GROOT N1.7 RTX full graph coverage: 16.61 ms E2E

[LiangSu8899]

Summary

• capture the existing RTX FP8 ViT -> DeepStack -> LLM -> VLSA backbone as a reusable CUDA Graph
• keep observation and language inputs in persistent buffers and refresh them before each replay
• allow infer(..., aux=fresh_aux) to run the backbone graph followed by the existing action graph in one public E2E call
• preserve the existing infer(...) behavior when aux is omitted

Root cause

The RTX action head already replayed a CUDA Graph covering cross-KV refresh, state encoding, all diffusion steps, and action decoding. The FP8 backbone still rebuilt its Python launch sequence for every observation, despite the Thor FP8 frontend already having the required persistent-buffer backbone graph pattern.

This change ports that established pattern to RTX. It does not change CUDA kernels, CMake targets, quantization, or action-head graph behavior.

RTX 5090 E2E validation

Measured with PyTorch 2.11.0, CUDA 13.0, the production SM120 FP8 frontend, a real two-view fixture, fixed initial noise, and 30 steady-state iterations.

The timing boundary is one fresh model-input bundle through backbone, action head, all four diffusion steps, decode, final normalized action, and CUDA synchronization.

Path	E2E P50
Before: eager backbone + action graph	19.47 ms
After: backbone graph + action graph	16.61 ms

Output validation:

• shape: (1, 40, 132)
• finite: yes
• fixed-noise comparison against the previous path: bit-exact
• max absolute error: 0.0
• cosine similarity: 1.0

Checks

• 75 passed, 84 skipped across the new graph tests and related GROOT N1.7/load-model tests
• Python compileall
• git diff --check