Repository navigation

All pull requests

Labels (9)Milestones (0)New pull request

Open

(137)

Closed

(621)

fix: bound SwiGLU-OAI tiles to avoid A3 UB overflow
#857PuppyLu opened Oct 13, 2026
1
Support NGRAM reconstruction and tree-aware greedy verification
#854McZyWu opened Oct 13, 2026Contributor
fix: preserve the inclusive left boundary in sink attention
#853PuppyLu opened Oct 13, 2026
2
feat(compressor/arch35): env-gated batch consistency (SGL_DSV4_COMPRE…
#852unclezhou486 opened Oct 13, 2026Contributor
[Perf] Fuse top-k/top-p renormalization with Triton and RN division
#850cl-vv-h 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
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
🐛 [NPU] Size the cache_loc kernel's out_cache_loc view from the tensor
#824TallMessiWu opened Sep 23, 2026Contributor
[DeepEP] Port native INT8 FuseEP mode 2 to Ascend A5
#822McZyWu opened Sep 23, 2026Contributor
Feat/compressor a5 explicit
#821unclezhou486 opened Sep 23, 2026Contributor
[kernels] new fused shared expert kernel for qwen
#817MatsueYu opened Sep 22, 2026Contributor
fix(npu): resolve UB overflow in mamba_state_update_triton for large state dims
#814snyi001 opened Sep 20, 2026
Codex/sync upstream main
#813wanerlee opened Sep 20, 2026

Pull requests