Repository navigation

All pull requests

Labels (9)Milestones (0)New pull request

feat(qy): QY ops backend with FA and longctx MoE/RoPE kernels
#1002Ceng23333 opened Oct 14, 2026Contributor
feat: add optimized Kunlun rotary embedding backend
#1001JoeZhang-0x000 opened Oct 10, 20264 tasks done
perf(ascend): accelerate packed Q/K and strided KV/SwiGLU kernels
#1000JoeZhang-0x000 opened Oct 9, 20265 of 8 tasks
perf(metax): reduce RMSNorm blocks for small head dimensions
#999JoeZhang-0x000 opened Oct 9, 2026
perf(metax): vectorize small fused add RMSNorm rows
#998JoeZhang-0x000 opened Oct 9, 2026
perf(metax): vectorize contiguous embedding lookups
#997JoeZhang-0x000 opened Oct 9, 2026
feat: add LightningAttentionInfinilm for MiniMax linear attention
#995y258dd opened Sep 19, 20265 of 19 tasks
feat(metax): add moe_topk_softmax operator for MoE top-k gating
#993rubik-hua opened Sep 18, 2026
feat(metax): add moe_fused_dense operator for fused MoE expert forward
#992rubik-hua opened Sep 18, 2026
feat(metax): add moe_align operator for MoE token alignment
#991rubik-hua opened Sep 18, 2026
fix(ascend): detect exact SoC version
#986baominghelly opened Sep 14, 2026Contributor3 of 18 tasks
feat(moore): support flash-attn via MooreThreads/mate v0.2.5
#983spike-zhu opened Sep 9, 2026Contributor2 of 19 tasks
feat(cambricon): add convolution provider
#967baominghelly opened Sep 2, 2026Contributor3 of 19 tasks
feat(cambricon): add topk softmax provider
#958baominghelly opened Aug 31, 2026Contributor2 of 19 tasks
fix: preserve implementation lookups under optimization
#929voltjia opened Aug 13, 2026Collaborator9 of 19 tasks
perf: optimize cache for tensor conversion
#928DockeyMint opened Aug 13, 20263 of 19 tasks
feat(nvidia): add top-k top-p sampling provider
#922voltjia opened Aug 11, 2026Collaborator3 of 19 tasks
refactor: simplify operator cache fast path
#920voltjia opened Aug 11, 2026Collaborator10 of 19 tasks
feat(nvidia): add Argmax provider
#918voltjia opened Aug 10, 2026Collaborator3 of 19 tasks
fix(rms_norm): support optional weight
#916voltjia opened Aug 9, 2026Collaborator9 of 19 tasks
feat(nvidia): link vLLM fused_marlin_moe
#914voltjia opened Aug 8, 2026Collaborator5 of 19 tasks
fix(linked): support multiple operator schemas
#913voltjia opened Aug 8, 2026Collaborator5 of 19 tasks
fix: support optional C in Gemm
#870voltjia opened Aug 4, 2026Collaborator11 of 20 tasks
perf: avoid redundant operator cache lookups
#858baominghelly opened Jul 30, 2026Contributor10 of 19 tasks
perf(torch): reuse ATen wrappers in generated operators
#816voltjia opened Jul 24, 2026Collaborator11 of 19 tasks