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

b52879a79f · Merge main prefill checkpoints into Qwen4-Exp backends · 最后更新于 2026-09-12 08:11:06 +00:00    frostbyte_neo

3
32

4400b41af1 · feat(amd): enable DeepSeek V4 on gfx1250 with portable Triton · 最后更新于 2026-09-11 19:04:31 +00:00    frostbyte_neo

13
1

0630294cba · fix(amd): narrow gfx1250 MXFP4 indices only at 32-bit operands · 最后更新于 2026-09-11 16:22:32 +00:00    frostbyte_neo

34
2

68b36bf653 · perf: enable DeepSeek V4.1 decode graphs and streaming indexing · 最后更新于 2026-09-11 10:37:36 +00:00    frostbyte_neo

40
4

612d0dbe9f · perf: improve PDL overlap in GDN, gated residual, and QSA · 最后更新于 2026-09-11 09:56:41 +00:00    frostbyte_neo

38
13

fc75f40bfb · Keep the L3 PR focused by removing unrelated CI changes · 最后更新于 2026-09-11 03:57:17 +00:00    frostbyte_neo

20
79

6fc9da5c5e · ci: retry unsuccessful Slurm cases at the original commit · 最后更新于 2026-09-11 03:17:08 +00:00    frostbyte_neo

22
1

9451891ad0 · Add AMD kernel benchmark regression CI · 最后更新于 2026-09-10 23:07:30 +00:00    frostbyte_neo

22
2

843fac0ef6 · fmt · 最后更新于 2026-09-10 20:19:50 +00:00    frostbyte_neo

33
15

0eab9e0366 · perf(amd): enable GLM-5.3 FP8 dense prefill · 最后更新于 2026-09-10 17:45:56 +00:00    frostbyte_neo

33
1

3d9dbd825e · perf(comm): tune TP4 Iris decode launches · 最后更新于 2026-09-10 17:41:22 +00:00    frostbyte_neo

33
6

eb0ecd7799 · perf(mhc): tune GLM-5.3-Flash reduction grid · 最后更新于 2026-09-10 16:59:16 +00:00    frostbyte_neo

33
5

017adde4e9 · perf(amd): reuse sigmoid routing kernels across row counts · 最后更新于 2026-09-10 16:59:10 +00:00    frostbyte_neo

33
3

e836cc5ce3 · merge: update Iris MoE policy stack · 最后更新于 2026-09-10 16:56:13 +00:00    frostbyte_neo

33
14

616c720b5c · merge: resolve Iris preallocation conflicts with main · 最后更新于 2026-09-10 16:55:29 +00:00    frostbyte_neo

33
12

d8929cd218 · perf(gptoss): add gfx1250 softmax top-k routing · 最后更新于 2026-09-10 16:36:53 +00:00    frostbyte_neo

37
1

be4da55446 · fix(comm): preserve zero producer-direct capacity · 最后更新于 2026-09-10 01:13:54 +00:00    frostbyte_neo

46
7

84478e96d2 · test(kda): vectorize MTP reference batch · 最后更新于 2026-09-10 00:27:26 +00:00    frostbyte_neo

52
2

f8f6f2b54a · fix(kimi3): tolerate missing gfx1250 AttnRes and snapshot graph replay · 最后更新于 2026-09-09 21:25:26 +00:00    frostbyte_neo

55
7

e9485f4c81 · fix(kimi3): keep auto shared-SiTU on gfx1250 large-M · 最后更新于 2026-09-09 21:25:25 +00:00    frostbyte_neo

55
4