heavy buffers. * feat(driver): prevent pinned host allocation size overflow Compute pinned allocation bytes without wrapping before calling CUDA while preserving the existing panic-based slice-size invariant. Add a regression test for overflowing typed lengths. * chore: document write-combined behavior on default pinned allocation Keep the configurable allocation API neutral while warning users of the CPU-read cost on the convenience method that selects write-combined memory. * chore(driver): clarify pinned allocation flags List the three cuMemHostAlloc options and mirror CUDA's orthogonal flag wording so the API docs avoid presenting raw zero as a named Driver API flag. | 1 个月前 | |
Adding safe `Group` api to nccl (#578) * #575 Adding safe nccl::Group api * Rework Group to use CudaView/CudaViewMut explicitly * Adding missing group_start call & docs | 4 个月前 | |
Add CUDA 13.3 support (#582) * Generated bindings: cuda-13030 * Non-generated support code for cuda-13030 * bindings_generator: support cuda 13.3 NVIDIA renamed the cuda_cccl redist module to cccl in the 13.3 manifest. Match on (major, minor) so 13.3+ uses the new name, and add 13.3.0 to the generated version list. | 3 个月前 | |
Add CUDA 13.3 support (#582) * Generated bindings: cuda-13030 * Non-generated support code for cuda-13030 * bindings_generator: support cuda 13.3 NVIDIA renamed the cuda_cccl redist module to cccl in the 13.3 manifest. Match on (major, minor) so 13.3+ uses the new name, and add 13.3.0 to the generated version list. | 3 个月前 | |
Add CUDA 13.3 support (#582) * Generated bindings: cuda-13030 * Non-generated support code for cuda-13030 * bindings_generator: support cuda 13.3 NVIDIA renamed the cuda_cccl redist module to cccl in the 13.3 manifest. Match on (major, minor) so 13.3+ uses the new name, and add 13.3.0 to the generated version list. | 3 个月前 | |
Supporting multiple NCCL/CUDNN versions via features (#564) * nccl, cudnn, cutensor specific versions * Tidy * Par iter for nccl/cudnn * tidy * Simplify review slightly * Tidy * Bug fixes & more parallel * Rework version keys * bindings generator now working * Finally fix cudnn version check * Fixes for cfg system * Cargo fmt * cargo clippy * Drop support for nccl 2.18-2.21 (2.22 first release without cuda 11 support) * Removing old nccl versions from bindings | 4 个月前 | |
feat: wire FP4 packed types from float4 0.2.0 into cudarc - Bump float4 dependency from 0.1.0 to 0.2.0 - Add DeviceRepr and ValidAsZeroBits for F4E2M1x2 (packed 2-per-byte) - Add CudaTypeName for F4E2M1, F4E2M1x2, and E8M0 - Add CudnnDataType for F4E2M1 (gated to cuda-12080+) | 6 个月前 |