[TransferEngine] Add Moore Threads GPUs Support

[popsiclexu]

This PR enables Moore Threads GPUDirect RDMA support for the Transfer Engine's RDMA transport. To compile and link the musa libraries, please set the USE_MUSA flag to "ON".

[alogfans]

Is there any mechanism similar to CUDA IPC?

Additionally, I think in the future we'd better allow the mixture use of Moore Threads GPUs and NVGPUs based on the next geenration of TE https://github.com/kvcache-ai/Mooncake/tree/dev/tent. Maybe we can figure out how to approach this goal.

[popsiclexu]

Is there any mechanism similar to CUDA IPC?

Additionally, I think in the future we'd better allow the mixture use of Moore Threads GPUs and NVGPUs based on the next geenration of TE https://github.com/kvcache-ai/Mooncake/tree/dev/tent. Maybe we can figure out how to approach this goal.

Yes, MUSA has mechanism similar to CUDA IPC.

The mixture use of MTGPUS and NVGPUs based on the next geenration of TE is a great work. The usage of MUSA closely resembles that of CUDA, requiring primarily the adaptation of CUDA APIs to MUSA APIs. For reference, mooncake-transfer-engine/include/musa_porting.h is a example

[stmatengss]

Hi, @popsiclexu . Have you tested GDR-based PD in SGLang with Moore Threads GPU? (sry, we don't have the necessary testbed.

[popsiclexu]

Hi, @popsiclexu . Have you tested GDR-based PD in SGLang with Moore Threads GPU? (sry, we don't have the necessary testbed.

@stmatengss Yes, I have, and it works fine.