Fix sparse MLA cache layout validation and eight-head decode - #3

Closed

ormandj wants to merge 9 commits into

lucifer1004:sm120-sparse-mla-runtime-row-stridefrom

ormandj:sparse-mla-layout-validation

Commits on Sep 10, 2026

feat(sm120): DeepSeek-V4.1 (DSV4_1) sparse MLA cache format support
lucifer1004
committed
2c71fb6
feat(sm120): pair-folded XV for 32-wide-group decode (DSV4_1 8-warp tile)
lucifer1004
committed
00e7865
Carry GLM eight-head decode fix and regression coverage
ormandj
committed
0795228
Validate sparse MLA cache origins and physical block strides
ormandj
committed
295ac64
Test flat GLM decode caches with page gaps
ormandj
committed
2f199af
Cover distinct 528-byte formats in refreshed sparse MLA tests
ormandj
committed
50e7d04
Test sparse MLA scale groups across decode chunk boundaries
ormandj
committed
5283c02
Invalidate sparse MLA calibration after canonical payload change
ormandj
committed
8dc1812
Cover footer cache rejection and avoid rereading stale tuning
ormandj
committed
e7cad40