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

ggml : store offset as opt arg for ggml_view_xd() operators

Browse source
This commit is contained in:
Georgi Gerganov 2023-06-01 19:21:08 +03:00
parent 7ca81e9e65
commit 94ea9e7bfe
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
1 changed files with 42 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

42
ggml.c
View file

@ -5802,10 +5802,18 @@ struct ggml_tensor * ggml_view_1d(
struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 1, &ne0, (char *) a->data + offset); struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 1, &ne0, (char *) a->data + offset);
ggml_scratch_save(ctx);
struct ggml_tensor * offs = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 2);
memcpy(offs->data, &offset, 2*sizeof(int32_t));
ggml_scratch_load(ctx);
result->op = GGML_OP_VIEW; result->op = GGML_OP_VIEW;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
result->src0 = a; result->src0 = a;
result->src1 = NULL; result->src1 = NULL;
result->opt[0] = offs;
if (is_node) { if (is_node) {
memcpy(result->padding, &offset, sizeof(offset)); memcpy(result->padding, &offset, sizeof(offset));
@ -5834,6 +5842,13 @@ struct ggml_tensor * ggml_view_2d(
struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 2, ne, (char *) a->data + offset); struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 2, ne, (char *) a->data + offset);
ggml_scratch_save(ctx);
struct ggml_tensor * offs = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 2);
memcpy(offs->data, &offset, 2*sizeof(int32_t));
ggml_scratch_load(ctx);
result->nb[1] = nb1; result->nb[1] = nb1;
result->nb[2] = result->nb[1]*ne1; result->nb[2] = result->nb[1]*ne1;
result->nb[3] = result->nb[2]; result->nb[3] = result->nb[2];
@ -5842,6 +5857,7 @@ struct ggml_tensor * ggml_view_2d(
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
result->src0 = a; result->src0 = a;
result->src1 = NULL; result->src1 = NULL;
result->opt[0] = offs;
if (is_node) { if (is_node) {
memcpy(result->padding, &offset, sizeof(offset)); memcpy(result->padding, &offset, sizeof(offset));
@ -5872,6 +5888,13 @@ struct ggml_tensor * ggml_view_3d(
struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 3, ne, (char *) a->data + offset); struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 3, ne, (char *) a->data + offset);
ggml_scratch_save(ctx);
struct ggml_tensor * offs = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 2);
memcpy(offs->data, &offset, 2*sizeof(int32_t));
ggml_scratch_load(ctx);
result->nb[1] = nb1; result->nb[1] = nb1;
result->nb[2] = nb2; result->nb[2] = nb2;
result->nb[3] = result->nb[2]*ne2; result->nb[3] = result->nb[2]*ne2;
@ -5880,6 +5903,7 @@ struct ggml_tensor * ggml_view_3d(
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
result->src0 = a; result->src0 = a;
result->src1 = NULL; result->src1 = NULL;
result->opt[0] = offs;
if (is_node) { if (is_node) {
memcpy(result->padding, &offset, sizeof(offset)); memcpy(result->padding, &offset, sizeof(offset));
@ -5912,6 +5936,13 @@ struct ggml_tensor * ggml_view_4d(
struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 4, ne, (char *) a->data + offset); struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 4, ne, (char *) a->data + offset);
ggml_scratch_save(ctx);
struct ggml_tensor * offs = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 2);
memcpy(offs->data, &offset, 2*sizeof(int32_t));
ggml_scratch_load(ctx);
result->nb[1] = nb1; result->nb[1] = nb1;
result->nb[2] = nb2; result->nb[2] = nb2;
result->nb[3] = nb3; result->nb[3] = nb3;
@ -5920,6 +5951,7 @@ struct ggml_tensor * ggml_view_4d(
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL; result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
result->src0 = a; result->src0 = a;
result->src1 = NULL; result->src1 = NULL;
result->opt[0] = offs;
if (is_node) { if (is_node) {
memcpy(result->padding, &offset, sizeof(offset)); memcpy(result->padding, &offset, sizeof(offset));
@ -15021,6 +15053,16 @@ struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context **
{ {
tensor = ggml_reshape_4d(*ctx_eval, args[0], ne[0], ne[1], ne[2], ne[3]); tensor = ggml_reshape_4d(*ctx_eval, args[0], ne[0], ne[1], ne[2], ne[3]);
} break; } break;
case GGML_OP_VIEW:
{
tensor = ggml_view_4d(*ctx_eval, args[0], ne[0], ne[1], ne[2], ne[3], 0, 0, 0, 0);
uint64_t offs;
memcpy(&offs, args[2]->data, sizeof(offs));
tensor->data = ((char *) tensor->data) + offs;
printf("xxxxxx offs: %zu\n", offs);
} break;
case GGML_OP_TRANSPOSE: case GGML_OP_TRANSPOSE:
{ {
tensor = ggml_transpose(*ctx_eval, args[0]); tensor = ggml_transpose(*ctx_eval, args[0]);