Repository navigation

[None][perf] Optimize GVR prefill windows and BF16 refinement - #19416

Draft

longcheng-nv wants to merge 5 commits into

NVIDIA:mainfrom

longcheng-nv:perf/gvr-v2-window-refinement

Draft

[None][perf] Optimize GVR prefill windows and BF16 refinement#19416
longcheng-nv wants to merge 5 commits into
NVIDIA:mainfrom
longcheng-nv:perf/gvr-v2-window-refinement

Commits on Sep 16, 2026
- [None][feat] Add native BF16 self-sampling GVR decode
  longcheng-nv
  committed
  c75cb10
- [None][refactor] Unify FP32 and BF16 GVR decode templates
  longcheng-nv
  committed
  351545d
Commits on Sep 17, 2026
- [None][perf] Preserve the original FP32 GVR host fast path
  longcheng-nv
  committed
  ecd61cb
- refactor: consolidate GVR dtype host and device specializations
  longcheng-nv
  committed
  c7f3437
Commits on Sep 18, 2026
- [None][perf] Optimize GVR prefill windows and BF16 refinement
  longcheng-nv
  committed
  7470e04