llama_cpp 0.3.8 → 0.5.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- checksums.yaml +4 -4
- data/CHANGELOG.md +19 -0
- data/README.md +1 -1
- data/examples/chat.rb +4 -6
- data/ext/llama_cpp/extconf.rb +3 -3
- data/ext/llama_cpp/llama_cpp.cpp +129 -124
- data/ext/llama_cpp/src/ggml-alloc.c +90 -113
- data/ext/llama_cpp/src/ggml-alloc.h +1 -1
- data/ext/llama_cpp/src/ggml-cuda.cu +350 -77
- data/ext/llama_cpp/src/ggml-cuda.h +13 -0
- data/ext/llama_cpp/src/ggml-metal.h +4 -0
- data/ext/llama_cpp/src/ggml-metal.m +226 -121
- data/ext/llama_cpp/src/ggml-metal.metal +157 -35
- data/ext/llama_cpp/src/ggml.c +2724 -584
- data/ext/llama_cpp/src/ggml.h +282 -31
- data/ext/llama_cpp/src/k_quants.c +112 -56
- data/ext/llama_cpp/src/llama.cpp +4857 -2986
- data/ext/llama_cpp/src/llama.h +180 -126
- data/lib/llama_cpp/version.rb +2 -2
- data/lib/llama_cpp.rb +2 -2
- data/sig/llama_cpp.rbs +12 -11
- metadata +2 -2
|
@@ -2,6 +2,14 @@
|
|
|
2
2
|
|
|
3
3
|
#include "ggml.h"
|
|
4
4
|
|
|
5
|
+
#ifdef GGML_USE_HIPBLAS
|
|
6
|
+
#define GGML_CUDA_NAME "ROCm"
|
|
7
|
+
#define GGML_CUBLAS_NAME "hipBLAS"
|
|
8
|
+
#else
|
|
9
|
+
#define GGML_CUDA_NAME "CUDA"
|
|
10
|
+
#define GGML_CUBLAS_NAME "cuBLAS"
|
|
11
|
+
#endif
|
|
12
|
+
|
|
5
13
|
#ifdef __cplusplus
|
|
6
14
|
extern "C" {
|
|
7
15
|
#endif
|
|
@@ -16,9 +24,14 @@ GGML_API bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const str
|
|
|
16
24
|
GGML_API void ggml_cuda_set_tensor_split(const float * tensor_split);
|
|
17
25
|
GGML_API void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor);
|
|
18
26
|
GGML_API void ggml_cuda_free_data(struct ggml_tensor * tensor);
|
|
27
|
+
|
|
19
28
|
GGML_API void ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
|
|
20
29
|
GGML_API void ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
|
|
21
30
|
GGML_API void ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor);
|
|
31
|
+
|
|
32
|
+
GGML_API void ggml_cuda_assign_buffers_no_alloc(struct ggml_tensor * tensor);
|
|
33
|
+
GGML_API void ggml_cuda_assign_scratch_offset(struct ggml_tensor * tensor, size_t offset);
|
|
34
|
+
|
|
22
35
|
GGML_API void ggml_cuda_set_main_device(int main_device);
|
|
23
36
|
GGML_API void ggml_cuda_set_mul_mat_q(bool mul_mat_q);
|
|
24
37
|
GGML_API void ggml_cuda_set_scratch_size(size_t scratch_size);
|
|
@@ -24,6 +24,7 @@
|
|
|
24
24
|
|
|
25
25
|
// max memory buffers that can be mapped to the device
|
|
26
26
|
#define GGML_METAL_MAX_BUFFERS 16
|
|
27
|
+
#define GGML_METAL_MAX_COMMAND_BUFFERS 32
|
|
27
28
|
|
|
28
29
|
struct ggml_tensor;
|
|
29
30
|
struct ggml_cgraph;
|
|
@@ -38,6 +39,9 @@ struct ggml_metal_context;
|
|
|
38
39
|
struct ggml_metal_context * ggml_metal_init(int n_cb);
|
|
39
40
|
void ggml_metal_free(struct ggml_metal_context * ctx);
|
|
40
41
|
|
|
42
|
+
void * ggml_metal_host_malloc(size_t n);
|
|
43
|
+
void ggml_metal_host_free (void * data);
|
|
44
|
+
|
|
41
45
|
// set the number of command buffers to use
|
|
42
46
|
void ggml_metal_set_n_cb(struct ggml_metal_context * ctx, int n_cb);
|
|
43
47
|
|