github-featured / lightseekorg--tokenspeed

分支列表

1,138 提交 264 分支 1 Git 标签

默认分支

main

0ba6b6aab7 · perf(mla): add opt-in packed Q decode for FP8 and FP16 (#1512) · 最后更新于 2026-09-12 09:04:01 +00:00
 ZIP  TAR.GZ

分支列表

perf/dist-argmax-cross-node

f2766637b2 · perf(sampling): probe NVLS for the distributed argmax and draft through it · 最后更新于 2026-08-26 22:09:47 +00:00    frostbyte_neo

205
1
 ZIP  TAR.GZ
bot/update-tokenspeed-smg-1.10.0.post20260826

51af38300a · Merge branch 'main' into bot/update-tokenspeed-smg-1.10.0.post20260826 · 最后更新于 2026-08-26 20:09:13 +00:00    frostbyte_neo

197
2
 ZIP  TAR.GZ
ts2/mnnvl-plain-ar

97e6b30dca · Merge branch 'main' into ts2/mnnvl-plain-ar · 最后更新于 2026-08-26 20:08:51 +00:00    frostbyte_neo

197
2
 ZIP  TAR.GZ
hongtaoc/k3-flashinfer-self-routing-situ

766dacd39a · fix(kimi-k3): align routing mode across DP ranks · 最后更新于 2026-08-26 16:51:49 +00:00    frostbyte_neo

201
3
 ZIP  TAR.GZ
zt/qwen3.8_next

efb223fac2 · fix(qwen4-exp): follow the upstream pdl_enabled move · 最后更新于 2026-08-26 13:20:18 +00:00    frostbyte_neo

202
3
 ZIP  TAR.GZ
xuchun/perf-glm

7991f9243c · fix(glm5): preserve AMD DSA prefill metadata support · 最后更新于 2026-08-26 11:14:28 +00:00    frostbyte_neo

204
3
 ZIP  TAR.GZ
ci/mi450-simulator

870579cd1e · fix(amd): harden MI450 simulator coverage · 最后更新于 2026-08-26 09:57:26 +00:00    frostbyte_neo

216
14
 ZIP  TAR.GZ
feat/dp-rank-affinity

b8d42ff536 · Merge branch 'main' into feat/dp-rank-affinity · 最后更新于 2026-08-26 04:24:28 +00:00    frostbyte_neo

205
3
 ZIP  TAR.GZ
ts2/fp8-blockscale-kmajor

d95c99d340 · perf(gemm): pass native K-major scales to the FlashInfer FP8 block-scale GEMM · 最后更新于 2026-08-25 23:52:54 +00:00    frostbyte_neo

215
1
 ZIP  TAR.GZ
bot/update-tokenspeed-smg-1.10.0.post20260825

e3bfdb7d4a · Update tokenspeed-smg* pins to 1.10.0.post20260825 / 0.4.15.post20260825 / 0.9.0.post20260825 · 最后更新于 2026-08-25 22:04:31 +00:00    frostbyte_neo

211
1
 ZIP  TAR.GZ
feat/gfx950-kda-fused-replay-cleanup

90479173ed · perf(kda): specialize gfx950 verify and replay · 最后更新于 2026-08-25 21:29:56 +00:00    frostbyte_neo

212
1
 ZIP  TAR.GZ
borontion/fix-pdl

4b14308860 · Merge branch 'main' into borontion/fix-pdl · 最后更新于 2026-08-25 20:52:42 +00:00    frostbyte_neo

212
17
 ZIP  TAR.GZ
refactor/device-handle-boundary

3bc020f2b4 · perf(pd): stop draining the PP chunk pipeline for every finished prefill · 最后更新于 2026-08-25 15:56:30 +00:00    frostbyte_neo

216
12
 ZIP  TAR.GZ
hongtaoc/chore-kimi-k3-agentic-tokenspeed-mla

ca9bc93b06 · chore(kimi-k3): use TokenSpeed MLA for agentic drafter · 最后更新于 2026-08-25 13:30:47 +00:00    frostbyte_neo

216
1
 ZIP  TAR.GZ
wb-deliverable

e202a9290c · Merge branch 'main' into wb-deliverable · 最后更新于 2026-08-25 04:56:36 +00:00    frostbyte_neo

216
5
 ZIP  TAR.GZ
bot/kimi-k3-beam-1m-ci

531fc45449 · ci(k3): add Beam 1M verifier on GB300 · 最后更新于 2026-08-25 01:26:45 +00:00    frostbyte_neo

224
1
 ZIP  TAR.GZ
ts2/kda-recover-ssm

1a37a6fb52 · perf(kda): commit from verify's cached corrections instead of replaying · 最后更新于 2026-08-25 00:50:45 +00:00    frostbyte_neo

229
2
 ZIP  TAR.GZ
ts2/kda-cached-producers

c9ed891db3 · perf(kda): hand replay the gate that verify already computed · 最后更新于 2026-08-25 00:49:56 +00:00    frostbyte_neo

229
1
 ZIP  TAR.GZ
perf/mnnvl-oneshot-traffic-dispatch

2749dfb971 · Merge branch 'main' into perf/mnnvl-oneshot-traffic-dispatch · 最后更新于 2026-08-24 23:50:52 +00:00    frostbyte_neo

221
2
 ZIP  TAR.GZ
bot/kimi-k3-mxfp4-deepswe-ci

e0695ee66a · Merge remote-tracking branch 'origin/main' into bot/kimi-k3-mxfp4-deepswe-ci · 最后更新于 2026-08-24 20:36:57 +00:00    frostbyte_neo

224
2
 ZIP  TAR.GZ