feat: FP8 quantization & TensorRT build infrastructure#6

Open

forkni wants to merge 9 commits intopr3/ipadapter-vram-depsfrom

pr2/fp8-tensorrt-build

Commits on Apr 4, 2026

fix: patch ByteSize() for >2GB ONNX in modelopt FP8 quantization
INTER-NYC
committed
fix: reduce FP8 calibration batches 128→8 (KVO cache OOM, 281GB→17GB)
INTER-NYC
committed
fix: merge calibration list-of-dicts into stacked dict for modelopt CalibrationDataProvider
INTER-NYC
committed
fix: add NVIDIA DLLs to PATH and retry without quantize_mha on ORT EP failure
INTER-NYC
committed
fix: use single calibration batch for modelopt (avoid rank mismatch), cleanup intermediates on retry
INTER-NYC
committed
fix: resolve 4 FP8 quantization bugs for TRT 10.12 cached attention engine build

INTER-NYC
and
claude
committed
fix: prevent FP8 engine build intermediate file bloat on Windows

INTER-NYC
and
claude
committed
feat: add FP8 quantization support to wrapper, compile_unet, and engine path

INTER-NYC
and
claude
committed
perf: pre-allocate KV clone buffers in CachedSTAttnProcessor2_0
INTER-NYC
committed