vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

cuBLAS: use host pinned memory and dequantize while copying (#1207)

Browse source 
* cuBLAS: dequantize simultaneously while copying memory

* cuBLAS: use host pinned memory

* cuBLAS: improve ggml_compute_forward_mul_mat_f16_f32 with pinned memory

* cuBLAS: also pin kv cache

* fix rebase
This commit is contained in:
slaren 2023-04-29 02:04:18 +02:00 • committed by GitHub
parent b1ee8f59b4
commit 7fc50c051a
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
6 changed files with 110 additions and 54 deletions
Download patch file Download diff file Expand all files Collapse all files

10
ggml-cuda.h
View file

@ -26,9 +26,14 @@ extern "C" {
} while (0)
extern cublasHandle_t g_cublasH;
extern cudaStream_t g_cudaStream;
extern cudaStream_t g_cudaStream;
extern cudaStream_t g_cudaStream2;
extern cudaEvent_t g_cudaEvent;
void ggml_init_cublas(void);
void * ggml_cuda_host_malloc(size_t size);
void ggml_cuda_host_free(void * ptr);
void * ggml_cuda_pool_malloc(size_t size, size_t * actual_size);
void ggml_cuda_pool_free(void * ptr, size_t size);
@ -41,6 +46,9 @@ void dequantize_row_q8_0_cuda(const void * vx, float * y, int k, cudaStream_t st
cudaError_t ggml_cuda_h2d_tensor_2d(void * dst, const struct ggml_tensor * src, uint64_t i3, uint64_t i2, cudaStream_t stream);
typedef void (*dequantize_row_q_cuda_t)(const void * x, float * y, int k, cudaStream_t stream);
dequantize_row_q_cuda_t ggml_get_dequantize_row_q_cuda(enum ggml_type type);
#ifdef __cplusplus
}
#endif