 Johannes Gäßler
|
17c97fb062
CUDA: mul_mat_vec_q max. batch size 8 -> 4 (#5370)
|
2 years ago |
|
Johannes Gäßler
|
2c516611f1
CUDA: mul_mat_vec_q for batch sizes > 1 (#5351)
|
2 years ago |
 slaren
|
8ca511cade
cuda : fix LLAMA_CUDA_F16 (#5262)
|
2 years ago |
 JidongZhang-THU
|
15606309a0
llava : add MobileVLM support (#5132)
|
2 years ago |
 Georgi Gerganov
|
8f8ddfcfad
sync : ggml (#0)
|
2 years ago |
 John Balis
|
625a699b54
`ggml_cuda_cpy` support for 4d tensors and float16->float32 upcasting (ggml/686)
|
2 years ago |
 Kawrakow
|
f4d7e54974
SOTA 3-bit quants (#5196)
|
2 years ago |
 0cc4m
|
2307523d32
ggml : add Vulkan backend (#2059)
|
2 years ago |
|
slaren
|
62fead3ea0
cuda : fix tensor size calculation for non-split buffer (#5145)
|
2 years ago |
 Engininja2
|
cd4fddb29f
cuda : fix 2-bit quants on amd hip (#5105)
|
2 years ago |
|
Johannes Gäßler
|
9ecdd12e95
CUDA: more info when no device code (#5088)
|
2 years ago |
 Kylin
|
cca894f16a
cuda : fix compile error in jetson platform (#4975)
|
2 years ago |
|
Georgi Gerganov
|
38566680cd
ggml : add IQ2 to test-backend-ops + refactoring (#4990)
|
2 years ago |
 Justine Tunney
|
a0b3ac8c48
ggml : introduce GGML_CALL function annotation (#4850)
|
2 years ago |
|
Georgi Gerganov
|
ddb008d845
cuda : fix dequantize kernel names (#4938)
|
2 years ago |
|
Kawrakow
|
4a3156de2f
CUDA: faster dequantize kernels for Q4_0 and Q4_1 (#4938)
|
2 years ago |
|
Johannes Gäßler
|
3fe81781e3
CUDA: faster q8_0 -> f16 dequantization (#4895)
|
2 years ago |
|
slaren
|
e7e4df031b
llama : ggml-backend integration (#4766)
|
2 years ago |
|
Johannes Gäßler
|
1b280c9fff
CUDA: fix softmax compile for old CUDA versions (#4862)
|
2 years ago |
|
Kawrakow
|
49662cbed3
ggml : SOTA 2-bit quants (add IQ2_XS) (#4856)
|
2 years ago |
 Erik Scholz
|
f34432ca1e
fix : cuda order of synchronization when setting a buffer (ggml/679)
|
2 years ago |
|
Johannes Gäßler
|
8f900abfc0
CUDA: faster softmax via shared memory + fp16 math (#4742)
|
2 years ago |
|
Kawrakow
|
dd5ae06405
SOTA 2-bit quants (#4773)
|
2 years ago |
|
Johannes Gäßler
|
d5a410e855
CUDA: fixed redundant value dequantization (#4809)
|
2 years ago |
 Konstantin Zhuravlyov
|
63ee677efd
ggml : use __builtin_amdgcn_sudot4 in __dp4a for gfx11 (#4787)
|
2 years ago |
 Finn Voorhees
|
1bf681f90e
ggml : add error handling to graph_compute (whisper/1714)
|
2 years ago |
|
Georgi Gerganov
|
7bed7eba35
cuda : simplify expression
|
2 years ago |
|
Georgi Gerganov
|
d55356d3ba
cuda : mark I16 and I32 ops as unsupported
|
2 years ago |
|
Johannes Gäßler
|
39d8bc71ed
CUDA: fixed tensor cores not being used on RDNA3 (#4697)
|
2 years ago |
|
Johannes Gäßler
|
a20f3c7465
CUDA: fix tensor core logic for Pascal and HIP (#4682)
|
2 years ago |
