Applicability
Product |
Supported |
|---|---|
Atlas 350 Accelerator Card |
√ |
√ |
|
√ |
|
☓ |
|
√ |
|
☓ |
For
For the
Function
Functions as the operation API of the ReduceScatterV operator, which is similar to the ReduceScatter operation. The difference is that the ReduceScatterV operation supports the configuration of different data sizes for different ranks in a communicator. (The data size can be configured for different IDs in a rank, but the data size of the same ID in different ranks must be the same.) After the reduction operation (sum, prod, max, and min) is performed on data with the same ID in each rank, the result is distributed to the output buffer of each rank based on the ID.
Prototype
1 | HcclResult HcclReduceScatterV(void *sendBuf, const void *sendCounts, const void *sendDispls, void *recvBuf, uint64_t recvCount, HcclDataType dataType, HcclReduceOp op, HcclComm comm, aclrtStream stream) |
Parameters
Parameter |
Input/Output |
Description |
|---|---|---|
sendBuf |
Input |
Address of the send buffer. |
sendCounts |
Input |
Size of data in sendBuf of each rank involved in the ReduceScatterV operation, which is an array of the uint64 type. The ith element of this array indicates the data size to send to rank i. |
sendDispls |
Input |
Offset of data of each rank in sendBuf involved in the ReduceScatterV operation (unit: dataType), which is an array of the uint64 type. The ith element of this array indicates the offset in sendBuf for the data sent to rank i. |
recvBuf |
Output |
Address of the buffer to receive collective communication results. The addresses configured for recvBuf and sendBuf must be different. |
recvCount |
Input |
Size of rank-specific data in recvBuf involved in the ReduceScatterV operation. If the rank ID is i, the value of recvCount must be the same as that of the element with the index of i in the sendCounts array. |
dataType |
Input |
Data type of the ReduceScatterV operation, which is of the HcclDataType type. Atlas 350 Accelerator Card: The supported data types are int8, int16, int32, float16, float32, and bfp16. Atlas 300I Duo Inference Card: The supported data types are int16, float16, and float32. |
op |
Input |
Reduction operation type. Atlas 350 Accelerator Card: The supported operation types are sum, max, and min. Atlas 300I Duo Inference Card: Only the sum operation type is supported. |
comm |
Input |
Communicator where the operation is performed. |
stream |
Input |
Stream of the rank. |
Returns
HcclResult: HCCL_SUCCESS on success, or else failure.
Constraints
- All ranks must have the same sendCounts, sendDispls, dataType, and op.
- Atlas 350 Accelerator Card: Only the single-server scenario is supported, and operators must be expanded on CCU.
Atlas A3 training product /Atlas A3 inference product : Only the single-server scenario is supported.Atlas A2 training product /Atlas A2 inference product : Only the symmetric multi-server distribution scenario is supported. The asymmetric distribution (that is, asymmetric number of devices) scenario is not supported.- Atlas 300I Duo Inference Card: Only the single-server scenario is supported, and a maximum of 2 Atlas 300I Duo Inference Cards (4 NPUs) can be deployed on a single server.
- The input and output addresses (sendBuf and recvBuf) of the operator must meet the following alignment requirements based on different data types:
- int8: 1-byte aligned
- int16, float16, bfp16: 2-byte aligned
- int32 and float32: 4-byte aligned
- int64: 8-byte aligned