Skip to content

Pull requests: lightseekorg/tokenspeed

New pull request New
33 Open 857 Closed
33 Open 857 Closed
Author
Filter by author
Loading
Label
Filter by label
Loading
Use alt + click/return to exclude labels
or ⇧ + click/return for logical OR
Projects
Filter by project
Loading
Milestones
Filter by milestone
Loading
Reviews
Filter by reviews
No reviews Review required Approved review Changes requested
Assignee
Filter by who’s assigned
Assigned to nobody Loading
Sort
Sort by
Newest Oldest Most commented Least commented Recently updated Least recently updated Best match
Most reactions
👍 👎 😄 🎉 😕 ❤️ 🚀 👀

Pull requests list

perf(kimi-k3): router cublas dispatch, grouped MoE-join reduce, FP8 NoPE prefill
#926 opened Aug 4, 2026 by dongjiyingdjy Contributor Loading…
7
fix(comm): gate MNNVL buffers on a real fabric-allocation probe
#925 opened Aug 4, 2026 by tuanzhangCS Contributor Loading…
5
feat(dspark): complete Kimi K3 draft execution
#924 opened Aug 4, 2026 by HarryWu99 Contributor Loading…
4
perf(kimi3): fuse semantic MLA decode stages
#919 opened Aug 4, 2026 by qedawkins Contributor Loading…
4
feat(amd): GFX950 Gluon KDA prefill
#918 opened Aug 3, 2026 by Yu-Zhewen Contributor Loading…
11
[WIP] feat(moe): MORI all-to-all expert-parallel MoE backend (AMD)
#915 opened Aug 3, 2026 by nirvedhmeshram Loading…
12
refactor: rewrite triton moe kernel
#914 opened Aug 3, 2026 by borontion Contributor Loading…
19
feat(dspark): Kimi-K3 DSpark speculative decoding on FlatKV
#905 opened Aug 3, 2026 by jhinpan Loading…
5 of 6 tasks
1
19
refactor(runtime): make ForwardContext tensor-free
#891 opened Aug 1, 2026 by rjzhb Collaborator Loading…
5 tasks done
4
(perf): Optimize Kimi K3 MoE input projections and low-token decode
#878 opened Jul 31, 2026 by panditsa Contributor Loading…
14
feat: add Qwen2.5-MoE model support
#866 opened Jul 31, 2026 by potoior Loading…
9
Support detailed token logprobs for RL and distillation
#861 opened Jul 31, 2026 by HJSang Collaborator Loading…
10
[WIP] perf(kda): CuTe DSL fused decode for KDA linear attention
#857 opened Jul 30, 2026 by nperrin-fr Collaborator • Draft
1
fix(runtime): reject ambiguous prompt sources
#854 opened Jul 30, 2026 by ShiroKSH Loading…
fix(spec): revert dflash-specific branches in speculative runtime
#853 opened Jul 30, 2026 by minedec Contributor • Draft
perf(kimi-k3): fuse AMD native MoE up projection epilogue
#841 opened Jul 29, 2026 by qedawkins Contributor Loading…
1
perf: fuse MiniMax sparse cache insertion
#839 opened Jul 29, 2026 by FlamingoPg Contributor • Draft
[WIP] perf(kimi3): retile warp decode
#831 opened Jul 28, 2026 by panditsa Contributor Loading…
3
feat(qwen3.5): support text-only qwen3.5 checkpoint config
#828 opened Jul 28, 2026 by minedec Contributor • Draft
fix(runtime): keep multimodal tensors inline across nodes
#826 opened Jul 28, 2026 by tuanzhangCS Contributor • Draft
2
feat: support torchspec training
#798 opened Jul 25, 2026 by Dogacel Contributor Loading…
4
feat(kernel): Add validation per family/mode for kernel registration
#769 opened Jul 22, 2026 by Max191 Contributor Loading…
12
feat(moe): R3 rollout routing replay — slot-indexed expert-routing pool (scaffold) inactive
#737 opened Jul 20, 2026 by qywu Collaborator • Draft
2
feat(engine): return_token_ids — echo raw prompt/output token ids (RL anti-drift) inactive
#736 opened Jul 20, 2026 by qywu Collaborator • Draft
1
feat(scheduler): per-adapter KV prefix-cache namespace + max_loras batch cap
#735 opened Jul 19, 2026 by qywu Collaborator Loading…
ProTip! Filter pull requests by the default branch with base:main.