onnxruntime
[CUDA] PagedAttention: add SM<80 fp16 fallback via memory-efficient attention
#28200

Merged

[CUDA] PagedAttention: add SM<80 fp16 fallback via memory-efficient attention #28200

tianleiwu merged 10 commits into microsoft:main from elwhyjay:feature/paged-attention-mea-fallback

[CUDA] PagedAttention: add GatherAndExpandPagedKVCache kernel for MEA…

566b1571

[CUDA] PagedAttention: add UnfusedAttention MEA fallback path

3e0dda62

[CUDA] PagedAttention: populate cumulative_seqlens_kv in UnfusedAtten…

23aa6eca

[CUDA] PagedAttention: wire MEA fallback dispatch + scratch allocation

f1338c78

[CUDA] PagedAttention: add MEA fallback test coverage

dad6e94e

[CUDA] PagedAttention: declare LaunchGetCumulativeSeqlensKV in impl h…

df076c88

[CUDA] PagedAttention: fix OrtStreamAdapter access in ComputeInternal

b815206c

tianleiwu requested a review from

copilot-pull-request-reviewer 104 days ago

copilot-pull-request-reviewer commented on 2026-04-23

[CUDA] PagedAttention: address Copilot review on PR #28200

93aff525

tianleiwu requested changes on 2026-04-24

[CUDA] PagedAttention: fix int32 overflow + heuristic max_query_len (…

f3458655

tianleiwu dismissed these changes on 2026-04-24

[CUDA] PagedAttention: early-return on empty query input (token_count…

7375578b

elwhyjay dismissed their stale review via 7375578b 103 days ago

tianleiwu approved these changes on 2026-04-28

tianleiwu merged 45f5aba5 into main 99 days ago

Reviewers

tianleiwu

Assignees

No one assigned

Labels

None yet

Milestone

No milestone