v2.0.5

What's Changed

• Optimize cache calculations by @LoserCheems in #326
• Add Triton decode top-k gather KV by @LoserCheems in #333
• Reduce Triton graph-capture temporary allocations by @LoserCheems in #335
• Refine Triton launch config controls by @LoserCheems in #337
• Support paged KV for Triton dense, sparse, and gated fwd/dec by @LoserCheems in #339
• [BUG FIX] Prevent PyPI publish failure on workflow reruns by @LoserCheems with @Copilot in #340

Full Changelog: v2.0.4...v2.0.5