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

028f741dc4 · Merge branch 'main' into jjd/fi_cutedsl_w4a16_linear · 最后更新于 2026-09-02 18:13:41 +00:00    frostbyte_neo

110
3

43c66a8a9e · Merge branch 'main' into jjd/ple_hc_refactor_opt · 最后更新于 2026-09-02 18:13:25 +00:00    frostbyte_neo

110
13

b1633015da · test(qsa): make persistent dispatch platform-independent · 最后更新于 2026-09-02 09:51:22 +00:00    frostbyte_neo

113
6

15d6800915 · fic(ci): assign Kimi K3 DSpark perf to model stage to enable manual runs · 最后更新于 2026-09-02 04:09:19 +00:00    frostbyte_neo

114
1

3d865b596f · feat(qwen4-exp): enable PLE embedding offload by default · 最后更新于 2026-09-02 04:05:20 +00:00    frostbyte_neo

114
5

129c991065 · perf(k3): optimize prefill MoE blocks on gfx950 · 最后更新于 2026-09-02 03:34:06 +00:00    frostbyte_neo

114
1

bb04864381 · test: preserve CuTe compile option binding · 最后更新于 2026-09-01 20:16:27 +00:00    frostbyte_neo

120
24

ef96140d3f · Fuse KDA decode state staging · 最后更新于 2026-09-01 17:13:38 +00:00    frostbyte_neo

142
5

a23c4bb1c9 · ci(amd): extend MI450 simulator job timeout · 最后更新于 2026-09-01 15:04:36 +00:00    frostbyte_neo

124
6

2d7272892d · code(rms): remove unused branch code. · 最后更新于 2026-09-01 04:02:08 +00:00    frostbyte_neo

127
2

070c458454 · perf(gemm): 256-bit skinny GEMV configs for the measured decode shapes · 最后更新于 2026-09-01 04:00:48 +00:00    frostbyte_neo

127
1

e6d90b26c5 · refactor(prefill graph): one capture-table width rule, no per-backend knobs · 最后更新于 2026-08-31 16:56:33 +00:00    frostbyte_neo

154
3

c48743137f · perf(k3): decode gates, GEMV route QA, and FlashInfer 0.6.18 re-tune · 最后更新于 2026-08-31 16:14:16 +00:00    frostbyte_neo

142
1

76ef6f3ad4 · Merge branch 'main' into hongtaoc/k3-latent-tail-fast-paths · 最后更新于 2026-08-31 13:38:21 +00:00    frostbyte_neo

142
4

62e5e537d7 · refactor(qwen4): remove redundant hyperconnection paths · 最后更新于 2026-08-31 12:17:31 +00:00    frostbyte_neo

142
2

cbce6b0e67 · fix(moe): pad NVFP4 expert shards for TRT-LLM · 最后更新于 2026-08-31 08:35:33 +00:00    frostbyte_neo

143
1

6219bd2eab · ci: add four-task B300 DeepSWE smoke option · 最后更新于 2026-08-31 06:59:44 +00:00    frostbyte_neo

146
12

fdfb93bd06 · fix(profile): make /start_profile actually collect GPU work · 最后更新于 2026-08-31 03:54:18 +00:00    frostbyte_neo

158
2

b0811a4685 · Merge branch 'main' into borontion/chore/attn-apis · 最后更新于 2026-08-30 22:13:59 +00:00    frostbyte_neo

146
3

bded67a7bf · perf(prefill): size the prefill graph to the chunk, and degrade instead of dying · 最后更新于 2026-08-30 15:03:12 +00:00    frostbyte_neo

151
1