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

f159e83903 · docs(scheduler): correct reclaim-floor comment, note admission alignment coupling · 最后更新于 2026-07-22 02:42:52 +00:00    frostbyte_neo

584
7

909a81835f · feat(flat): add sharded Mamba state paging for Qwen3.5 MTP · 最后更新于 2026-07-21 07:42:01 +00:00    frostbyte_neo

655
2

3bb827df5a · test(rl): cover weight version lifecycle · 最后更新于 2026-07-21 03:55:45 +00:00    frostbyte_neo

593
3

e47363dfb8 · perf(gdn): update to gdn decode API to flashinfer 0.6.15 · 最后更新于 2026-07-20 07:16:41 +00:00    frostbyte_neo

597
3

239ac75e5c · feat(moe): wire R3 routing capture into the model forward (guarded) · 最后更新于 2026-07-20 01:05:06 +00:00    frostbyte_neo

598
2

65bb02f4d9 · feat(engine): add return_token_ids to echo raw prompt/output token ids · 最后更新于 2026-07-20 00:36:04 +00:00    frostbyte_neo

598
1

6e9d8290b8 · Merge branch 'main' into perf/deepseek-v4-prefill-graph-dsa-writes · 最后更新于 2026-07-18 14:59:01 +00:00    frostbyte_neo

609
2

9caffdf179 · refactor(kernel): migrate GDN Triton kernels to tensor descriptors · 最后更新于 2026-07-18 07:51:38 +00:00    frostbyte_neo

611
1

7faa322d5b · fix(runtime): stabilize Qwen3.5 MoE routing under MTP · 最后更新于 2026-07-18 06:49:35 +00:00    frostbyte_neo

611
2

965bf28434 · fix(multimodal): remove stale model imports · 最后更新于 2026-07-15 03:53:13 +00:00    frostbyte_neo

635
10

a59addeb9f · fix(lint): clang-format kv event C++ and skip vendored xxhash · 最后更新于 2026-07-14 17:54:27 +00:00    frostbyte_neo

642
25

4df591e189 · refactor(comm): arming policy moves into TrtllmAllReduceBackend · 最后更新于 2026-07-14 17:09:07 +00:00    frostbyte_neo

642
2

4327a009a7 · perf(gdn): collapse flat decode head-group loop into one per-head launch · 最后更新于 2026-07-14 04:55:49 +00:00    frostbyte_neo

655
20

7afff1105d · fix(multimodal): order TP input broadcast on model stream · 最后更新于 2026-07-13 14:13:34 +00:00    frostbyte_neo

655
6

d5bc88f3b9 · perf(deepseek-v4): extend two-stage fused mHC launch to small token counts · 最后更新于 2026-07-13 10:26:48 +00:00    frostbyte_neo

738
6

6caad4b1e1 · perf(deepseek-v4): widen sparse compress cache launch for large ratios · 最后更新于 2026-07-10 08:22:02 +00:00    frostbyte_neo

679
1

e7d5a7c041 · style: format stacked DeepSeek V4 tests · 最后更新于 2026-07-10 05:00:15 +00:00    frostbyte_neo

703
9

271bd32b8b · fix(deepseek-v4): route large-token fused mHC around the allinone cliff · 最后更新于 2026-07-10 04:47:01 +00:00    frostbyte_neo

738
3

68e23cb628 · refactor(multimodal): narrow Qwen vision runtime scope · 最后更新于 2026-07-08 04:43:05 +00:00    frostbyte_neo

703
3

228e4483cb · fix(runtime): restore EnvField overrides after exceptions · 最后更新于 2026-07-07 13:20:49 +00:00    frostbyte_neo

701
1