perf(minimax-h3): add conservative DiT velocity reuse

[L17807593]

This PR contains the Conservative DiT velocity cache, disabled by default, and depends on the lossless optimizations in PR #37.

Please review and merge #37 first. Once #37 is merged, this PR will contain only the Conservative optimization relative to main.

TeleFuser MiniMax-H3: Conservative DiT Velocity Cache

1. Summary

This implementation adds a Conservative DiT velocity cache on top of the five lossless optimizations from PR #37. The cache is disabled by default. Under a four-GPU setup with five prompt/seed pairs, the final results are:

Configuration	Avg. wall time	Speedup vs. baseline	Avg. denoise time	DiT calls	Cache hits
Baseline	78.546 s	0.000%	75.737 s	49	0
Five lossless optimizations	75.766 s	3.539%	72.934 s	49	0
Conservative only	55.450 s	29.404%	52.612 s	34	15
All optimizations	53.380 s	32.040%	50.601 s	34	15

2. Lossless optimizations from PR #37

PR #37 provides the following five lossless optimizations:

• Overlap Ulysses communication with FlashAttention 4.
• Fuse Q/K RMSNorm, RoPE, and Ulysses packing.
• Add fixed-valid and valid-only attention modes.
• Fuse valid-only gather with zero-tail merge.
• Fuse RMSNorm with indexed AdaLN modulation.

These changes are described and validated in detail in PR #37.

3. Conservative DiT velocity cache

Added MiniMaxH3DiTCacheConfig:

mode="off"
start_ratio=0.2
end_ratio=0.8
refresh_interval=2
max_consecutive_reuse=1
threshold=None

conservative is accepted as an alias for velocity. In the default validation schedule, caching is active only between 20% and 80% of the denoising process. A DiT refresh is forced every other step, and at most one consecutive step reuses the previous video/audio velocity.

When the cache is hit, the previous DiT-predicted video and audio velocities are reused and the current DiT invocation is skipped.

4. Per-case performance

Case	Baseline (s)	Lossless (s)	Conservative only (s)	All optimizations (s)
P0/S0	78.426	75.745	55.347	53.503
P0/S42	78.420	75.893	55.618	53.334
P0/S1234	78.801	75.740	55.423	53.413
P1/S0	78.612	75.706	55.400	53.366
P1/S42	78.473	75.745	55.463	53.284

5. Quality verification

All quality comparisons use the baseline generated with the same prompt and seed, comparing the final video and audio MP4 outputs.

Configuration	Exact MP4 SHA match	Video PSNR	Video SSIM	Audio cosine	Audio SNR	Audio RMSE
Five lossless optimizations	5/5	inf	1.000000	1.000000	inf	0
Conservative only	0/5	35.860 dB	0.964553	0.988936	18.811 dB	0.015783
All optimizations	0/5	35.860 dB	0.964553	0.988936	18.811 dB	0.015783

Additional hash checks:

• baseline == lossless: 5/5.
• lossy_only == all: 5/5.

The lossless optimizations are bit-exact with the baseline at the final video and audio file level. The Conservative cache changes the denoising trajectory and final outputs, so it is a lossy optimization. The average PSNR, SSIM, and audio metrics show that the quality difference is limited under the tested settings.

[lzx1413]

MiniMax h3已经支持了AdaTaylorCache的feature cache策略，应该包括了这种特殊情况

[L17807593]

感觉优化方向很类似但实现上没有完全包含，PR #38 中的 Conservative cache 位于 denoising loop 层，使用上一次 DiT 预测的完整 velocity 来跳过当前 DiT 调用；AdaTaylorCache 是在 block/feature 层进行缓存和近似。

[lzx1413]

@L17807593 AdaTaylorCache也是用来跳步加速的