| .. |
|
template-instances
|
7d1a378b8f
CUDA: refactor mmq, dmmv, mmvq (#7716)
|
1 éve |
|
acc.cu
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
acc.cuh
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
arange.cu
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
arange.cuh
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
argsort.cu
|
76d66ee0be
CUDA: faster q2_K, q3_K MMQ + int8 tensor cores (#7921)
|
1 éve |
|
argsort.cuh
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
binbcast.cu
|
0d56246f4b
ggml : group all experts in a single ggml_mul_mat_id (#6505)
|
1 éve |
|
binbcast.cuh
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
clamp.cu
|
bc4bba364f
Introduction of CUDA Graphs to LLama.cpp (#6766)
|
1 éve |
|
clamp.cuh
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
common.cuh
|
a818f3028d
CUDA: use MMQ instead of cuBLAS by default (#8075)
|
1 éve |
|
concat.cu
|
cce3dcffc5
cuda : non-cont concat support (#7610)
|
1 éve |
|
concat.cuh
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
convert.cu
|
e84b71c2c6
ggml : drop support for QK_K=64 (#7473)
|
1 éve |
|
convert.cuh
|
5dc9dd7152
llama : add Command R Plus support (#6491)
|
1 éve |
|
cpy.cu
|
bc4bba364f
Introduction of CUDA Graphs to LLama.cpp (#6766)
|
1 éve |
|
cpy.cuh
|
bc4bba364f
Introduction of CUDA Graphs to LLama.cpp (#6766)
|
1 éve |
|
dequantize.cuh
|
5dc9dd7152
llama : add Command R Plus support (#6491)
|
1 éve |
|
diagmask.cu
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
diagmask.cuh
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
dmmv.cu
|
7d1a378b8f
CUDA: refactor mmq, dmmv, mmvq (#7716)
|
1 éve |
|
dmmv.cuh
|
d48ccf3ad4
sync : ggml (#6351)
|
1 éve |
|
fattn-common.cuh
|
1f0dabda8d
CUDA: use tensor cores for MMQ (#7676)
|
1 éve |
|
fattn-tile-f16.cu
|
1f0dabda8d
CUDA: use tensor cores for MMQ (#7676)
|
1 éve |
|
fattn-tile-f16.cuh
|
0fc1e820a9
CUDA: faster large batch FA without tensor cores (#7314)
|
1 éve |
|
fattn-tile-f32.cu
|
750f60c03e
CUDA: fix Pascal FA, deq. KV to FP16 for batch > 8 (#7681)
|
1 éve |
|
fattn-tile-f32.cuh
|
0fc1e820a9
CUDA: faster large batch FA without tensor cores (#7314)
|
1 éve |
|
fattn-vec-f16.cuh
|
1f0dabda8d
CUDA: use tensor cores for MMQ (#7676)
|
1 éve |
|
fattn-vec-f32.cuh
|
963552903f
CUDA: fix broken oob check for FA vec f32 kernel (#7904)
|
1 éve |
|
fattn-wmma-f16.cuh
|
1f0dabda8d
CUDA: use tensor cores for MMQ (#7676)
|
1 éve |
|
fattn.cu
|
750f60c03e
CUDA: fix Pascal FA, deq. KV to FP16 for batch > 8 (#7681)
|
1 éve |
|
fattn.cuh
|
9c67c2773d
ggml : add Flash Attention (#5021)
|
1 éve |
|
getrows.cu
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
getrows.cuh
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
im2col.cu
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
im2col.cuh
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
mma.cuh
|
9a590c8226
CUDA: optimize MMQ int8 tensor core performance (#8062)
|
1 éve |
|
mmq.cu
|
a818f3028d
CUDA: use MMQ instead of cuBLAS by default (#8075)
|
1 éve |
|
mmq.cuh
|
3b099bcd9c
CUDA: fix MMQ writeback for int8 tensor cores (#8100)
|
1 éve |
|
mmvq.cu
|
19b7a836f6
cuda : fix bounds check for src0 rows in MMVQ kernel (whisper/2231)
|
1 éve |
|
mmvq.cuh
|
a818f3028d
CUDA: use MMQ instead of cuBLAS by default (#8075)
|
1 éve |
|
norm.cu
|
fb76ec31a9
ggml : fix YARN + add tests + add asserts (#7617)
|
1 éve |
|
norm.cuh
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
pad.cu
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
pad.cuh
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
pool2d.cu
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
pool2d.cuh
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
quantize.cu
|
42b53d192f
CUDA: revise q8_1 data layout for mul_mat_q (#7824)
|
1 éve |
|
quantize.cuh
|
42b53d192f
CUDA: revise q8_1 data layout for mul_mat_q (#7824)
|
1 éve |
|
rope.cu
|
2b3389677a
ggml : refactor rope norm/neox (#7634)
|
1 éve |
|
rope.cuh
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
scale.cu
|
bc4bba364f
Introduction of CUDA Graphs to LLama.cpp (#6766)
|
1 éve |
|
scale.cuh
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
softmax.cu
|
76d66ee0be
CUDA: faster q2_K, q3_K MMQ + int8 tensor cores (#7921)
|
1 éve |
|
softmax.cuh
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
sumrows.cu
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
sumrows.cuh
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
tsembd.cu
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
tsembd.cuh
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
unary.cu
|
43b35e38ba
Add support for sqrt on CUDA (#7953)
|
1 éve |
|
unary.cuh
|
43b35e38ba
Add support for sqrt on CUDA (#7953)
|
1 éve |
|
upscale.cu
|
48aa8fd1f2
ggml : add `ggml_upscale_ext` (ggml/814)
|
1 éve |
|
upscale.cuh
|
ae1f211ce2
cuda : refactor into multiple files (#6269)
|
1 éve |
|
vecdotq.cuh
|
76d66ee0be
CUDA: faster q2_K, q3_K MMQ + int8 tensor cores (#7921)
|
1 éve |
Johannes Gäßler