Default Branch

main 30381fc1fc · Faster hybrid inference when shared experts (#1191) · Updated 2026-01-26 05:22:05 +00:00

ZIP  TAR.GZ

Branches

ik/fix_reduce_race d35cf5a92d · Fix data races in the reduce op · Updated 2026-01-09 08:32:00 +00:00	4147 4100		ZIP  TAR.GZ
ik/ernie_graph 37caf11f2c · Cleanup · Updated 2026-01-08 08:18:34 +00:00	49 2		ZIP  TAR.GZ
ik/dont_abort_on_nccl_init_failure 8308320bca · Do not abort on NCCL initizalization failure · Updated 2026-01-08 07:16:23 +00:00	4147 4098		ZIP  TAR.GZ
ik/gpt_oss_graph 646fe94085 · Force split_mode_f16 to false · Updated 2026-01-07 16:34:41 +00:00	4147 4098		ZIP  TAR.GZ
ik/hunyuan_graph edd56b1bf7 · Split mode "graph" for Hunyuan-MoE · Updated 2026-01-07 09:12:46 +00:00	4147 4095		ZIP  TAR.GZ
ik/mimo2_4_gpus a29f62fc50 · Enable up to 4 GPUs for Mimo2-Flash · Updated 2026-01-07 07:36:00 +00:00	4147 4094		ZIP  TAR.GZ
ik/fix_ring_reduction 10c531c8de · Actually enable it · Updated 2026-01-07 05:55:10 +00:00	4147 4094		ZIP  TAR.GZ
ik/reduce_race_quick_fix 289aadb9d4 · Disable ring reduction for now · Updated 2026-01-06 13:14:30 +00:00	4147 4092		ZIP  TAR.GZ
ik/qwen3vl_graph b41f2c3ffe · Split mode 'graph' fpr Qwen3-VL · Updated 2026-01-05 13:21:10 +00:00	57 1		ZIP  TAR.GZ
ik/qwen3_graph a725f15d9d · Split mode graph for Qwen3 · Updated 2026-01-05 08:00:30 +00:00	58 1		ZIP  TAR.GZ
ik/mimo2_graph b586f89e50 · Set max_gpu to 2 for Mimo2 · Updated 2026-01-05 06:49:17 +00:00	59 3		ZIP  TAR.GZ
ik/fix_fa_192_128 ae3498dabd · Fix race in CUDA FA for head sizes 192/128 · Updated 2026-01-05 06:17:10 +00:00	4147 4088		ZIP  TAR.GZ
ik/mimo2 ba0e88a5e3 · Minor · Updated 2025-12-28 08:57:24 +00:00	4147 4087		ZIP  TAR.GZ
ik/graph_reuse_on bf3ff8ec41 · Turn on graph reuse by default · Updated 2025-12-27 07:22:46 +00:00	4147 4083		ZIP  TAR.GZ
ik/nccl3_async 29d323117c · Command line option to turn on async. Set to false by defualt for now · Updated 2025-12-27 06:24:01 +00:00	4147 4109		ZIP  TAR.GZ
ik/more_set_device 0e059879b7 · Be more careful with having set the device before using a stream · Updated 2025-12-26 18:18:28 +00:00	4147 4081		ZIP  TAR.GZ
ik/ring_reduce f109274859 · Graph parallel: better PP performance for 3 and more GPUs · Updated 2025-12-26 15:57:19 +00:00	4147 4080		ZIP  TAR.GZ
ik/fix_no_p2p_case d2f52ec104 · Fix split mode graph when p2p is not enabled · Updated 2025-12-25 07:52:38 +00:00	4147 4079		ZIP  TAR.GZ
ik/reduce_no_nccl 723e18bb98 · Reduce add improvemens without NCCL · Updated 2025-12-24 14:23:10 +00:00	4147 4078		ZIP  TAR.GZ
ik/split_mode_f32 c6a3903571 · Be able to set reduce op data type for split mode "graph" · Updated 2025-12-24 10:57:41 +00:00	4147 4075		ZIP  TAR.GZ