[perf] LTX2 FP4-linear on DGX Spark (GB10/sm_121): ~24% faster denoise at 1080p#1594

Open

Mister-Raggs wants to merge 3 commits into

hao-ai-lab:mainfrom

Mister-Raggs:spark/ltx2-fp4-linear

Commits on Jul 12, 2026

[bench] LTX2 FP4-linear vs bf16 A/B harness (Spark/GB10)
Mister-Raggs
committed

Commits on Jul 13, 2026

[bench] Default the FP4-linear A/B to 2 warmup + 5 measured runs
Mister-Raggs
committed
[bench] Address review: validate warmup<runs, guard CUDA sync, dict-safe stages
Mister-Raggs
committed