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

ggml : reduce memory buffer for F16 mul_mat when not using cuBLAS

Browse source
This commit is contained in:
Georgi Gerganov 2023-04-29 12:41:54 +03:00
parent 150e135858
commit 0ffcd89870
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
1 changed files with 7 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

11
ggml.c
View file

@ -11571,10 +11571,13 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
node->n_tasks = 1; // TODO: this actually is doing nothing
// the threads are still spinning
cur = GGML_TYPE_SIZE[GGML_TYPE_F32]*MAX(ggml_nelements(node->src1), ggml_nelements(node->src0));
//printf("src0: ne0 = %d, ne1 = %d, ne = %d\n", node->src0->ne[0], node->src0->ne[1], node->src0->ne[0]*node->src0->ne[1]);
//printf("src1: ne0 = %d, ne1 = %d, ne = %d\n", node->src1->ne[0], node->src1->ne[1], node->src1->ne[0]*node->src1->ne[1]);
//printf("cur = %zu\n", cur);
#if defined(GGML_USE_CUBLAS)
// with cuBLAS, we need memory for the full 3D / 4D data of src1
cur = GGML_TYPE_SIZE[GGML_TYPE_F16]*ggml_nelements(node->src1);
#else
// here we need memory just for single 2D matrix from src0
cur = GGML_TYPE_SIZE[GGML_TYPE_F32]*(node->src0->ne[0]*node->src0->ne[1]);
#endif
} else {
cur = GGML_TYPE_SIZE[GGML_TYPE_F16]*ggml_nelements(node->src1);
}