[Diffusion][HunyuanVideo] Fuse eager QKV packing and high-quality QKNorm

[BBuf]

Summary

Reduce HunyuanVideo eager overhead with three scoped paths:

• a lossless Triton kernel that applies image RoPE while packing image/text Q, K and V into contiguous attention inputs
• a quality=high strided QK RMSNorm path that avoids input contiguous copies
• the existing quality-gated fused linear+GELU epilogue for the Hunyuan MLP

The QKV/RoPE pack uses per-signature first-use bit-exact verification. The numerically relaxed RMSNorm and GELU paths are mounted only for quality=high. The cleanup keeps mounting all-or-nothing when the public FlashInfer CuTe QK RMSNorm entry point is unavailable, so no partial quality configuration leaks into the model.

B300 performance

HunyuanVideo, 848x480x65 frames, seed 42, 30 steps, native backend:

configuration	denoise	e2e	vs latest-main eager
latest main eager	23.7185 s	26.7040 s	baseline
PR lossless eager	22.8496 s	25.8321 s	+3.66% / +3.26%
PR quality=high eager	22.3054 s	25.2885 s	+5.96% / +5.30%
PR torch.compile	21.6219 s	23.7775 s	+8.84% / +10.96%

quality=high is 2.38% faster in denoise than the PR's lossless path. It remains 3.16% behind compile in denoise and 6.35% behind end to end; the table intentionally reports that remaining gap rather than implying parity.

The first compile attempt hit the benchmark driver's 180-second warmup limit while compiling the VAE. A retry with the same generated cache and a 900-second limit completed successfully; compilation itself took about 11 minutes, while the displayed measurements exclude warmup.

Measured against main dbebc1deb42b00befa3d0de67265d7003994c1ad. The B300 validation tree is identical to this PR's final Git tree.

Output comparison

All videos use the same prompt, seed, resolution, frame count and step count.

Lossless path

The latest-main and PR lossless MP4 files are byte-identical (0345e7c0ace3ed8adeebbf967a068c58a2e757f6be9b58a1672f77056164a67a).

latest main eager (before)	PR lossless eager (after)

quality=high path

The high-quality output versus the latest-main eager video measures PSNR 43.35 dB and SSIM 0.9851 over all 65 frames.

latest main eager	PR quality=high eager

Click either frame sheet to open the corresponding MP4.

Validation

• B300 focused QKV/RoPE, mount/unmount and QK RMSNorm tests: 4 passed
• exactness covers two production-relevant QKV shapes and the unmounted reference path
• changed-file pre-commit suite passed on the final tree
• weight cache cleanup removed 39.03 GiB after the model runs

---

CI States

Latest PR Test (Base): ❌ Run #31687604964
Latest PR Test (Extra): ✅ Run #31693045443