[DSv4.1] Add FROST prepared BF16 tail RoPE for prefill - #1116

Merged

YangXu1990uiuc merged 18 commits into

NVIDIA:developfrom

YangXu1990uiuc:frost/tail-rope-prefill

Sep 22, 2026

Commits on Sep 17, 2026

Add FROST tail RoPE and group32 FP4/FP8 QDQ
YangXu1990uiuc
committed
4016c94
Add FROST prepared BF16 tail RoPE for prefill
YangXu1990uiuc
committed
941ee62

Commits on Sep 22, 2026

[DSv4.1] Refresh RoPE QDQ against develop
YangXu1990uiuc
committed
213fca8
[DSv4.1] Refresh stacked RoPE PR #1116 with updated group32 base
YangXu1990uiuc
committed
4d58128
Merge commit '22f075b2bc356e330ad4c9dbac73826a8a62af57' into refresh-pr1116
YangXu1990uiuc
committed
af84985
[DSv4.1] Preserve benchmark guard and overflow failures
YangXu1990uiuc
committed
22f075b
[DSv4.1] Enforce pinned RoPE source hashes under optimized Python
YangXu1990uiuc
committed
4a6ed00
Keep QDQ API separate from vision RoPE backward
YangXu1990uiuc
committed
7df0585
Merge branch 'refs/heads/refresh-pr1114' into refresh-pr1116
YangXu1990uiuc
committed
837eeac
Point QDQ cache regression at its relocated API module
YangXu1990uiuc
committed
14de284
Merge branch 'refs/heads/refresh-pr1114' into refresh-pr1116
YangXu1990uiuc
committed
b6bc4f7
Merge commit '9e9050c4013eff07f438c903d685cb5cd14f5641' into HEAD
YangXu1990uiuc
committed
cec99ae
Fix tail RoPE allocating wrapper stream resolution
YangXu1990uiuc
committed
776c487
Reject optimized Python in RoPE QDQ benchmark admission
YangXu1990uiuc
committed
d69d09e
Merge commit 'd69d09e7f' into HEAD
YangXu1990uiuc
committed
c8b944e
Merge develop and preserve optional API exports
YangXu1990uiuc
committed
82701b5
Merge remote-tracking branch 'refs/remotes/upstream/develop' into HEAD
YangXu1990uiuc
committed
546b511
Merge landed Vision API and preserve independent lazy exports
YangXu1990uiuc
committed
1fa4825