Gate TensorRT finite diagnostics behind CF_TRT_DIAGNOSTICS

Enable full SLat DAE TensorRT by default

Add SLat DAE full TensorRT finite-value diagnostics

Fix skeleton grouping dtype contract for TensorRT SLat DAE

Use canonical x4 head coordinates for full SLat DAE TensorRT composition

Expand full SLat DAE runtime head profile to 8192 and log token counts

Add opt-in checkpoint-free full SLat DAE custom TensorRT runtime

Add checkpoint-free full SLat DAE TensorRT decoder runtime

Add zero-copy TensorRT runner with data-dependent output allocation

Fix TensorRT skin decoder runtime dtype bindings

Fix dynamic TensorRT skin output allocation

Use robust CPU SLat shell construction for positional buffers

Fix SLat TensorRT adapter to exact AniGen forward_stage API

Fix top-level runtime imports in HF Jobs

Activate shell-only SLat TensorRT production runtime

Support shell-only SLat TensorRT runtime with lazy native fallback

Add lightweight AniGen SLat sparse-shell loader

Add hybrid TensorRT adapter for AniGen SLat transformer core

Avoid downloading inactive SLat TensorRT plan

Use validated TensorRT SS Flow decoder and skin decoder in AniGen runtime

Add TensorRT wrappers for AniGen SS and skin models

Expose BF16 facade around FP16 FLUX TensorRT plan

Make FLUX TensorRT wrapper compatible with Accelerate CPU offload

Lazy-load FLUX.2 TensorRT engine after prompt encoding

Add Diffusers-compatible FLUX.2 TensorRT transformer wrapper

Avoid downloading FLUX engine in AniGen 3D Jobs

Remove generated pycache from runtime repo

Remove generated pycache from runtime repo

Remove generated pycache from runtime repo

Stage dense conditioning before sparse AniGen model load

Gate TRT by validation and prefer native DSINE fallback

Lazy-download ONNX only when TensorRT plan is unavailable

Select TensorRT or ORT CUDA independently per AniGen dense component

Add zero-copy ORT CUDA fallback before PyTorch

Add zero-copy TensorRT PyTorch runtime and AniGen hybrid loader

Add L4 ONNX TensorRT hybrid runtime scaffold