github-featured / lightseekorg--tokenspeed

分支列表

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

ts2/multimem-decode-axis 4ff6a358b1 · perf(k3): keep the multimem tail window out of decode · 最后更新于 2026-08-30 11:22:56 +00:00 frostbyte_neo	151 1	ZIP TAR.GZ
ci/mi450-rocjitsu-bump-2 99adb44f78 · ci(amd): bump rocJITsu to 2ab62717 and raise the MI450 per-test timeout · 最后更新于 2026-08-30 08:38:12 +00:00 frostbyte_neo	152 1	ZIP TAR.GZ
feat/gfx1250-kimi-k3-mxfp4-situ 5b82a3b2ca · feat(moe): add gfx1250 MXFP4 SiTU support · 最后更新于 2026-08-30 08:30:15 +00:00 frostbyte_neo	153 1	ZIP TAR.GZ
jjd/fi_bf16_cutedsl_latency_gemm edcf785819 · feat(gemm): add the ll bf16 decode-gemv backend and its measured tp4 route · 最后更新于 2026-08-30 04:36:41 +00:00 frostbyte_neo	180 4	ZIP TAR.GZ
agent/add-b300-deepswe-ci b4df61a611 · style: format DeepSWE summary helper · 最后更新于 2026-08-30 03:09:55 +00:00 frostbyte_neo	156 2	ZIP TAR.GZ
perf/drafter-gemv-route 186ee0e35c · chore(gemm): log the shapes the decode-GEMV router actually sees · 最后更新于 2026-08-29 21:05:01 +00:00 frostbyte_neo	163 2	ZIP TAR.GZ
fix/dsv4-flashinfer-precomputed-topk e43154902c · refactor(moe): simplify FlashInfer routed dispatch · 最后更新于 2026-08-29 18:24:11 +00:00 frostbyte_neo	158 3	ZIP TAR.GZ
ts2/tail-fusion-token-cap 846699bc85 · perf(k3): end the fused MoE tail at its profit edge, not its capacity · 最后更新于 2026-08-29 16:39:03 +00:00 frostbyte_neo	163 1	ZIP TAR.GZ
bot/slurm-flashinfer-0.6.18 0b98000f49 · Apply contract test formatting · 最后更新于 2026-08-29 13:14:58 +00:00 frostbyte_neo	160 8	ZIP TAR.GZ
feat/kda-tp-phase0 b0516a2501 · refactor(config): split AttnConfig into a model-wide container and component specs · 最后更新于 2026-08-29 11:17:34 +00:00 frostbyte_neo	159 2	ZIP TAR.GZ
ts2/fp8-scale-nocopy 5eba49868e · perf(gemm): keep GB300 on the copy-free MN-major FP8 block-scale scales · 最后更新于 2026-08-28 18:56:10 +00:00 frostbyte_neo	189 1	ZIP TAR.GZ
ts2/moe-topk-pdl 6bfe525f1d · fix(moe): let the Kimi-K3 sigmoid top-k router use PDL · 最后更新于 2026-08-28 17:52:12 +00:00 frostbyte_neo	189 1	ZIP TAR.GZ
bot/kimi-k3-gb300-nightly 3853a86fa5 · ci: run Kimi K3 vision evals nightly · 最后更新于 2026-08-28 09:23:17 +00:00 frostbyte_neo	177 1	ZIP TAR.GZ
bot/tokenspeed-fa4-nvidia-cutlass-dsl-4.7.1 136f8c9d9e · Revert Cutlass DSL resolver bypass · 最后更新于 2026-08-28 07:59:47 +00:00 frostbyte_neo	177 3	ZIP TAR.GZ
perf/dspark-hoist-vocab-gemm 82dbfeba9c · perf(dspark): project the whole proposal block in one GEMM · 最后更新于 2026-08-28 00:25:54 +00:00 frostbyte_neo	186 1	ZIP TAR.GZ
feat/dist-topk d9a5b341d9 · feat(sampling): distributed top-k on the argmax exchange protocol · 最后更新于 2026-08-27 22:28:27 +00:00 frostbyte_neo	186 1	ZIP TAR.GZ
borontion/amd-dsv4-2 0f67860d62 · Merge branch 'main' into borontion/amd-dsv4-2 · 最后更新于 2026-08-27 20:51:32 +00:00 frostbyte_neo	187 15	ZIP TAR.GZ
Max191/glm-5.3-flash-rebase 24ff5ab3df · add import/target guards for amd tests · 最后更新于 2026-08-27 20:11:21 +00:00 frostbyte_neo	188 2	ZIP TAR.GZ
ts2/retire-prepacked-mn d620bc6772 · refactor(gemm): retire the FlashInfer prepacked-MN scale machinery · 最后更新于 2026-08-27 17:49:11 +00:00 frostbyte_neo	189 1	ZIP TAR.GZ
fix/kimi-k3-dp-mla-packing 4189260b5c · Merge branch 'main' into fix/kimi-k3-dp-mla-packing · 最后更新于 2026-08-27 01:10:38 +00:00 frostbyte_neo	197 3	ZIP TAR.GZ

... 4 5 6 7 8 ...

默认分支

分支列表