Add TurboT2AV SageSLA, FastNorm, and W8A8 acceleration submodule

[liuyuxiang1021]

Summary

This PR adds TurboT2AV as an optional TurboDiffusion submodule for accelerated text-to-audio-video inference.

The submodule main branch is based on the clean origin_inference branch and contains inference-only integration:

• SageAttention and SLA/SageSLA adapters for selected LTX self-attention modules
• FastNorm replacements for module and functional norm paths, plus fused modulation, residual, and RoPE helpers
• W8A8 Linear support with TurboDiffusion's strict prequantized backend and an H20-oriented TileLang post-scale backend
• attention scope, SLA top-k, resolution override, timing, and prequantized-checkpoint CLI controls
• a reproducible H20 latency breakdown and a documented speed/quality tradeoff

Training scripts, training workflow changes, and trained SLA adapter checkpoints are not included.

Defaults

The original inference behavior remains the CLI default:

• --attention_type default
• --fast_norm disabled
• --quant_linear disabled
• --sla_topk 1.0 when SLA/SageSLA is selected without an explicit ratio

The recommended accelerated command is opt-in and uses self-attention SageSLA with topk=0.3, FastNorm, and TileLang W8A8 Linear.

Cumulative H20 validation

The breakdown follows TurboDiffusion's cumulative order: optimize the teacher with W8A8/FastNorm, switch to the rCM-distilled student while retaining those optimizations, and finally add SageSLA. Generator-only measurements use one NVIDIA H20, 1024x1792, 121 frames, and --skip_decode.

Stage	Generator latency	Speedup vs previous	Speedup vs teacher
LTX-2-19B teacher (40 steps)	318.7405s/video	-	1.00x
+ W8A8 and FastNorm	233.3424s/video	1.37x	1.37x
+ rCM (4-step student)	11.7628s/video	19.84x	27.10x
+ SageSLA topk=0.3	5.8689s/video	2.00x	54.31x

The long teacher rows use one measured generation; student rows report medians from repeated generations. The non-cumulative pure 4-step student takes 16.1096s/video, making the final path 2.75x faster than the pure student as well.

At this resolution the video latent contains 28,672 self-attention tokens. Component checks measured SageSLA self-attention at 4.82x over SDPA and the two dominant TileLang W8A8 FFN GEMM shapes at 1.56x and 1.61x over BF16.

topk=0.3 is the documented high-resolution speed/quality tradeoff based on decoded prompt comparisons. SageSLA is approximate sparse attention, so lower ratios can change generated content more visibly.

Submodule

• Repository: liuyuxiang1021/turbo-t2av
• Branch: main
• Current commit: b57fb9bbbe92324b1aebc5b7e5b17ded3538aa8d
• Clean inference base: origin_inference at c99f03b7b615661f63513b3816ea6c62b754c5ce

The parent .gitmodules entry tracks the submodule's main branch, while the gitlink pins the exact commit above.

Validation

• git diff --check
• Python byte-compilation of the acceleration, FastNorm, TileLang W8A8, prequantization, inference-runner, and plotting modules
• regenerated and inspected the 1024x1792 acceleration figure
• verified both local worktrees match their pushed remote branches
• scanned submitted files for tokens, proxy settings, and workspace-local secret paths