Added custom low_latency operators for dispatch/combine in the A2 dec…

[oagniqgnat]

Modification：

1. A2 and A3 are packaged separately. To package on A2, you need to run bash build.sh -a deepep2.
2. If HCCL_INTRA_PCIE_ENABLE=1 and HCCL_INTRA_ROCE_ENABLE=0 are configured on A2, a hierarchical implementation of dispatch_low_latency/combine_low_latency will be executed; otherwise, a non-hierarchical implementation will be performed.
3. In the A2 dispatch_low_latency hierarchical implementation, an additional parameter topk_weights needs to be passed. In addition, an extra 1D Tensor expand_scales with shape (A,) will be returned. expand_scales will replace topk_weights as the weight parameter for the internal kernel in low_latency_combine.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!