llama.cpp
ggml-metal: add chunked SSD MMA for Mamba-2 prefill optimization
#26647
Merged
Merged
ggml-metal: add chunked SSD MMA for Mamba-2 prefill optimization #26647
ggerganov
merged
15 commits into
ggml-org:master
from
dpantaleoni:ssd-metal
metal: WIP chunked SSD SSM_SCAN kernels for multi-token prefill
c375a5a2
metal: drop scalar SSD path; MMA + sequential tail
778f64cd
drop WIP ssm scan test noise
6edd9bb5
remove state_from_dst and rename CS and NSG constants
895c55ec
remove unrelated added whitespace padding
b978c8ff
added clarity to mma_tokens calculation
a4725e4f
added clarity to use_mma bool checks
15668be5
added comments to metal ssd op constants for clarity
5b2d1521
reserve K tokens for sequential kernel rollback snapshots
4820ff8f
reset concurrency between mma and seq tail
559a766f
remove print args no longer used
61922f6a
fixed comment to no longer point to specific line
0c4872d2
add FC_SSM_SCAN so seq path skips token offlset unless it's mma tail
0bc1e226
added changes to new ssm.metal for rebase after ggml-metal.metal refa…
f19941d2
specialize ssm_scan tail with a template instead of a function constant
3954a13c
forforever73
approved these changes
on 2026-08-26
ggerganov
approved these changes
on 2026-08-26
Reviewers
Assignees
Labels
testing
ggml
merge ready
Apple Metal
Milestone
No milestone
Login to write a write a comment.
Login via GitHub