Default Branch

76f46ad29d · hexagon: add CLAMP op (#25934) · Updated 2026-07-20 23:12:09 +00:00

Branches

40ec4882c4 · ggml : use F16C conversion when available · Updated 2023-05-17 17:05:51 +00:00    furyhawk

9517
1

a3e6d62283 · cuda : alternative q4_q8 kernel · Updated 2023-05-12 14:02:39 +00:00    furyhawk

9551
8

e116eb638c · ggml : speed-up Q5_0 + Q5_1 at 4 threads · Updated 2023-05-11 15:51:56 +00:00    furyhawk

9553
20

4baa85633a · Fix build · Updated 2023-05-07 01:44:07 +00:00    furyhawk

9561
5

31ff9e2e83 · ci : add cublas to windows release · Updated 2023-05-03 21:21:20 +00:00    furyhawk

9576
1

102cd98074 · ggml : Q4_3c using 2x "Full range" approach · Updated 2023-04-23 11:56:44 +00:00    furyhawk

9657
8

71e6ae3779 · ggml : continue from #729 (wip) · Updated 2023-04-22 15:49:07 +00:00    furyhawk

9657
7

a0242a833c · Minor, plus rebase on master · Updated 2023-04-22 14:07:10 +00:00    furyhawk

9657
2

4b8d5e3890 · llama : quantize attention results · Updated 2023-04-22 08:35:13 +00:00    furyhawk

9662
1

1506737499 · Add mmap pages stats (disabled by default) · Updated 2023-04-16 16:22:30 +00:00    furyhawk

9712
1

36ddd12924 · llama : add flash attention (demo) · Updated 2023-04-05 19:12:04 +00:00    furyhawk

9778
1

c9c820ff36 · Added support for _POSIX_MAPPED_FILES if defined in source (#564) · Updated 2023-03-28 21:26:25 +00:00    furyhawk

10012
8

4aeee216fd · Regroup q4_1 dot addition for better numerics. · Updated 2023-03-24 20:20:57 +00:00    furyhawk

9893
2

66ea164e1d · Kahan summation on Q4_1 · Updated 2023-03-23 03:28:51 +00:00    furyhawk

9920
2

711224708d · Break up loop for numeric stability · Updated 2023-03-23 02:14:44 +00:00    furyhawk

9920
2

3a0dcb3920 · Implement server mode. · Updated 2023-03-22 17:34:19 +00:00    furyhawk

9921
5
dev

a169bb889c · Gate signal support on being on a unixoid system. (#74) · Updated 2023-03-13 03:08:01 +00:00    furyhawk

10027
0
Included