 Georgi Gerganov
|
0548a4187f
ggml : generalize GGML_OP_CONCAT (#7563)
|
il y a 1 an |
|
Georgi Gerganov
|
1d8fca72ae
metal : add GGML_OP_REPEAT kernels (#7557)
|
il y a 1 an |
|
Georgi Gerganov
|
62bfef5194
metal : disable FA kernel for HS=256 (#7556)
|
il y a 1 an |
|
Georgi Gerganov
|
e84b71c2c6
ggml : drop support for QK_K=64 (#7473)
|
il y a 1 an |
 liuwei-git
|
201cc11afa
llama : add phi3 128K model support (#7225)
|
il y a 1 an |
 John Balis
|
48aa8fd1f2
ggml : add `ggml_upscale_ext` (ggml/814)
|
il y a 1 an |
|
Georgi Gerganov
|
e8a7fd4fb0
metal : support FA without mask + add asserts (#7278)
|
il y a 1 an |
|
Georgi Gerganov
|
f308ea7059
metal : tune soft_max number of threads (whisper/0)
|
il y a 1 an |
|
Georgi Gerganov
|
6aeff24f8b
metal : fix indent (ggml/0)
|
il y a 1 an |
 Justina Cho
|
f5ef34e428
feat: implemented sigmoid function (ggml/806)
|
il y a 1 an |
|
Georgi Gerganov
|
9cb317f77e
ggml : full ALiBi support (#7192)
|
il y a 1 an |
|
Georgi Gerganov
|
18e437665c
metal : fix flash attention kernel requirements (#7169)
|
il y a 1 an |
 Gilad S
|
26458af1d6
metal : use `vm_allocate` instead of `posix_memalign` on macOS (#7078)
|
il y a 1 an |
 Justine Tunney
|
3855416027
ggml : introduce bfloat16 support (#6412)
|
il y a 1 an |
 Kevin Gibbons
|
f364eb6fb5
switch to using localizedDescription (#7010)
|
il y a 1 an |
|
Georgi Gerganov
|
77e15bec62
metal : remove deprecated error code (#7008)
|
il y a 1 an |
|
Kevin Gibbons
|
a68a1e7ed0
metal : log more info on error (#6987)
|
il y a 1 an |
|
Georgi Gerganov
|
9c67c2773d
ggml : add Flash Attention (#5021)
|
il y a 1 an |
 slaren
|
0d56246f4b
ggml : group all experts in a single ggml_mul_mat_id (#6505)
|
il y a 1 an |
 Shijie
|
f4dea7da18
llama : add qwen2moe (#6074)
|
il y a 1 an |
 Dave
|
422c2aff1c
Added support for GGML_OP_CLAMP in Metal (#6662)
|
il y a 1 an |
|
slaren
|
fbbc030ba9
metal : unify mul_mv_id kernels (#6556)
|
il y a 1 an |
|
slaren
|
08a0c02060
ggml : mul_mat_id use the same tensor for all the experts (#6387)
|
il y a 1 an |
 compilade
|
557410b8f0
llama : greatly reduce output buffer memory usage (#6122)
|
il y a 1 an |
 Kawrakow
|
55c1b2a3bb
IQ1_M: 1.75 bpw quantization (#6302)
|
il y a 1 an |
|
Georgi Gerganov
|
b3e94f26ba
metal : proper assert for mat-mat memory alignment (#6225)
|
il y a 1 an |
|
Kawrakow
|
76aa30a263
Add ability to use Q5_0, Q5_1, and IQ4_NL for quantized K cache (#6183)
|
il y a 1 an |
|
slaren
|
2bf8d0f7c4
backend : offload large batches to GPU (#6083)
|
il y a 1 an |
|
Georgi Gerganov
|
381da2d9f0
metal : build metallib + fix embed path (#6015)
|
il y a 1 an |
|
slaren
|
f30ea47a87
llama : add pipeline parallelism support (#6017)
|
il y a 1 an |
