onnxruntime
[CUDA] Extend DynamicSparseAttention INT8 KV cache and PackedSparseAttentionIndexer QSA
#33472
Merged
Merged
[CUDA] Extend DynamicSparseAttention INT8 KV cache and PackedSparseAttentionIndexer QSA #33472
apsonawane
merged
26 commits into
main
from
asonawane/main-ops-sparse-attention-indexer
Extend sparse attention caches and optimize packed indexing
52c73b51
Align QSA position overrides and CUDA-only outputs across providers
7be8d682
Update generated QSA position override documentation
b538d1fc
Fix sparse attention workspace and optional interface handling
538bd982
copilot-pull-request-reviewer
commented
on 2026-10-19
Fix WebGPU binding-limit test and clarify mixed key-state support
93565558
Add compact QSA selection and specialized INT8 attention dispatch
d21946e6
Merge main into asonawane/main-ops-sparse-attention-indexer
b42576e6
Apply lintrunner formatting to PR changes
b71498b2
Enable default-on sparse attention optimizations, S<=8 grouped decode…
c0c4e7dd
[CUDA] Fold #33546 and S=4 prefix-state outputs (mid_kv_buffer2/3, mi…
6548c21c
Merge remote-tracking branch 'origin/main' into tlwu/pr33472-updates
58b6da62
github-actions
commented
on 2026-10-23
Apply lintrunner formatting
3b52694c
[CUDA] Revert unused QSA position_ids and WebGPU shader changes in Pa…
5ed039c7
tianleiwu
changed the title
github-actions
commented
on 2026-10-24
fix(cuda): fix shared-provider include order, Windows macro clash, te…
6511e8a3
Merge remote-tracking branch 'origin/main' into pr33472-work
a9afd694
Optimize GroupedDecodeDynamicSparseAttentionKernel for 12 query heads…
a523af45
Remove legacy mid-state snapshot outputs from PackedSparseAttentionIn…
61c5d87f
perf(cuda): optimize INT8 KV DynamicSparseAttention prefill and Packe…
aa787761
fix(cuda): fix FP16 ratio-12 grouped decode MMA and expand prefill te…
18a13c70
fix(cuda): avoid narrowing in sparse attention test
fea46be4
[CUDA] Support per-batch committed_length in DynamicSparseAttention (…
9116cb9d
apsonawane
commented
on 2026-10-26
apsonawane
commented
on 2026-10-26
[CUDA] DSA: fail strict validation on inconsistent committed cursor
a4151cc4
test(cuda): run alternate DSA/indexer switches in fresh processes
977e9813
docs: document DSA/indexer dispatch switches, INT8 encoding and statu…
0639efa7
[CUDA] Remove packed sparse indexer radix top-K path
cd4e2e42
kunal-vaishnavi
commented
on 2026-10-27
fix(cuda): guard DSA wide loads on buffer alignment and fix iOS test …
c8c7b77c
jiafatom
approved these changes
on 2026-10-28
Assignees
No one assigned
Labels
model:qwen
Milestone
No milestone
Login to write a write a comment.
Login via GitHub