perf: Inference performance & pipeline correctness#4

Open

forkni wants to merge 11 commits intoSDTD_031_devfrom

pr1/inference-performance

Commits on Apr 3, 2026

Fix SDXL TensorRT engine build failure on Windows

dotsimulate
authored and
INTER-NYC
committed
feat: clone decode_image output to prevent TRT VAE buffer reuse

INTER-NYC
and
claude
committed
perf: Tier 2 inference performance optimizations

INTER-NYC
and
claude
committed
Add cuda-python 13.x compatibility fix for cudart import

INTER-NYC
and
claude
committed
Quick-win CUDA optimizations: pre-allocated buffers + L2 cache persistence

INTER-NYC
and
claude
committed
Fix hardcoded float16 autocasts and add fp32 precision for scheduler division

INTER-NYC
and
claude
committed
Fix L2 cache: second reserve call was resetting reservation to 0

INTER-NYC
and
claude
committed
fix: report inference FPS separately from output FPS when similar image filter is active

INTER-NYC
and
claude
committed
perf: pre-allocate image output buffers, replace .clone() with .copy_() in pipeline hot path
INTER-NYC
committed

Commits on Apr 4, 2026