nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 28 Packages Projects Releases Wiki Activity

ggml : add ggml_top_k (#17365)

Browse Source 
* ggml : add ggml_top_k

* cont : add ggml_argsort_top_k

* metal : add top_k support

* ggml : cleanup

* tests : add virtual err() function for test_case

* ggml : add comments
This commit is contained in:
Georgi Gerganov
2025-11-25 15:31:43 +02:00
committed by GitHub
parent 05872ac885
commit 583cb83416
15 changed files with 516 additions and 85 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml.c
+48 -29
View File
@@ -990,6 +990,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"ARANGE",
"TIMESTEP_EMBEDDING",
"ARGSORT",
"TOP_K",
"LEAKY_RELU",
"TRI",
"FILL",
@@ -1023,7 +1024,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"GLU",
};
static_assert(GGML_OP_COUNT == 94, "GGML_OP_COUNT != 94");
static_assert(GGML_OP_COUNT == 95, "GGML_OP_COUNT != 95");
static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"none",
@@ -1098,6 +1099,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"arange(start, stop, step)",
"timestep_embedding(timesteps, dim, max_period)",
"argsort(x)",
"top_k(x)",
"leaky_relu(x)",
"tri(x)",
"fill(x, c)",
@@ -1131,7 +1133,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"glu(x)",
};
static_assert(GGML_OP_COUNT == 94, "GGML_OP_COUNT != 94");
static_assert(GGML_OP_COUNT == 95, "GGML_OP_COUNT != 95");
static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
@@ -5036,28 +5038,6 @@ struct ggml_tensor * ggml_roll(
return result;
}
// ggml_arange
struct ggml_tensor * ggml_arange(
struct ggml_context * ctx,
float start,
float stop,
float step) {
GGML_ASSERT(stop > start);
const int64_t steps = (int64_t) ceilf((stop - start) / step);
struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, steps);
ggml_set_op_params_f32(result, 0, start);
ggml_set_op_params_f32(result, 1, stop);
ggml_set_op_params_f32(result, 2, step);
result->op = GGML_OP_ARANGE;
return result;
}
// ggml_timestep_embedding
struct ggml_tensor * ggml_timestep_embedding(
@@ -5139,6 +5119,7 @@ struct ggml_tensor * ggml_argsort(
struct ggml_tensor * a,
enum ggml_sort_order order) {
GGML_ASSERT(a->ne[0] <= INT32_MAX);
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_I32, GGML_MAX_DIMS, a->ne);
ggml_set_op_params_i32(result, 0, (int32_t) order);
@@ -5149,6 +5130,24 @@ struct ggml_tensor * ggml_argsort(
return result;
}
// ggml_argsort_top_k
struct ggml_tensor * ggml_argsort_top_k(
struct ggml_context * ctx,
struct ggml_tensor * a,
int k) {
GGML_ASSERT(a->ne[0] >= k);
struct ggml_tensor * result = ggml_argsort(ctx, a, GGML_SORT_ORDER_DESC);
result = ggml_view_4d(ctx, result,
k, result->ne[1], result->ne[2], result->ne[3],
result->nb[1], result->nb[2], result->nb[3],
0);
return result;
}
// ggml_top_k
struct ggml_tensor * ggml_top_k(
@@ -5157,12 +5156,32 @@ struct ggml_tensor * ggml_top_k(
int k) {
GGML_ASSERT(a->ne[0] >= k);
struct ggml_tensor * result = ggml_argsort(ctx, a, GGML_SORT_ORDER_DESC);
struct ggml_tensor * result = ggml_new_tensor_4d(ctx, GGML_TYPE_I32, k, a->ne[1], a->ne[2], a->ne[3]);
result = ggml_view_4d(ctx, result,
k, result->ne[1], result->ne[2], result->ne[3],
result->nb[1], result->nb[2], result->nb[3],
0);
result->op = GGML_OP_TOP_K;
result->src[0] = a;
return result;
}
// ggml_arange
struct ggml_tensor * ggml_arange(
struct ggml_context * ctx,
float start,
float stop,
float step) {
GGML_ASSERT(stop > start);
const int64_t steps = (int64_t) ceilf((stop - start) / step);
struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, steps);
ggml_set_op_params_f32(result, 0, start);
ggml_set_op_params_f32(result, 1, stop);
ggml_set_op_params_f32(result, 2, step);
result->op = GGML_OP_ARANGE;
return result;
}