Default Branch

2fb371e318 · feat(kvcache): add fp8_ds_mla and nvfp4_ds_mla kv formats for deepseek-v4 / v4.1 (#920) · Updated 2026-11-19 14:04:03 +08:00

Branches

5eb7597d27 · refactor(engine): drop the model field from the dsv4 kv formats · Updated 2026-11-19 07:27:17 +08:00

1
2

530965688d · fix(attention): bound multi-row triton decode tiles by the LDS size on ROCm · Updated 2026-11-14 08:40:10 +08:00

9
1

7e353d11cd · fix(attention): keep multi-row triton decode free of register spills on sm_80 to sm_120 · Updated 2026-11-13 09:52:17 +08:00

12
2

d4a9f0e1ad · perf(qwen4_exp): write the q8_1 input of the hc projections in the norm and silu kernels · Updated 2026-11-12 07:05:29 +08:00

13
2

f4838cad01 · Merge remote-tracking branch 'public/main' into feat/kv-quant-per-token-head · Updated 2026-11-12 06:49:16 +08:00

13
6

785b086c86 · Merge remote-tracking branch 'origin/main' into df11-online-quant · Updated 2026-11-12 05:09:37 +08:00

14
3

48b57b05c6 · fix(spec): correct the token map vram note and seed its test · Updated 2026-11-12 04:07:05 +08:00

15
3

42553c5463 · fix(quant): accept the modelopt_mixed quant method · Updated 2026-11-10 03:24:14 +08:00

19
1

3673da8534 · chore(release): 0.2.2 · Updated 2026-11-09 09:51:15 +08:00

20
1

987015cf62 · fix(ggml): cap the MoE route sort's shared memory at the device limit · Updated 2026-11-08 10:19:44 +08:00

21
1

2d2dca0e04 · test(engine): check graph capture runs on the dummy page · Updated 2026-11-08 08:23:59 +08:00

22
2

9d87bd7d92 · test(models): give the ricursive test's stand-in batch a features field · Updated 2026-11-07 09:39:02 +08:00

24
48

1a78bb2e7b · fix(utils): read stop ids from config.json without generation_config.json · Updated 2026-11-06 11:21:28 +08:00

27
1

6a23d89f27 · chore(ggml): replace freetoken.patch with the llama.cpp pin · Updated 2026-11-06 06:48:37 +08:00

28
1

ed27d13443 · perf(ggml): fold the MoE MMQ tile map into the route sort · Updated 2026-11-05 17:17:07 +08:00

29
2

92ceb27d2b · fix(shell): reset the first-delta count per turn and ignore zero running totals · Updated 2026-11-05 14:16:01 +08:00

30
4

0e4a1cf77d · docs: document gguf support and ask for checkpoint links · Updated 2026-11-05 12:01:20 +08:00

31
1

382bda03d1 · docs: add a guide to speeding up decode · Updated 2026-11-04 08:12:09 +08:00

37
1

70af8532ae · feat(utils): add once-per-message logger methods · Updated 2026-11-04 07:20:43 +08:00

39
1

54416c9e8f · perf(cpu-moe): speed up the nvfp4 W4A8 kernels · Updated 2026-11-03 00:27:23 +08:00    trending

44
1