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

cuda : use CUBLAS_COMPUTE_16F for non-attention ops

Browse source
This commit is contained in:
Georgi Gerganov 2023-10-27 20:15:21 +03:00
parent 3b9ea655d4
commit 0f2498f25d
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
1 changed files with 12 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

16
ggml-cuda.cu
View file

@ -6385,8 +6385,11 @@ inline void ggml_cuda_op_mul_mat_cublas(
} }
const half * src1_ptr = src1->type == GGML_TYPE_F16 ? (const half *) src1_ddq_i : src1_as_f16; const half * src1_ptr = src1->type == GGML_TYPE_F16 ? (const half *) src1_ddq_i : src1_as_f16;
const float alpha = 1.0f; size_t dst_as = 0;
const float beta = 0.0f; half * dst_f16 = (half *) ggml_cuda_pool_malloc(row_diff*src1_ncols * sizeof(half), &dst_as);
const half alpha = 1.0f;
const half beta = 0.0f;
CUBLAS_CHECK(cublasSetStream(g_cublas_handles[id], stream)); CUBLAS_CHECK(cublasSetStream(g_cublas_handles[id], stream));
CUBLAS_CHECK( CUBLAS_CHECK(
@ -6394,10 +6397,15 @@ inline void ggml_cuda_op_mul_mat_cublas(
row_diff, src1_ncols, ne10, row_diff, src1_ncols, ne10,
&alpha, src0_ptr, CUDA_R_16F, ne00, &alpha, src0_ptr, CUDA_R_16F, ne00,
src1_ptr, CUDA_R_16F, ne10, src1_ptr, CUDA_R_16F, ne10,
&beta, dst_dd_i, CUDA_R_32F, ldc, &beta, dst_f16, CUDA_R_16F, ldc,
CUBLAS_COMPUTE_32F, CUBLAS_COMPUTE_16F,
CUBLAS_GEMM_DEFAULT_TENSOR_OP)); CUBLAS_GEMM_DEFAULT_TENSOR_OP));
const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
to_fp32_cuda(dst_f16, dst_dd_i, row_diff*src1_ncols, stream);
ggml_cuda_pool_free(dst_f16, dst_as);
if (src0_as != 0) { if (src0_as != 0) {
ggml_cuda_pool_free(src0_as_f16, src0_as); ggml_cuda_pool_free(src0_as_f16, src0_as);
} }