Pull requests / #1335

#1335 docs: supporting QFUSE one-token commit evidence for #1209

open · @CC-David-CC · 0 comentários · No GitHub

NVIDIA / CUDADocumentation

Descrição

Supporting evidence for #1209, with no engine changes. Seven Q8/FP16-KV cases on RTX PRO 6000 Blackwell reproduce the broken one-token commit and verify an isolated equivalent guard, including MTP T4 and exact 512-token comparisons. This supports the commit guard, not the full #1209 implementation. Reproduction and results: docs/experiments/QFUSE_ONE_TOKEN_COMMIT.md.

No site

Links install, modelos, releases.