CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-17 11:37:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

ggml : add ops SOFTPLUS, EXPM1, TRI, SOLVE_TRI, CUMSUM (#17063)

Browse Source 
* Add ops needed for new hybrid models: SOFTPLUS, EXPM1, TRI, SOLVE_TRI, CUMSUM

* Update ggml/include/ggml.h

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* Update tests/test-backend-ops.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* Code review

* Whitespace

* Update tests/test-backend-ops.cpp

Co-authored-by: Diego Devesa <slarengh@gmail.com>

* This is actually sigmoid, duh.

* Add CONST, remove TRI_KEEP, other changes from review

* Update tests/test-backend-ops.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* Update ggml/src/ggml.c

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* Update ggml/src/ggml.c

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* Update ggml/src/ggml-cuda/unary.cu

Co-authored-by: Aman Gupta <amangupta052@gmail.com>

* Remove extra script

* Update ggml/src/ggml.c

Co-authored-by: Diego Devesa <slarengh@gmail.com>

* Update tests/test-backend-ops.cpp

Co-authored-by: Diego Devesa <slarengh@gmail.com>

* moving changes from laptop [no ci]

* pre-rebase

* Update tests/test-backend-ops.cpp

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* Update tests/test-backend-ops.cpp

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

* Refactor tests

* ggml : cleanup

* cont : fix ggml_fill srcs

* tests : add note

* ggml : add ggml_fill_inplace

* ggml : add asserts

* ggml : fix ggml_fill constant cast

* cont : ggml_tri minor

* Use TENSOR_LOCALS

* Fix regression from #14596, regenerate

* Don't make commits at night...

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
Co-authored-by: Diego Devesa <slarengh@gmail.com>
Co-authored-by: Aman Gupta <amangupta052@gmail.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
This commit is contained in:
Piotr Wilkin (ilintar)
2025-11-13 19:54:47 +01:00
committed by GitHub
parent a19bd6f7ce
commit 389ac78b26
16 changed files with 33051 additions and 12051 deletions
Download Patch File Download Diff File Expand all files Collapse all files

159
ggml/src/ggml.c
View File

@@ -935,6 +935,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"COS",
"SUM",
"SUM_ROWS",
"CUMSUM",
"MEAN",
"ARGMAX",
"COUNT_EQUAL",
@@ -990,6 +991,8 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"TIMESTEP_EMBEDDING",
"ARGSORT",
"LEAKY_RELU",
"TRI",
"FILL",
"FLASH_ATTN_EXT",
"FLASH_ATTN_BACK",
@@ -1002,6 +1005,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"RWKV_WKV6",
"GATED_LINEAR_ATTN",
"RWKV_WKV7",
"SOLVE_TRI",
"UNARY",
@@ -1019,7 +1023,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"GLU",
};
static_assert(GGML_OP_COUNT == 90, "GGML_OP_COUNT != 90");
static_assert(GGML_OP_COUNT == 94, "GGML_OP_COUNT != 94");
static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"none",
@@ -1039,6 +1043,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"cos(x)",
"Σx",
"Σx_k",
"cumsum(x)",
"Σx/n",
"argmax(x)",
"count_equal(x)",
@@ -1094,6 +1099,8 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"timestep_embedding(timesteps, dim, max_period)",
"argsort(x)",
"leaky_relu(x)",
"tri(x)",
"fill(x, c)",
"flash_attn_ext(x)",
"flash_attn_back(x)",
@@ -1106,6 +1113,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"rwkv_wkv6(k, v, r, tf, td, s)",
"gated_linear_attn(k, v, q, gate, s)",
"rwkv_wkv7(r, w, k, v, a, b, s)",
"A X = B, A triangular, solve X",
"unary(x)",
@@ -1123,7 +1131,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"glu(x)",
};
static_assert(GGML_OP_COUNT == 90, "GGML_OP_COUNT != 90");
static_assert(GGML_OP_COUNT == 94, "GGML_OP_COUNT != 94");
static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
@@ -1142,6 +1150,8 @@ static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = {
"HARDSWISH",
"HARDSIGMOID",
"EXP",
"EXPM1",
"SOFTPLUS",
"GELU_ERF",
"XIELU",
"FLOOR",
@@ -1150,7 +1160,7 @@ static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = {
"TRUNC",
};
static_assert(GGML_UNARY_OP_COUNT == 20, "GGML_UNARY_OP_COUNT != 20");
static_assert(GGML_UNARY_OP_COUNT == 22, "GGML_UNARY_OP_COUNT != 22");
static const char * GGML_GLU_OP_NAME[GGML_GLU_OP_COUNT] = {
"REGLU",
@@ -2258,6 +2268,30 @@ struct ggml_tensor * ggml_log_inplace(
return ggml_log_impl(ctx, a, true);
}
struct ggml_tensor * ggml_expm1(
struct ggml_context * ctx,
struct ggml_tensor * a) {
return ggml_unary(ctx, a, GGML_UNARY_OP_EXPM1);
}
struct ggml_tensor * ggml_expm1_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a) {
return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_EXPM1);
}
struct ggml_tensor * ggml_softplus(
struct ggml_context * ctx,
struct ggml_tensor * a) {
return ggml_unary(ctx, a, GGML_UNARY_OP_SOFTPLUS);
}
struct ggml_tensor * ggml_softplus_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a) {
return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_SOFTPLUS);
}
// ggml_sin
static struct ggml_tensor * ggml_sin_impl(
@@ -2341,6 +2375,21 @@ struct ggml_tensor * ggml_sum_rows(
return result;
}
// ggml_cumsum
struct ggml_tensor * ggml_cumsum(
struct ggml_context * ctx,
struct ggml_tensor * a) {
GGML_ASSERT(a->type == GGML_TYPE_F32);
struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
result->op = GGML_OP_CUMSUM;
result->src[0] = a;
return result;
}
// ggml_mean
struct ggml_tensor * ggml_mean(
@@ -2668,8 +2717,8 @@ struct ggml_tensor * ggml_xielu(
struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
ggml_set_op_params_i32(result, 0, (int32_t) GGML_UNARY_OP_XIELU);
ggml_set_op_params_f32(result, 1, beta + ggml_softplus(alpha_n));
ggml_set_op_params_f32(result, 2, ggml_softplus(alpha_p));
ggml_set_op_params_f32(result, 1, beta + ggml_compute_softplus_f32(alpha_n));
ggml_set_op_params_f32(result, 2, ggml_compute_softplus_f32(alpha_p));
ggml_set_op_params_f32(result, 3, beta);
ggml_set_op_params_f32(result, 4, eps);
@@ -5028,6 +5077,61 @@ struct ggml_tensor * ggml_timestep_embedding(
return result;
}
// ggml_tri
struct ggml_tensor * ggml_tri(
struct ggml_context * ctx,
struct ggml_tensor * a,
enum ggml_tri_type type) {
GGML_ASSERT(a->type == GGML_TYPE_F32);
GGML_ASSERT(ggml_is_contiguous(a));
GGML_ASSERT(a->ne[0] == a->ne[1]);
struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
ggml_set_op_params_i32(result, 0, type);
result->op = GGML_OP_TRI;
result->src[0] = a;
return result;
}
// ggml_fill
static struct ggml_tensor * ggml_fill_impl(
struct ggml_context * ctx,
struct ggml_tensor * a,
float c,
bool inplace) {
GGML_ASSERT(a->type == GGML_TYPE_F32);
GGML_ASSERT(ggml_is_contiguous(a));
struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
ggml_set_op_params_f32(result, 0, c);
result->op = GGML_OP_FILL;
result->src[0] = a;
return result;
}
struct ggml_tensor * ggml_fill(
struct ggml_context * ctx,
struct ggml_tensor * a,
float c) {
return ggml_fill_impl(ctx, a, c, false);
}
struct ggml_tensor * ggml_fill_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a,
float c) {
return ggml_fill_impl(ctx, a, c, true);
}
// ggml_argsort
struct ggml_tensor * ggml_argsort(
@@ -5882,6 +5986,41 @@ struct ggml_tensor * ggml_opt_step_sgd(
return result;
}
// solve_tri
struct ggml_tensor * ggml_solve_tri(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
bool left,
bool lower,
bool uni) {
GGML_ASSERT(a->type == GGML_TYPE_F32);
GGML_ASSERT(b->type == GGML_TYPE_F32);
// A must be square and lower diagonal
GGML_ASSERT(a->ne[0] == a->ne[1]);
// B must have same outer dimension as A
GGML_ASSERT(a->ne[1] == b->ne[1]);
// batch dimensions must be equal
GGML_ASSERT(a->ne[2] == b->ne[2]);
GGML_ASSERT(a->ne[3] == b->ne[3]);
GGML_ASSERT(ggml_is_contiguous(a));
GGML_ASSERT(ggml_is_contiguous(b));
GGML_ASSERT(lower && left && !uni); // TODO: support other variants
struct ggml_tensor * result = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, b->ne[0], b->ne[1], b->ne[2], b->ne[3]);
result->op = GGML_OP_SOLVE_TRI;
result->src[0] = a;
result->src[1] = b;
return result;
}
////////////////////////////////////////////////////////////////////////////////
struct ggml_hash_set ggml_hash_set_new(size_t size) {
@@ -6454,6 +6593,16 @@ static void ggml_compute_backward(
ggml_add_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, tensor, grad));
}
} break;
case GGML_UNARY_OP_EXPM1: {
if (src0_needs_grads) {
ggml_add_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, grad, ggml_exp(ctx, src0)));
}
} break;
case GGML_UNARY_OP_SOFTPLUS: {
if (src0_needs_grads) {
ggml_add_or_set(ctx, cgraph, isrc0, ggml_mul(ctx, grad, ggml_sigmoid(ctx, src0)));
}
} break;
default: {
fprintf(stderr, "%s: unsupported unary op for backward pass: %s\n",
__func__, ggml_unary_op_name(ggml_get_unary_op(tensor)));