* The MSCCL feature is now disabled by default. The `--disable-msccl-kernel` build flag is replaced with `--enable-msccl-kernel` in the `rccl/install.sh` script.
* MSCCL and NPKIT are deprecated and will be removed in a future release of RCCL.
* RCCL error messages have been made more verbose in several cases. RCCL now prints out fatal error messages by default. Fatal error messages can be suppressed by setting `NCCL_DEBUG=NONE`.
* Experimental support for traffic shaping using warp specialization (also known as WarpSpeed) is now available for the Ring algorithm.
* Enabling WarpSpeed in auto mode using RCCL_WARP_SPEED_AUTO optimizes performance and reduces the CU count by 50% on a single node for AllReduce, AllGather from 64MB, and ReduceScatter from 256MB.
* The following configuration knobs control WarpSpeed behavior for debugging purposes: `RCCL_WARP_SPEED_ENABLE`, `RCCL_UNROLL_FACTOR`, `RCCL_WARP_SPEED_CU_COUNT`, and `RCCL_THREADS_PER_BLOCK`. Note that the effective unroll factor is calculated as 2 raised to the value of `RCCL_UNROLL_FACTOR`.
* Added `RCCL_IB_QPS_PER_P2P` to set the number of QPs per connection for P2P operations. When set (≥1), P2P operations (Send/Recv) use `RCCL_IB_QPS_PER_P2P`, while other collective operations continue to use `NCCL_IB_QPS_PER_CONNECTION`. When not set, `NCCL_IB_QPS_PER_CONNECTION` applies to all operations.
* Added `RCCL_P2P_BATCH_THRESHOLD` to set the message size limit for batching P2P operations. This mainly affects small message performance for alltoall at a large scale but also applies to alltoallv.
* Added `RCCL_P2P_BATCH_ENABLE` to enable batching P2P operations to receive performance gains for smaller messages up to 4MB for alltoall when the workload requires it. This is to avoid performance dips for larger messages.
* The MSCCL++ feature is now disabled by default. The `--disable-mscclpp` build flag is replaced with `--enable-mscclpp` in the `rccl/install.sh` script.
* Fixed the known issue "When splitting a communicator using `ncclCommSplit` in some GPU configurations, MSCCL initialization can cause a segmentation fault." with a design change to use `comm` instead of `rank` for `mscclStatus`. The Global map for `comm` to `mscclStatus` is still not thread safe but should be explicitly handled by mutexes for read writes. This is tested for correctness, but there is a plan to use a thread-safe map data structure in upcoming changes.
* Added MSCCL support for AllGather multinode gfx942/gfx950 (i.e., 16 and 32 GPUs). To enable, set the environment variable `RCCL_MSCCL_FORCE_ENABLE=1`. Max message size for MSCCL AllGather usage is `12292 * sizeof(datatype) * nGPUs`.
* Thread thresholds for LL/LL128 are selected in Tuning Models for the MI300X. This impacts the number of channels used for AG and RS. Channel tuning model is bypassed if `NCCL_THREAD_THRESHOLDS`, `NCCL_MIN_NCHANNELS', or 'NCCL_MAX_NCHANNELS` are set.
* Multi-node tuning for AllGather, AllReduce, and ReduceScatter that leverages LL/LL64/LL128 protocol to use nontemporal vector load/store for tunable message size ranges.
* LL/LL128 usage ranges for AR, AG, and RS are part of the tuning models, which enable architecture-specific tuning in conjunction with the existing Rome Models scheme in RCCL.
* Two new APIs are exposed as part of an initiative to separate RCCL code. These APIs are `rcclGetAlgoInfo` and `rcclFuncMaxSendRecvCount`. However, user-level invocation requires that RCCL be built with `RCCL_EXPOSE_STATIC` enabled.
* Enabled double-buffering in `reduceCopyPacks` to trigger pipelining, especially to overlap `bf16` arithmetic and bridge the gap between `fp32` performance and `bf16` for both `gfx942` and `gfx950`. Pipelining has been made tunable via `rcclSetPipelining`, similar to algorithms/protocols so that regression is avoided in certain message sizes.
* Added `RCCL_OVERRIDE_PROTO` and `RCCL_OVERRIDE_ALGO` to allow direct replacement of protocol and algorithm choices. Unlike `NCCL_PROTO` and `NCCL_ALGO`, which re-run the model across enabled combinations and may not guarantee the intended override, these new options enforce the specified selections explicitly.
* Within the RCCL-UnitTests test suite, failures occur in tests ending with the `ManagedMem` and `ManagedMemGraph` suffixes. These failures only affect the test results and do not affect the RCCL component itself. This issue will be resolved in the next major release.
- On systems running Linux kernel 6.8.0, such as Ubuntu 24.04, Direct Memory Access (DMA) transfers between the GPU and NIC are disabled and impacts multi-node RCCL performance.
- This issue was reproduced with RCCL 2.20.5 (ROCm 6.2.0 and 6.2.1) on systems with Broadcom Thor-2 NICs and affects other systems with RoCE networks using Linux 6.8.0 or newer.
- Older RCCL versions are also impacted.
- This issue will be addressed in a future ROCm release.
- Packaging split into a runtime package called rccl and a development package called rccl-devel. The development package depends on runtime. The runtime package suggests the development package for all supported OSes except CentOS 7 to aid in the transition. The suggests feature in packaging is introduced as a deprecated feature and will be removed in a future rocm release.
- Ability to select the number of channels to use for clique-based all reduce (RCCL_CLIQUE_ALLREDUCE_NCHANNELS). This can be adjusted to tune for performance when computation kernels are being executed in parallel.
- Clique-based kernels are currently experimental and have not been fully tested on all topologies. By default, clique-based kernels are disabled if the detected topology is not supported (override with RCCL_FORCE_ENABLE_CLIQUE)
- Clique-based kernels may hang if there are differences between environment variables set across ranks.
- Clique-based kernels may fail if the input / output device pointers are not the base device pointers returned by hipMalloc.
