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

perf/gfx1250-sigmoid-topk 8e8fffde5b · fix(moe): keep mixed-dtype gfx950 sigmoid top-k eligible · 最后更新于 2026-09-09 21:13:28 +00:00 frostbyte_neo	55 3	ZIP TAR.GZ
perf/gfx1250-projected-mla 9e827e828a · fix(ci): update MI450 sim node id for projected MLA decode · 最后更新于 2026-09-09 21:05:06 +00:00 frostbyte_neo	55 3	ZIP TAR.GZ
perf/gfx1250-latent-input-m1 ff22b1ade8 · style: format gfx1250 latent-input decode · 最后更新于 2026-09-09 20:54:07 +00:00 frostbyte_neo	55 2	ZIP TAR.GZ
Max191/glm53-flash-stack-04-dsa-0902 904d43ddff · perf(dsa): reuse slot-conversion kernels across workspace sizes · 最后更新于 2026-09-09 16:33:44 +00:00 frostbyte_neo	58 9	ZIP TAR.GZ
Max191/glm53-flash-stack-03-sparse-kv-0902 b6073e7e48 · perf(cache): reuse MLA prefill writes across row alignment · 最后更新于 2026-09-09 16:33:40 +00:00 frostbyte_neo	58 6	ZIP TAR.GZ
Max191/glm53-flash-stack-02-kda-recurrent-0902 2b18432f4d · perf(kda): reuse prefill convolution across batch sizes · 最后更新于 2026-09-09 16:33:36 +00:00 frostbyte_neo	58 5	ZIP TAR.GZ
zt/gdn_opt f10033ca76 · fix(ci): retry transient readiness connection failures · 最后更新于 2026-09-09 11:55:26 +00:00 frostbyte_neo	60 8	ZIP TAR.GZ
v4-current-best-optimizations b531c7442f · refactor(v4): trim redundant support code · 最后更新于 2026-09-09 09:51:38 +00:00 frostbyte_neo	62 20	ZIP TAR.GZ
refactor/l2-named-streams-no-empty-ops cc0d1c6a5b · refactor(l2): name every stream the executor orders against; refuse cache ops with no transfers · 最后更新于 2026-09-09 07:23:40 +00:00 frostbyte_neo	60 1	ZIP TAR.GZ
perf/deepseek-v4-flash-agentic d359d12390 · test(v4): remove redundant topology flags · 最后更新于 2026-09-09 06:53:25 +00:00 frostbyte_neo	74 4	ZIP TAR.GZ
efric/iris-buffer-sizing-refactor e74767cf13 · fix(comm): preserve zero producer-direct capacity · 最后更新于 2026-09-09 06:07:03 +00:00 frostbyte_neo	60 7	ZIP TAR.GZ
refactor/explicit-cache-op-streams cc3f13c135 · refactor(executor): drop the forward prologue's default-stream wait on the execution stream · 最后更新于 2026-09-09 04:06:51 +00:00 frostbyte_neo	62 3	ZIP TAR.GZ
jjd/fix_mtp_accept_rate 624a98522c · Merge branch 'main' into jjd/fix_mtp_accept_rate · 最后更新于 2026-09-09 02:26:23 +00:00 frostbyte_neo	62 2	ZIP TAR.GZ
lei/gfx950-scaled-downcast 66b1d307a4 · fix(gfx950): preserve zero MX scale payloads · 最后更新于 2026-09-08 16:24:59 +00:00 frostbyte_neo	77 2	ZIP TAR.GZ
jjd/support_fp8_draft ff3c539b0e · feat(quantization): support ModelOpt block-FP8 MTP experts · 最后更新于 2026-09-08 16:22:09 +00:00 frostbyte_neo	74 1	ZIP TAR.GZ
Max191/glm53-flash-stack-01-kpool-0902 8d4381aa1b · perf(kpool): reuse decode metadata kernels across tail batches · 最后更新于 2026-09-08 14:50:34 +00:00 frostbyte_neo	73 5	ZIP TAR.GZ
codex/qsa-verify-state 6461af3b0c · Move QSA verify state ownership to the root attention backend · 最后更新于 2026-09-08 13:17:57 +00:00 frostbyte_neo	74 11	ZIP TAR.GZ
refactor/bind-cache-groups-from-plan 4df638af12 · fix(attention): keep window_left 0 as a one-token window · 最后更新于 2026-09-08 08:48:52 +00:00 frostbyte_neo	74 6	ZIP TAR.GZ
yweng/dev/opt_dflash2 b0905a8880 · Merge branch 'main' into yweng/dev/opt_dflash2 · 最后更新于 2026-09-08 08:05:19 +00:00 frostbyte_neo	74 29	ZIP TAR.GZ
fix/pd-bootstrap-stall 495dcd04d6 · fix(pd): stop failing prefill rooms that are only queued on the decode · 最后更新于 2026-09-08 06:25:17 +00:00 frostbyte_neo	77 1	ZIP TAR.GZ

1 2 3 4 5 ...

默认分支

分支列表