What's Changed

• Fixing bug with nccl-version-from-build-system by @chelsea0x3b in #574

Full Changelog: v0.19.5...v0.19.6

What's Changed

• Add CUDA stream priority API by @w4nderlust in #554
• Peer to peer by @gregszumel in #561
• Replace CUmemAllocationHandleType enum with u32 by @roykuper13 in #563
• Supporting multiple NCCL/CUDNN versions via features by @chelsea0x3b in #564
• Load symbol on use instead of at initialization by @chelsea0x3b in #573

New Contributors

• @gregszumel made their first contribution in #561
• @roykuper13 made their first contribution in #563

Full Changelog: v0.19.4...v0.19.5

What's Changed

• Add total_mem() and mem_get_info() safe methods to CudaContext by @OneThing98 in #534
• driver: add upload() and raw accessors to CudaGraph by @OneThing98 in #535
• docs: add cufft by @mayocream in #538
• feat: wire FP4 packed types from float4 0.2.0 by @jorgeantonio21 in #540
• Add cuCtxCreate_v4 bindings, CudaContext::new_non_primary() and new_cig() by @w4nderlust in #539
• driver: add result-level memory pool wrappers by @OneThing98 in #544
• Remove cuCtxCreate_v4 from blocklist by @chelsea0x3b in #550
• Add CUDA 13.2 Support by @TannerRogalsky in #546
• Exposing has_async_alloc field in CudaContext by @LateinCecer in #553
• Adds cuda 13.2 support by @chelsea0x3b in #551
• Parallelize bindings-generator by @chelsea0x3b in #552

New Contributors

• @OneThing98 made their first contribution in #534
• @jorgeantonio21 made their first contribution in #540
• @w4nderlust made their first contribution in #539
• @TannerRogalsky made their first contribution in #546
• @LateinCecer made their first contribution in #553

Full Changelog: v0.19.3...v0.19.4

What's Changed

• Implement DeviceRepr for arrays by @kaathewisegit in #523
• feat: cufft safe API by @mayocream in #532

Full Changelog: v0.19.2...v0.19.3

What's Changed

• fix: add support for cufft 12.x by @mayocream in #530
• Add lib{name}.so.9 by @chelsea0x3b in #531

Full Changelog: v0.19.1...v0.19.2

What's Changed

• Bump float8 to 0.7.0 by @EricLBuehler in #527
• Bump libloading 0.9.0 by @chelsea0x3b in #528

Full Changelog: v0.19.0...v0.19.1

What's Changed

• Fix memory safety issue in CudaSlice::leak and optimize Drop by @wizenink in #516
• [Breaking] get_global returns CudaViewMut by @chelsea0x3b in #517
• Add fallback for loading like there is for version by @wingertge in #518
• fixes a few issues with multi gpu usage in both candle and mistralrs by @krampenschiesser in #520
• unify memcpy peer & memcpy dtod by @chelsea0x3b in #522

New Contributors

• @wingertge made their first contribution in #518
• @krampenschiesser made their first contribution in #520

Full Changelog: v0.18.2...v0.19.0

What's Changed

• Add CUBIN binary format support to module loading by @SpenserCai in #505
• Allow accessing PTX image bytes by @TheNewJavaman in #502
• feat: cufft bindings by @mayocream in #500
• fix drop impls for cusolver by @chelsea0x3b in #506
• cublas Grouped gemm by @zackangelo in #508
• Add cuda-13010 by @chelsea0x3b in #507

New Contributors

• @SpenserCai made their first contribution in #505
• @TheNewJavaman made their first contribution in #502
• @mayocream made their first contribution in #500

Full Changelog: v0.18.1...v0.18.2

What's Changed

• More cublas filters for cuda-13000 by @chelsea0x3b in #496

Full Changelog: v0.18.0...v0.18.1

What's Changed

• Adding CudaContext::per_thread_stream() by @chelsea0x3b in #491
• adds more filters for cuda 13 by @chelsea0x3b in #492
• Deprecate memcpy_stod,memcpy_dtov. Adds clone_htod,clone_dtoh by @chelsea0x3b in #493
• Small docs updates by @chelsea0x3b in #494
• [Breaking] Fixing mutable aliasing soundness issue by @chelsea0x3b in #495

Full Changelog: v0.17.8...v0.18.0