ark: add BF16/FP16 Sparge variants #2145
yiliu30
wants to merge
28 commits into
main
from
feat/sparse-bf16-prefill-v2
test: update sparse bf16 validation status
2d425b6c
feat: add bf16 sparse mode to flux example
ddaad538
feat: add bf16 sparse kernel sources
2e2ecb54
chore: ignore XPU build artifact directories
0850714f
feat: wire native BF16 sparse SDPA kernel and launchers
95f2c060
fix: BF16 sparse native kernel element-aware MMA and softmax scale
f96ec416
fix: sparse BF16 LUT crash, lossless round-trip, and preprocess speed
f9e86122
feat: BF16 sparse SDPA benchmark support and repro script
c94a1b80
test: BF16 sparse preprocess and LUT regression tests
e34327a6
feat: include ark-xbuild in wan sparse binding search roots
9c86dda6
refactor: remove dense-mask fallback from sage_sparse_bf16
743cece1
docs: add documented BF16 sparse benchmark runner script
c6bd14a5
fix: work around triton SPV_INTEL_predicated_io on Intel XPU
de2eca03
docs: default BF16 benchmark runner to triton preprocess backend
31fad6e3
feat: add independent native-precision sparse SDPA kernel path
2e635bc1
feat: sparse-SDPA Python wrappers, switch sage_sparse_bf16, benchmark…
2127635a
test: sparse-SDPA correctness tests; fix XPU softmax reference
952ef649
refactor: separate native-precision sparse SDPA from INT8 sparse SAGE…
1da474b3
docs: refresh sparse BF16 handoff after int8/bf16 path separation
333e2426
chore: add bf16-sparse FLUX.1-dev output image reference
e39a2acb
![pre-commit-ci[bot]](https://avatars.githubusercontent.com/in/68672?v=4&s=60)
[pre-commit.ci] auto fixes from pre-commit.com hooks
18ef86fd
clean
0ca035fb
chore: drop dev handoff and generated images from bf16 sparse branch
a559dc82
Merge branch 'feat/sparse-bf16-prefill-v2' of https://github.com/inte…
97ce5c7d
fix: use k32 sparse sdpa tiles
cecc8a1b
bench: add ark sdpa baseline
79bef558
tools: add sparse bf16 sweep helpers
ee565a16
docs: record sparse bf16 profiling results
9b1a9512
Reviewers
No reviews
Assignees
No one assigned
Labels
None yet
Milestone
No milestone
Login to write a write a comment.
Login via GitHub