CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-03 09:22:01 +00:00
Code Issues Packages Projects Releases 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

263
ggml/src/ggml.c
View File

@@ -982,23 +982,18 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"UNARY",
"MAP_UNARY",
"MAP_BINARY",
"MAP_CUSTOM1_F32",
"MAP_CUSTOM2_F32",
"MAP_CUSTOM3_F32",
"MAP_CUSTOM1",
"MAP_CUSTOM2",
"MAP_CUSTOM3",
"CUSTOM",
"CROSS_ENTROPY_LOSS",
"CROSS_ENTROPY_LOSS_BACK",
"OPT_STEP_ADAMW",
};
static_assert(GGML_OP_COUNT == 85, "GGML_OP_COUNT != 85");
static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");
static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"none",
@@ -1081,23 +1076,18 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"unary(x)",
"f(x)",
"f(x,y)",
"custom_f32(x)",
"custom_f32(x,y)",
"custom_f32(x,y,z)",
"map_custom(x)",
"map_custom(x,y)",
"map_custom(x,y,z)",
"custom(x)",
"custom(x,y)",
"custom(x,y,z)",
"cross_entropy_loss(x,y)",
"cross_entropy_loss_back(x,y)",
"adamw(x)",
};
static_assert(GGML_OP_COUNT == 85, "GGML_OP_COUNT != 85");
static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");
static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
@@ -4842,179 +4832,6 @@ struct ggml_tensor * ggml_unary_inplace(
return ggml_unary_impl(ctx, a, op, true);
}
// ggml_map_unary
static struct ggml_tensor * ggml_map_unary_impl_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
const ggml_unary_op_f32_t fun,
bool inplace) {
struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
ggml_set_op_params(result, (const void *) &fun, sizeof(fun));
result->op = GGML_OP_MAP_UNARY;
result->src[0] = a;
return result;
}
struct ggml_tensor * ggml_map_unary_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
const ggml_unary_op_f32_t fun) {
return ggml_map_unary_impl_f32(ctx, a, fun, false);
}
struct ggml_tensor * ggml_map_unary_inplace_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
const ggml_unary_op_f32_t fun) {
return ggml_map_unary_impl_f32(ctx, a, fun, true);
}
// ggml_map_binary
static struct ggml_tensor * ggml_map_binary_impl_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
const ggml_binary_op_f32_t fun,
bool inplace) {
GGML_ASSERT(ggml_are_same_shape(a, b));
struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
ggml_set_op_params(result, (const void *) &fun, sizeof(fun));
result->op = GGML_OP_MAP_BINARY;
result->src[0] = a;
result->src[1] = b;
return result;
}
struct ggml_tensor * ggml_map_binary_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
const ggml_binary_op_f32_t fun) {
return ggml_map_binary_impl_f32(ctx, a, b, fun, false);
}
struct ggml_tensor * ggml_map_binary_inplace_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
const ggml_binary_op_f32_t fun) {
return ggml_map_binary_impl_f32(ctx, a, b, fun, true);
}
// ggml_map_custom1_f32
static struct ggml_tensor * ggml_map_custom1_impl_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
const ggml_custom1_op_f32_t fun,
bool inplace) {
struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
ggml_set_op_params(result, (const void *) &fun, sizeof(fun));
result->op = GGML_OP_MAP_CUSTOM1_F32;
result->src[0] = a;
return result;
}
struct ggml_tensor * ggml_map_custom1_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
const ggml_custom1_op_f32_t fun) {
return ggml_map_custom1_impl_f32(ctx, a, fun, false);
}
struct ggml_tensor * ggml_map_custom1_inplace_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
const ggml_custom1_op_f32_t fun) {
return ggml_map_custom1_impl_f32(ctx, a, fun, true);
}
// ggml_map_custom2_f32
static struct ggml_tensor * ggml_map_custom2_impl_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
const ggml_custom2_op_f32_t fun,
bool inplace) {
struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
ggml_set_op_params(result, (const void *) &fun, sizeof(fun));
result->op = GGML_OP_MAP_CUSTOM2_F32;
result->src[0] = a;
result->src[1] = b;
return result;
}
struct ggml_tensor * ggml_map_custom2_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
const ggml_custom2_op_f32_t fun) {
return ggml_map_custom2_impl_f32(ctx, a, b, fun, false);
}
struct ggml_tensor * ggml_map_custom2_inplace_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
const ggml_custom2_op_f32_t fun) {
return ggml_map_custom2_impl_f32(ctx, a, b, fun, true);
}
// ggml_map_custom3_f32
static struct ggml_tensor * ggml_map_custom3_impl_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
const ggml_custom3_op_f32_t fun,
bool inplace) {
struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
ggml_set_op_params(result, (const void *) &fun, sizeof(fun));
result->op = GGML_OP_MAP_CUSTOM3_F32;
result->src[0] = a;
result->src[1] = b;
result->src[2] = c;
return result;
}
struct ggml_tensor * ggml_map_custom3_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
const ggml_custom3_op_f32_t fun) {
return ggml_map_custom3_impl_f32(ctx, a, b, c, fun, false);
}
struct ggml_tensor * ggml_map_custom3_inplace_f32(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
struct ggml_tensor * c,
const ggml_custom3_op_f32_t fun) {
return ggml_map_custom3_impl_f32(ctx, a, b, c, fun, true);
}
// ggml_map_custom1
static struct ggml_tensor * ggml_map_custom1_impl(
@@ -5033,7 +4850,7 @@ static struct ggml_tensor * ggml_map_custom1_impl(
/*.n_tasks =*/ n_tasks,
/*.userdata =*/ userdata
};
ggml_set_op_params(result, (const void *) &params, sizeof(params));
ggml_set_op_params(result, &params, sizeof(params));
result->op = GGML_OP_MAP_CUSTOM1;
result->src[0] = a;
@@ -5078,7 +4895,7 @@ static struct ggml_tensor * ggml_map_custom2_impl(
/*.n_tasks =*/ n_tasks,
/*.userdata =*/ userdata
};
ggml_set_op_params(result, (const void *) &params, sizeof(params));
ggml_set_op_params(result, &params, sizeof(params));
result->op = GGML_OP_MAP_CUSTOM2;
result->src[0] = a;
@@ -5127,7 +4944,7 @@ static struct ggml_tensor * ggml_map_custom3_impl(
/*.n_tasks =*/ n_tasks,
/*.userdata =*/ userdata
};
ggml_set_op_params(result, (const void *) &params, sizeof(params));
ggml_set_op_params(result, &params, sizeof(params));
result->op = GGML_OP_MAP_CUSTOM3;
result->src[0] = a;
@@ -5159,6 +4976,66 @@ struct ggml_tensor * ggml_map_custom3_inplace(
return ggml_map_custom3_impl(ctx, a, b, c, fun, n_tasks, userdata, true);
}
struct ggml_tensor * ggml_custom_4d(
struct ggml_context * ctx,
enum ggml_type type,
int64_t ne0,
int64_t ne1,
int64_t ne2,
int64_t ne3,
struct ggml_tensor ** args,
int n_args,
ggml_custom_op_t fun,
int n_tasks,
void * userdata) {
GGML_ASSERT(n_args < GGML_MAX_SRC);
struct ggml_tensor * result = ggml_new_tensor_4d(ctx, type, ne0, ne1, ne2, ne3);
struct ggml_custom_op_params params = {
/*.fun =*/ fun,
/*.n_tasks =*/ n_tasks,
/*.userdata =*/ userdata
};
ggml_set_op_params(result, &params, sizeof(params));
result->op = GGML_OP_CUSTOM;
for (int i = 0; i < n_args; i++) {
result->src[i] = args[i];
}
return result;
}
struct ggml_tensor * ggml_custom_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor ** args,
int n_args,
ggml_custom_op_t fun,
int n_tasks,
void * userdata) {
GGML_ASSERT(n_args < GGML_MAX_SRC - 1);
struct ggml_tensor * result = ggml_view_tensor(ctx, a);
struct ggml_custom_op_params params = {
/*.fun =*/ fun,
/*.n_tasks =*/ n_tasks,
/*.userdata =*/ userdata
};
ggml_set_op_params(result, &params, sizeof(params));
result->op = GGML_OP_CUSTOM;
result->src[0] = a;
for (int i = 0; i < n_args; i++) {
result->src[i + 1] = args[i];
}
return result;
}
// ggml_cross_entropy_loss
struct ggml_tensor * ggml_cross_entropy_loss(