ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

ggml : add more generic custom op, remove deprecated custom ops (ggml/1183)

Browse source 
* ggml : add more generic ggml_custom op

* ggml : remove deprecated custom ops
This commit is contained in:
Diego Devesa 2025-04-09 12:31:34 +02:00 • committed by Georgi Gerganov
parent e4bf72d631
commit 459895c326
6 changed files with 132 additions and 485 deletions
Download patch file Download diff file Expand all files Collapse all files

55
ggml/src/ggml-cpu/ggml-cpu.c
View file

@ -2027,41 +2027,6 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
{
ggml_compute_forward_rwkv_wkv7(params, tensor);
} break;
case GGML_OP_MAP_UNARY:
{
ggml_unary_op_f32_t fun;
memcpy(&fun, tensor->op_params, sizeof(fun));
ggml_compute_forward_map_unary(params, tensor, fun);
}
break;
case GGML_OP_MAP_BINARY:
{
ggml_binary_op_f32_t fun;
memcpy(&fun, tensor->op_params, sizeof(fun));
ggml_compute_forward_map_binary(params, tensor, fun);
}
break;
case GGML_OP_MAP_CUSTOM1_F32:
{
ggml_custom1_op_f32_t fun;
memcpy(&fun, tensor->op_params, sizeof(fun));
ggml_compute_forward_map_custom1_f32(params, tensor, fun);
}
break;
case GGML_OP_MAP_CUSTOM2_F32:
{
ggml_custom2_op_f32_t fun;
memcpy(&fun, tensor->op_params, sizeof(fun));
ggml_compute_forward_map_custom2_f32(params, tensor, fun);
}
break;
case GGML_OP_MAP_CUSTOM3_F32:
{
ggml_custom3_op_f32_t fun;
memcpy(&fun, tensor->op_params, sizeof(fun));
ggml_compute_forward_map_custom3_f32(params, tensor, fun);
}
break;
case GGML_OP_MAP_CUSTOM1:
{
ggml_compute_forward_map_custom1(params, tensor);
@ -2077,6 +2042,11 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
ggml_compute_forward_map_custom3(params, tensor);
}
break;
case GGML_OP_CUSTOM:
{
ggml_compute_forward_custom(params, tensor);
}
break;
case GGML_OP_CROSS_ENTROPY_LOSS:
{
ggml_compute_forward_cross_entropy_loss(params, tensor);
@ -2328,11 +2298,6 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
case GGML_OP_WIN_PART:
case GGML_OP_WIN_UNPART:
case GGML_OP_GET_REL_POS:
case GGML_OP_MAP_UNARY:
case GGML_OP_MAP_BINARY:
case GGML_OP_MAP_CUSTOM1_F32:
case GGML_OP_MAP_CUSTOM2_F32:
case GGML_OP_MAP_CUSTOM3_F32:
{
n_tasks = 1;
} break;
@ -2366,6 +2331,16 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
n_tasks = MIN(p.n_tasks, n_threads);
}
} break;
case GGML_OP_CUSTOM:
{
struct ggml_custom_op_params p;
memcpy(&p, node->op_params, sizeof(p));
if (p.n_tasks == GGML_N_TASKS_MAX) {
n_tasks = n_threads;
} else {
n_tasks = MIN(p.n_tasks, n_threads);
}
} break;
case GGML_OP_CROSS_ENTROPY_LOSS:
case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
case GGML_OP_OPT_STEP_ADAMW:

158
ggml/src/ggml-cpu/ops.cpp
View file

@ -8268,152 +8268,6 @@ void ggml_compute_forward_rwkv_wkv7(
}
}
// ggml_compute_forward_map_unary
static void ggml_compute_forward_map_unary_f32(
const ggml_compute_params * params,
ggml_tensor * dst,
const ggml_unary_op_f32_t fun) {
const ggml_tensor * src0 = dst->src[0];
if (params->ith != 0) {
return;
}
assert(ggml_is_contiguous_1(src0));
assert(ggml_is_contiguous_1(dst));
assert(ggml_are_same_shape(src0, dst));
const int n = ggml_nrows(src0);
const int nc = src0->ne[0];
for (int i = 0; i < n; i++) {
fun(nc,
(float *) ((char *) dst->data + i*( dst->nb[1])),
(float *) ((char *) src0->data + i*(src0->nb[1])));
}
}
void ggml_compute_forward_map_unary(
const ggml_compute_params * params,
ggml_tensor * dst,
const ggml_unary_op_f32_t fun) {
const ggml_tensor * src0 = dst->src[0];
switch (src0->type) {
case GGML_TYPE_F32:
{
ggml_compute_forward_map_unary_f32(params, dst, fun);
} break;
default:
{
GGML_ABORT("fatal error");
}
}
}
// ggml_compute_forward_map_binary
static void ggml_compute_forward_map_binary_f32(
const ggml_compute_params * params,
ggml_tensor * dst,
const ggml_binary_op_f32_t fun) {
const ggml_tensor * src0 = dst->src[0];
const ggml_tensor * src1 = dst->src[1];
if (params->ith != 0) {
return;
}
assert(ggml_is_contiguous_1(src0));
assert(ggml_is_contiguous_1(src1));
assert(ggml_is_contiguous_1(dst));
assert(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
const int n = ggml_nrows(src0);
const int nc = src0->ne[0];
for (int i = 0; i < n; i++) {
fun(nc,
(float *) ((char *) dst->data + i*( dst->nb[1])),
(float *) ((char *) src0->data + i*(src0->nb[1])),
(float *) ((char *) src1->data + i*(src1->nb[1])));
}
}
void ggml_compute_forward_map_binary(
const ggml_compute_params * params,
ggml_tensor * dst,
const ggml_binary_op_f32_t fun) {
const ggml_tensor * src0 = dst->src[0];
switch (src0->type) {
case GGML_TYPE_F32:
{
ggml_compute_forward_map_binary_f32(params, dst, fun);
} break;
default:
{
GGML_ABORT("fatal error");
}
}
}
// ggml_compute_forward_map_custom1
void ggml_compute_forward_map_custom1_f32(
const ggml_compute_params * params,
ggml_tensor * dst,
const ggml_custom1_op_f32_t fun) {
const ggml_tensor * a = dst->src[0];
if (params->ith != 0) {
return;
}
fun(dst, a);
}
// ggml_compute_forward_map_custom2
void ggml_compute_forward_map_custom2_f32(
const ggml_compute_params * params,
ggml_tensor * dst,
const ggml_custom2_op_f32_t fun) {
const ggml_tensor * a = dst->src[0];
const ggml_tensor * b = dst->src[1];
if (params->ith != 0) {
return;
}
fun(dst, a, b);
}
// ggml_compute_forward_map_custom3
void ggml_compute_forward_map_custom3_f32(
const ggml_compute_params * params,
ggml_tensor * dst,
const ggml_custom3_op_f32_t fun) {
const ggml_tensor * a = dst->src[0];
const ggml_tensor * b = dst->src[1];
const ggml_tensor * c = dst->src[1];
if (params->ith != 0) {
return;
}
fun(dst, a, b, c);
}
// ggml_compute_forward_map_custom1
void ggml_compute_forward_map_custom1(
@ -8459,6 +8313,18 @@ void ggml_compute_forward_map_custom3(
p.fun(dst, a, b, c, params->ith, params->nth, p.userdata);
}
// ggml_compute_forward_custom
void ggml_compute_forward_custom(
const struct ggml_compute_params * params,
struct ggml_tensor * dst) {
struct ggml_custom_op_params p;
memcpy(&p, dst->op_params, sizeof(p));
p.fun(dst, params->ith, params->nth, p.userdata);
}
// ggml_compute_forward_cross_entropy_loss
static void ggml_compute_forward_cross_entropy_loss_f32(

21
ggml/src/ggml-cpu/ops.h
View file

@ -96,29 +96,10 @@ void ggml_compute_forward_add_rel_pos(const struct ggml_compute_params * params,
void ggml_compute_forward_rwkv_wkv6(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_rwkv_wkv7(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_gla(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_map_unary(
const struct ggml_compute_params * params,
struct ggml_tensor * dst,
const ggml_unary_op_f32_t fun);
void ggml_compute_forward_map_binary(
const struct ggml_compute_params * params,
struct ggml_tensor * dst,
const ggml_binary_op_f32_t fun);
void ggml_compute_forward_map_custom1_f32(
const struct ggml_compute_params * params,
struct ggml_tensor * dst,
const ggml_custom1_op_f32_t fun);
void ggml_compute_forward_map_custom2_f32(
const struct ggml_compute_params * params,
struct ggml_tensor * dst,
const ggml_custom2_op_f32_t fun);
void ggml_compute_forward_map_custom3_f32(
const struct ggml_compute_params * params,
struct ggml_tensor * dst,
const ggml_custom3_op_f32_t fun);
void ggml_compute_forward_map_custom1(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_map_custom2(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_map_custom3(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_custom(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_cross_entropy_loss(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_cross_entropy_loss_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
void ggml_compute_forward_opt_step_adamw(const struct ggml_compute_params * params, struct ggml_tensor * dst);