llama.cpp
Initial changes for Recurrent state rollback for nemotron for cuda
#26623
Merged
Merged
Initial changes for Recurrent state rollback for nemotron for cuda #26623
ggerganov
merged
17 commits into
ggml-org:master
from
lnigam:lnigam/recurrent-state-rollback
lnigam
changed the title 
ggerganov
commented
on 2026-08-07
am17an
commented
on 2026-08-10
ggerganov
commented
on 2026-08-10
ggerganov
commented
on 2026-08-11
ggerganov
commented
on 2026-08-14
Initial changes for Recurrent state rollback for nemotron for cpu and…
3953c27f
Removing CPU RS rollback. Will enable it in subsequent PRs
80492a00
addition of test case
18936761
Removing assert and calling runtime API to check if op is supported
6f0bbe86
removing extra API and updating the call sites for K
9f55d1ea
replace static cuda detection to runtime fused_op api
3c5cf687
address review comments and fallback when SSM rollback not supprted
169ca77e
Adding changes for supporting RS-rollback in CPU. Also added test-bac…
b81a776f
removing memory manipulation as rs rollback is now supported in CPU
afd1bb2d
removing the static probe which is not needed now
037104a2
correcting the format
19c1fb36
address review comments
5ef2209d
enabling test for all the backends, unsupported backends will fallbac…
f7de7039
Apply suggestions from code review
d1c9fb02
choose different graph based on the result of fused_ssm_op is support…
25cc903b
Support K > 1 in ssm_scan for all backends
b749728c
Fix CI Issues
a078c8a9
ggerganov
approved these changes
on 2026-08-14
gaugarg-nv
approved these changes
on 2026-08-14
Reviewers
Assignees
Labels
model
testing
Vulkan
ggml
merge ready
SYCL
Apple Metal
CUDA
WebGPU
Milestone
No milestone
Login to write a write a comment.
Login via GitHub