分支列表

默认分支

ffaf57832d · fix(glm5_next): remap quantized forget-gate scales/biases in sanitize (#3570) · 最后更新于 2026-09-11 05:31:37 +00:00

分支列表

bbb34b0b9c · fix: construct exact FP8 scales in Metal (#3574) · 最后更新于 2026-09-11 09:00:08 +00:00    frostbyte_neo

0
4

244eba941e · fix: separate partial prefill costs from representative estimates · 最后更新于 2026-09-08 06:03:09 +00:00    frostbyte_neo

36
1

d9ea483a6b · merge: preserve SDPA memory guards when syncing main (#3333) · 最后更新于 2026-09-06 14:15:15 +00:00    frostbyte_neo

57
3

74fcc595ca · build(deps): update transformers requirement · 最后更新于 2026-09-06 02:22:53 +00:00    frostbyte_neo

93
1

794f848b37 · fix(vlm): preserve GLM batch convolution state · 最后更新于 2026-08-27 13:17:36 +00:00    frostbyte_neo

158
3

dece501e62 · fix: defer VLM parameter eval to chunked materialization (#3179) · 最后更新于 2026-08-27 12:18:56 +00:00    frostbyte_neo

155
0
已包含

4349a29690 · fix: support Qwen4 FP8 PLE in oQ conversion · 最后更新于 2026-08-27 08:34:51 +00:00    frostbyte_neo

162
7

c0503a9e9a · fix(ane): schedule tiled GDN projections together (#3117) · 最后更新于 2026-08-26 04:12:50 +00:00    frostbyte_neo

166
3

e48c43da21 · fix: address MLX 0.32.2 CI regressions · 最后更新于 2026-08-25 15:51:13 +00:00    frostbyte_neo

166
6

5923c55e36 · chore(deps): bump ddgs from 9.14.4 to 9.15.0 · 最后更新于 2026-08-23 02:23:33 +00:00    frostbyte_neo

215
1

e5d487ed14 · chore(deps-dev): bump nanobind from 2.13.0 to 2.15.0 · 最后更新于 2026-08-23 02:23:19 +00:00    frostbyte_neo

215
1

7dc7c0e32b · chore(deps): update transformers requirement · 最后更新于 2026-08-23 02:23:01 +00:00    frostbyte_neo

215
1

06e9431a0a · chore(deps): bump mlx from 0.32.0 to 0.32.1 · 最后更新于 2026-08-23 02:22:45 +00:00    frostbyte_neo

215
1

8146265b3c · fix: re-enter Lightning MTP after performance parking (#2997) · 最后更新于 2026-08-21 15:04:36 +00:00    frostbyte_neo

219
1

0b9edfcf0c · feat(deepseek): add experimental hybrid ANE prompt processing · 最后更新于 2026-08-18 16:06:23 +00:00    frostbyte_neo

295
1

12f2266106 · fix(qwen3.5): align ANE prefill scheduler shape · 最后更新于 2026-08-17 17:47:52 +00:00    frostbyte_neo

322
2

48c8d7573e · perf: fuse qwen MoE decode router top-k · 最后更新于 2026-08-17 15:39:38 +00:00    frostbyte_neo

332
1

6cf105607c · feat(cluster): store snapshot chains of segments instead of cumulative copies · 最后更新于 2026-08-13 08:34:36 +00:00    frostbyte_neo

412
0
已包含

eca7d42c8d · Merge main into feat/chat-stt · 最后更新于 2026-08-12 06:36:35 +00:00    frostbyte_neo

427
4

93accf79eb · perf: use upstream MLX QMM for faster Lightning MTP · 最后更新于 2026-08-04 16:58:08 +00:00    frostbyte_neo

499
1