Skip to content

Pull requests: lightseekorg/tokenspeed

New pull request New
24 Open 1,549 Closed
24 Open 1,549 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

docs: clarify tokenspeed-scheduler release workflow
#1653 opened Sep 20, 2026 by lightseek-bot Contributor Loading…
feat: support tp shard on MoE shared expert
#1652 opened Sep 20, 2026 by byshiue Collaborator Loading…
perf(moe): fuse FlashInfer NVFP4 routing-map padding initialization
#1644 opened Sep 19, 2026 by byshiue Collaborator Loading…
3
perf(amd): give gfx1250 MXFP4 decode's spare warp bit to N
#1641 opened Sep 18, 2026 by jerryyin Member Loading…
2
refactor(kernel): unify PyTorch references under numerics/reference
#1626 opened Sep 17, 2026 by antiagainst Member • Draft
1
[Don't Merge][Draft] feat(kimi-k3) DEP + TP o projection on attention
#1610 opened Sep 17, 2026 by byshiue Collaborator Loading…
fix(runtime): aggregate cache flush results across data parallel schedulers
#1608 opened Sep 17, 2026 by yechank-nvidia Collaborator Loading…
@yechank-nvidia
1
perf(sampling): skip unused fallback-index reduction in target-only v…
#1607 opened Sep 17, 2026 by yechank-nvidia Collaborator Loading…
@yechank-nvidia
1
perf(moe): pass unpacked routes to FlashInfer
#1606 opened Sep 17, 2026 by yechank-nvidia Collaborator Loading…
@yechank-nvidia
1
fix(runtime): preserve pending drains on empty memory resume
#1605 opened Sep 17, 2026 by yechank-nvidia Collaborator Loading…
@yechank-nvidia
1
[Speculative Decoding] add synthetic acceptance length
#1601 opened Sep 16, 2026 by edwingao28 Loading…
4
fix(cache): allow uneven ordinary cache groups
#1589 opened Sep 16, 2026 by yechank-nvidia Collaborator Loading…
@yechank-nvidia
3
perf(kpool): overlap prefill pooled-cache writes with indexer query preparation
#1588 opened Sep 16, 2026 by Dovis01 Contributor Loading…
@FlamingoPg
3
feat(scheduler): keep ordinary state chunks request-local
#1585 opened Sep 16, 2026 by dongjiyingdjy Contributor Loading…
1
Fuse the K3 MoE tail across medium and large token ranges
#1575 opened Sep 15, 2026 by Njuapp Contributor • Draft
5
Add Kimi K3 AgentX Slurm benchmark harness
#1564 opened Sep 15, 2026 by Xiangyi1996 Collaborator • Draft
27
perf(mla): tokenspeed_mla GB200 decode optimizations (-14.6%)
#1525 opened Sep 12, 2026 by Dogacel Collaborator Loading…
12
feat(memory): reserve the CUDA-graph pools in the KV cache budget
#1509 opened Sep 12, 2026 by nperrin-fr Collaborator • Draft
feat(sampling): add the sonic-sampler backend
#1506 opened Sep 11, 2026 by nperrin-fr Collaborator • Draft
Integrate petit megamoe
#1475 opened Sep 10, 2026 by Marcozhumh Loading…
37
perf(autotune): tune exact decode shapes and refresh GEMM routes
#1421 opened Sep 7, 2026 by chenht2022 Contributor Loading…
30
feat(ple): support qwen3.8-flash-next PLE cpu offload
#1362 opened Sep 2, 2026 by minedec Contributor • Draft
perf(prefill-graph): share bucket outputs and size the ceiling to a chunk
#1350 opened Sep 1, 2026 by nperrin-fr Collaborator • Draft
11
[WIP] Try new Triton package
#631 opened Jul 10, 2026 by antiagainst Member Loading…
19
ProTip! Add no:assignee to see everything that’s not assigned.