Repository navigation

All pull requests

Labels (9)Milestones (0)New pull request

Open

(139)

Closed

(626)

[Feat] Add block verification mode to chain speculative sampling kernel
#864Estrella-xx opened Oct 16, 2026
[compressor][arch35] shape-invariant K-split / h-stagger (SGL_DSV4_CO…
#863unclezhou486 opened Oct 16, 2026Contributor
[Test] Add isolated target-only acceptance threshold cases
#860JinyanYi opened Oct 15, 2026
[feat]ops transformer integration
#8591329009851 opened Oct 15, 2026Contributor
Add experimental Ascend A2 support for aclnnDispatchFFNCombine
#858MirkoDeVita98 opened Oct 14, 2026
fix: bound SwiGLU-OAI tiles to avoid A3 UB overflow
#857PuppyLu opened Oct 13, 2026
5
[Feat] speculative: support NGRAM trees and paged attention masks
#854McZyWu opened Oct 13, 2026Contributor
fix: preserve the inclusive left boundary in sink attention
#853PuppyLu opened Oct 13, 2026
4
feat(compressor/arch35): env-gated batch consistency (SGL_DSV4_COMPRE…
#852unclezhou486 opened Oct 13, 2026Contributor
[Feat]a5 layered dispatch and combine
#8491329009851 opened Oct 12, 2026Contributor4 tasks
feat: port hc_pre operator for Ascend A3 and A5
#846AndyLi429 opened Oct 12, 2026Contributor
perf: scatter only nonzero speculative probabilities on NPU
#845w1ida opened Oct 11, 20264 of 6 tasks
feat: add probation-aware LRU metadata ops for sparse KV offloading
#844wangbiao0814 opened Oct 10, 2026
[BugFix] [DeepEP] Synchronize A2 dispatch layout scratch reuse
#842INBreezefall opened Oct 10, 20264 tasks done
[Perf] Add four-query batching for ratio-4 LightningIndexer prefill
#839w1ida opened Oct 3, 2026
[Perf] Fuse BF16 gate multiply and output cast for GDN
#838w1ida opened Oct 3, 2026
[Feature] Add Ascend C D256 QSA prefill with exact query sharing
#837w1ida opened Oct 3, 2026
1
Move Attentions kernels to csrc
#836Napkin-AI opened Oct 2, 2026Contributor
8
fix: synchronize FuseEP merge-sort output before extraction
#835MirkoDeVita98 opened Sep 30, 2026
[BugFix] Bound L2Norm tile size to avoid UB overflow
#833w1ida opened Sep 30, 2026
1
[BugFix] Avoid L2Norm recompilation for runtime sequence sizes
#832w1ida opened Sep 30, 2026
1
improve move cache and roll back for kda low concurrency
#831zhaozx-cn opened Sep 29, 2026Contributor
[BugFix] Avoid alloc_extend recompilation for runtime token counts
#827ltaodream opened Sep 25, 2026
[DeepEP] Port native INT8 FuseEP mode 2 to Ascend A5
#822McZyWu opened Sep 23, 2026Contributor
Feat/compressor a5 explicit
#821unclezhou486 opened Sep 23, 2026Contributor

Pull requests