perf: TensorRT engine builder — static shapes, profiling, CUDA graphs#8

Open

forkni wants to merge 10 commits intopr7/tier1-optimizationsfrom

pr8/trt-engine-builder

Commits on Apr 6, 2026

perf: clean up deprecated TRT 10.x API usage in engine builder and preprocessing
INTER-NYC
committed
fix(fp8): remove direct_io_types/simplify, make allocate_buffers FP8-safe

INTER-NYC
and
claude
committed
perf(trt): static spatial shapes + tactic cleanup for engine builder

INTER-NYC
and
claude
committed
perf(trt): fully static batch profiles to unlock l2tc on UNet

INTER-NYC
and
claude
committed
feat(trt): add TRT profiling infrastructure gated by STREAMDIFFUSION_PROFILE_TRT

INTER-NYC
and
claude
committed
perf(trt): document builder_optimization_level=4; tactic 0x3e9 is TRT 10.12 bug, level 3 unsafe
INTER-NYC
committed
fix(trt): guard aten::copy behind _use_prealloc to unblock ONNX export with use_cached_attn

INTER-NYC
and
claude
committed
fix: guard ControlNet TRT engine compilation behind acceleration check
INTER-NYC
committed
fix(controlnet): pass pipeline resolution to TRT engine builder

INTER-NYC
and
claude
committed
perf(controlnet): enable CUDA graphs for ControlNet TRT engine

INTER-NYC
and
claude
committed