TurboDiffusion Integration Notes

TurboT2AV adapts TurboDiffusion's inference stack to the LTX-2 joint audio-video transformer. The integration is intentionally limited to inference; the TurboT2AV distillation and training pipeline remains independent.

Reused Components

• SageSLA: TurboDiffusion's sparse-linear attention design and SLA compensation path are wrapped for LTX-2 video and audio self-attention.
• FastNorm: module and functional RMSNorm/LayerNorm operations are replaced with fused implementations, together with LTX-specific modulation, residual, and rotary-embedding helpers.
• W8A8 Linear: selected weights are quantized to INT8 at model initialization, and activations are dynamically quantized before transformer Linear operations. A BF16 copy is also retained because unsupported shapes use native Linear, so this path targets compute speed rather than checkpoint compression.

The recommended path combines SageSLA topk=0.3, FastNorm, and TileLang W8A8. The one-command installer pins the tested SageAttention and SpargeAttn commits and applies build-only compatibility patches. They keep architecture-specific CUDA sources scoped to their valid targets and do not change attention kernels. It also pins the tested TileLang and TVM FFI versions so a fresh environment does not resolve an incompatible TVM FFI runtime.

Differences From TurboDiffusion

Area	TurboDiffusion	TurboT2AV integration
Model	Wan text-to-video transformer	LTX-2 joint audio-video transformer
Tested GPU	RTX 5090 in the published latency figure	NVIDIA H20
Memory	Includes a CPU-offload stage	Full model remains on the H20; no CPU offload
Attention	Wan self-attention layout	LTX video/audio self-attention adapters; masked text cross-attention remains native
W8A8 kernel	Original CUTE/SM80 blockwise kernel	TileLang post-scale kernel tuned for H20 and LTX FFN shapes
Distillation	rCM is a separate acceleration stage	Uses the existing four-step TurboT2AV student checkpoint

The TileLang kernel keeps INT8 accumulation continuous across the K dimension and applies activation/weight scales in the epilogue. This adaptation is used because TurboDiffusion's original W8A8 kernel did not outperform BF16 Linear on the tested H20 LTX shapes.

End-to-End Result

The following generator-only measurements use one H20, 121 frames, and a 1024x1792 output resolution. Stages are cumulative.

Stage	Latency	Speedup vs previous	Speedup vs teacher
LTX-2-19B teacher, 40 steps	318.7405s	-	1.00x
+ TileLang W8A8 and FastNorm	233.3424s	1.37x	1.37x
+ four-step TurboT2AV student	12.1655s	19.18x	26.20x
+ SageSLA topk=0.3 and text trimming	5.8505s	2.08x	54.48x

The fully accelerated four-step student is 2.82x faster than the pure four-step student (16.5245s/video) and 54.48x faster than the 40-step teacher (318.7405s/video). topk=0.3 was selected as the practical speed/quality setting from paired visual comparisons; sparse attention is not numerically lossless and should be revalidated for a different resolution or prompt distribution.