v2.0 — exact_kv / exact_kv_and_rows + dense trailing steps

v2 — exact_kv / exact_kv_and_rows + dense trailing steps

Two new quality-preserving options, guided by recommendations from @Kijai.

1. Last steps dense — dense_steps / step_off

Sol-Attn's sparse error is concentrated in the early denoising steps. At the very end of sampling the noise is weakest, so a sparse approximation there is the most visible. Setting dense_steps=1 (or step_off=1.0) runs the final step(s) with full, exact attention — removing the last of the clamping noise at a tiny cost.

2. Exact KV sink — exact_mode (MiniMax-H3 only)

MiniMax H3's packed sequence is [text | conditioned rows | audio | video]. The rows before the video are the conditioning the model must honor faithfully — and the generated audio stream lives at the end of that prefix.

exact_mode	Cost	Effect
exact_kv	~+3%	Every query attends to the prefix KV rows exactly (no sparsification on the conditioning).
exact_kv_and_rows	~+20%	Also runs the prefix query rows dense, making the generated audio stream exact.

Recommended starting points

• Max quality (+~20% cost): exact_mode=exact_kv_and_rows, dense_steps=1
• Cheap quality boost: exact_mode=exact_kv, dense_steps=1
• Baseline (pure v1): exact_mode=off, dense_steps=0

Credits

Thanks to @Kijai for the two design suggestions behind this release.