v0.2.0

What's Changed

• dsv32: stock gather decode now on by default, mask path remains as opt-in by @asher in #11
• feat(apc): hybrid checkpoint tier, preserve_reasoning support by @asher in #12
• perf(gemma4): batched decode global attention in one ragged kernel call by @asher in #13
• apc: evaluate layer-major shard arrays before they cross to the disk writer thread by @asher in #14
• perf(serve): quantized-KV prefill and decode take the fused kq paths by @asher in #15
• perf(serve): shared-prefix cascade decode for concurrent streams by @asher in #16
• experimental(serve): opt-in sparse top-k page decode for dense-attention models by @asher in #17
• config: improved model defaults and thinking control helpers by @asher in #18
• fix(pull): download progress bar fix, improved cli help by @asher in #19
• feat(mtp): DSpark speculative decoding for DeepSeek-V4-Flash 0731 by @asher in #21
• fix(chat): enable combining --mtp with --stream-experts by @asher in #22
• feat(kimi-k3): kimi-k3 support, expert streaming memory optimizations by @asher in #20
• release 0.2 by @asher in #25

Full Changelog: v0.1.4...v0.2.0