DSA kernels

[motinwing]

PR Category

[ Operator ]

Type of Change

[ New Feature ]

Description

Added three Triton operators required for DeepSeekV3.2 inference (Deepseek Sparse Attention, DSA), including bin_topk, indexer, and sparse_mla. The operators have been validated.

Issue

This PR addresses the lack of core operators required for DeepSeek-V3.2 in the operator library, thereby providing the necessary low-level support for inference with this model.

Progress

• [√] Change is properly reviewed (1 reviewer required, 2 recommended).
• [√] Change is responded to an issue.
• [√] Change is fully covered by a UT.

Performance

Using the TileLang operator implementations as the benchmark, tests show that these Triton operators achieve over 80% of the benchmark's performance.

[CLAassistant]

All committers have signed the CLA.