Refactor Sparse MLA SM120 - #4802

Merged

bkryu merged 57 commits into

flashinfer-ai:mainfrom

lucifer1004:sm120-sparse-mla-decode-consolidated

Sep 3, 2026

Commits on Aug 28, 2026

fix(sm120): return empty outputs for zero-token sparse MLA decode

lucifer1004
and
XingSong
committed
71923b4
fix(sm120): explain sparse-MLA decode dispatch misses; add config query API

lucifer1004
and
SamMausberg
committed
d22075d
feat(sm120): accept row-strided indices in sparse-MLA decode kernels
lucifer1004
committed
1bc82e5
perf(sm120): replace per-shape cpb autotuning with a calibrated analytical model
lucifer1004
committed
12c4acb
perf(mla): add warp-specialized swapAB SM120 sparse-MLA prefill kernel

Lemon7-UP
and
lucifer1004
committed
9c35030
fix(sm120): relax sparse-MLA prefill num_tokens assert to >= 1
lucifer1004
committed
b7ddf6a
feat(sm120): add per-call prefill_impl override for the swapAB dispatch
lucifer1004
committed
132133f
feat(sm120): calibrated decode/prefill crossover dispatch
lucifer1004
committed
18ca95a
refactor(sm120): plan-layer dispatch with single-source envelopes
lucifer1004
committed
d7d5ddb
refactor(sm120): split prefill kernels by family
lucifer1004
committed
4ad7390
feat(mla): support native NoPE sparse MLA on SM120

lucamotz
and
lucifer1004
committed
4dad92e
feat(sm120): extend GLM53_NOPE to H=64 decode, swapAB prefill, cpb coverage
lucifer1004
committed
4d9b44e
feat(sparse-mla-sm120): add DOTS3_SWA model type (decode + prefill + runner)

lucifer1004
and
claude
committed
dc57682
feat(sm120): cpb model coverage for dots3_swa
lucifer1004
committed
87caf91
feat(sm120): export sparse-MLA runner as SparseMLASm120Wrapper
lucifer1004
committed
bc60731
perf(sm120): single-pass vectorized quantize_q_to_smem
lucifer1004
committed
a348fe3
perf(sm120): single-read candidate indices in decode-dsv4 IO gather
lucifer1004
committed
2d28a34
feat(sm120): runtime-NUM_HEADS decode instantiation, any H in [1, 128]
lucifer1004
committed
a0330c2
test(sm120): runtime-H decode coverage — arbitrary-H GPU cases + dispatch semantics
lucifer1004
committed
99b043a
perf(sm120): keep dedicated H in {8,16,32,64,128} decode instantiations, runtime-H as fallback
lucifer1004
committed
397b632

Commits on Aug 29, 2026

Commits on Sep 1, 2026

Commits on Sep 2, 2026

Commits on Sep 3, 2026

Merge branch 'main' into sm120-sparse-mla-decode-consolidated
bkryu
authored
ffa47d5