分支列表

默认分支

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

分支列表

01589f33f9 · perf(moe): reduce Kimi K3 grouped decode padding · 最后更新于 2026-08-24 19:01:30 +00:00    frostbyte_neo

242
6

2e4014ee41 · feat(kda): add AMD fused MTP verification · 最后更新于 2026-08-24 13:23:14 +00:00    frostbyte_neo

226
1

f7d22cd9b5 · feat(mla): serve a block drafter from the CuteDSL MLA backend · 最后更新于 2026-08-24 02:49:18 +00:00    frostbyte_neo

226
1

64a2b0ea3e · fix(metadata): stage a tape's registers through a per-call pinned block · 最后更新于 2026-08-23 19:06:05 +00:00    frostbyte_neo

231
2

9d258bf8c7 · Merge branch 'main' into hongtaoc/attnres-no-t-limit-103120 · 最后更新于 2026-08-23 16:51:27 +00:00    frostbyte_neo

230
3

5de4bd6980 · test(kimi-k3): dedup first turns and guard shrunken samples in pd_sim · 最后更新于 2026-08-23 14:23:46 +00:00    frostbyte_neo

232
2

94604367aa · perf(gfx1250): fill the WMMA tile in the KDA state scan · 最后更新于 2026-08-23 10:36:08 +00:00    frostbyte_neo

232
4

9ba79c95a2 · fix(dsv4): reset stale decode plans · 最后更新于 2026-08-23 09:47:30 +00:00    frostbyte_neo

239
42

df3183f0da · feat(runtime): prefill chunk pipeline + event-loop cohesion refactor · 最后更新于 2026-08-23 09:26:35 +00:00    frostbyte_neo

233
1

263bcfc1d7 · perf(kda): size the batched commit launches from the grid they actually have · 最后更新于 2026-08-23 09:16:11 +00:00    frostbyte_neo

238
1

4d868c3492 · docs(k3): document split shared shard contract · 最后更新于 2026-08-23 07:50:00 +00:00    frostbyte_neo

234
5

35594c74f2 · Merge branch 'main' into perf/k3-kda-payload-copy · 最后更新于 2026-08-23 01:55:29 +00:00    frostbyte_neo

239
2

7ff10897c8 · Merge branch 'main' into ts2/kda-triton-tuning · 最后更新于 2026-08-23 01:41:26 +00:00    frostbyte_neo

239
2

8cfe3d7984 · ci(k3): give the Slurm aime26 gates the budget their context window allows · 最后更新于 2026-08-22 23:50:23 +00:00    frostbyte_neo

246
1

efcadf1adc · fix(cli): disable smg load monitoring by default · 最后更新于 2026-08-22 16:48:36 +00:00    frostbyte_neo

244
1

ef0cf55ac4 · style(runtime): format scheduler communicator · 最后更新于 2026-08-22 16:48:01 +00:00    frostbyte_neo

244
2

4b9766f6cc · chore: merge main and drop CI test changes · 最后更新于 2026-08-22 09:32:38 +00:00    frostbyte_neo

246
12

6d83175f85 · fix(gemm): admit sm100 to the measured decode route · 最后更新于 2026-08-22 04:22:40 +00:00    frostbyte_neo

250
5

627f6474bb · refactor(attn): import Triton through the vendored facade · 最后更新于 2026-08-22 02:07:35 +00:00    frostbyte_neo

250
3

29047cd9fa · perf(kda): strong-form V-major replay tiles, thin replay fold, launch params · 最后更新于 2026-08-21 19:43:39 +00:00    frostbyte_neo

256
2