DeepSeek ports sparse indexer to Ascend

This is ecosystem strategy disguised as kernel code: API compatibility lowers the cost of moving from CUDA to Ascend, but it does not eliminate hardware access or CANN’s deployment friction.

• –MQA logits gives Ascend a core piece of DeepSeek’s sparse-attention token-selection pipeline.
• –The project targets Ascend 950 hardware and CANN 9.20, limiting immediate usefulness outside Huawei’s ecosystem.
• –DeepSeek reports near-peak hardware performance and 99% FIX-pipe utilization, but independent cross-platform benchmarks are still needed.
• –Matching DeepGEMM’s package and API shape reduces porting work, yet a single kernel library does not make the full model-serving stack portable.
• –The 34-day lag signals that Huawei support is following DeepSeek’s NVIDIA path rather than evolving as a parallel first-class target.