Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

ggml: add SWIGLU_CLAMP (#27930)

Browse Source 
* ggml: add SWIGLU_CLAMP

* add vulkan shader
This commit is contained in:
Aman Gupta
2026-08-30 23:00:02 +08:00
committed by GitHub
parent 2578138397
commit 0190529ec4
40 changed files with 768 additions and 35 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/test-backend-ops.cpp
+76 -4
View File
@@ -2243,6 +2243,63 @@ struct test_swiglu_oai : public test_case {
}
};
struct test_swiglu_clamp : public test_case {
const ggml_type type;
const std::array<int64_t, 4> ne_a;
int v; // view (1 : non-contiguous a)
float limit;
std::string vars() override {
return VARS_TO_STR4(type, ne_a, v, limit);
}
test_swiglu_clamp(ggml_type type = GGML_TYPE_F32,
std::array<int64_t, 4> ne_a = {128, 2, 2, 2},
int v = 0,
float limit = 7.0f)
: type(type), ne_a(ne_a), v(v), limit(limit) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * a;
ggml_tensor * b;
if (v & 1) {
auto ne = ne_a; ne[0] *= 3;
a = ggml_new_tensor(ctx, type, 4, ne.data());
ggml_set_param(a);
ggml_set_name(a, "a");
a = ggml_view_4d(ctx, a, ne_a[0], ne_a[1], ne_a[2], ne_a[3], a->nb[1], a->nb[2], a->nb[3], 0);
ggml_set_name(a, "view_of_a");
b = ggml_new_tensor(ctx, type, 4, ne.data());
ggml_set_param(b);
ggml_set_name(b, "b");
b = ggml_view_4d(ctx, b, ne_a[0], ne_a[1], ne_a[2], ne_a[3], b->nb[1], b->nb[2], b->nb[3], 0);
ggml_set_name(b, "view_of_b");
} else {
a = ggml_new_tensor(ctx, type, 4, ne_a.data());
ggml_set_param(a);
ggml_set_name(a, "a");
b = ggml_new_tensor(ctx, type, 4, ne_a.data());
ggml_set_param(b);
ggml_set_name(b, "b");
}
ggml_tensor * out = ggml_swiglu_clamp(ctx, a, b, limit);
ggml_set_name(out, "out");
return out;
}
void initialize_tensors(ggml_context * ctx) override {
for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
init_tensor_uniform(t, -150.f, 150.f);
}
}
};
// GGML_OP_GET_ROWS
struct test_get_rows : public test_case {
const ggml_type type;
@@ -6380,6 +6437,9 @@ struct test_mul_mat_vec_fusion : public test_case {
constexpr float alpha = 1.702f;
constexpr float limit = 7.0f;
out = ggml_swiglu_oai(ctx, ffn_gate, ffn_up, alpha, limit);
} else if (glu_op == GGML_GLU_OP_SWIGLU_CLAMP) {
constexpr float limit = 10.0f;
out = ggml_swiglu_clamp(ctx, ffn_gate, ffn_up, limit);
} else {
out = ggml_glu_split(ctx, ffn_gate, ffn_up, glu_op);
}
@@ -8376,8 +8436,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
for (ggml_type type : {GGML_TYPE_F16, GGML_TYPE_F32}) {
for (int v : {0, 1}) {
for (int op = 0; op < GGML_GLU_OP_COUNT; op++) {
if (op == GGML_GLU_OP_SWIGLU_OAI) {
// SWIGLU_OAI is handled separately
if (op == GGML_GLU_OP_SWIGLU_OAI || op == GGML_GLU_OP_SWIGLU_CLAMP) {
continue;
}
@@ -8400,6 +8459,14 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
}
}
for (ggml_type type : {GGML_TYPE_F16, GGML_TYPE_F32}) {
for (int v : {0, 1}) {
for (float limit : {2.0f, 10.0f}) {
test_cases.emplace_back(new test_swiglu_clamp(type, { 128, 2, 2, 2 }, v, limit));
}
}
}
for (ggml_type type : {GGML_TYPE_F32, GGML_TYPE_Q4_0}) {
test_cases.emplace_back(new test_get_rows(type, 300*256, 5, 4, 1, 2, false));
test_cases.emplace_back(new test_get_rows(type, 256, 80000, 70000, 2, 1, false));
@@ -10026,7 +10093,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
if (!with_gate && !with_bias) {
continue;
}
for (ggml_glu_op glu_op : {GGML_GLU_OP_SWIGLU, GGML_GLU_OP_GEGLU}) {
for (ggml_glu_op glu_op : {GGML_GLU_OP_SWIGLU, GGML_GLU_OP_GEGLU, GGML_GLU_OP_SWIGLU_CLAMP}) {
if (!with_bias && glu_op == GGML_GLU_OP_SWIGLU_OAI) {
continue;
}
@@ -10041,7 +10108,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
use_id, 16, 8, b, with_bias, with_gate, with_lane_scale));
test_cases.emplace_back(new test_mul_mat_vec_fusion(type, glu_op, 1, 32, 256,
use_id, 16, 8, b, with_bias, with_gate, with_lane_scale, {1, 1}));
if (!use_id && with_gate && !with_bias) {
if (!use_id && with_gate && !with_bias && glu_op != GGML_GLU_OP_SWIGLU_CLAMP) {
// small multi-token batches (speculative decoding / MTP verify)
for (int64_t m_batch : { 2, 4, 8 }) {
test_cases.emplace_back(new test_mul_mat_vec_fusion(type, glu_op, m_batch, 32, 256,
@@ -10056,6 +10123,11 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
}
}
for (bool b : {false, true}) {
test_cases.emplace_back(new test_mul_mat_vec_fusion(GGML_TYPE_IQ2_S, GGML_GLU_OP_SWIGLU_CLAMP, 1, 32, 256,
true, 16, 8, b, false, true, false));
}
for (auto gate : {GATING_FUNC_SOFTMAX, GATING_FUNC_SIGMOID, GATING_FUNC_SOFTMAX_WEIGHT, GATING_FUNC_SQRT_SOFTPLUS}) {
for (bool with_norm : {false, true}) {
for (bool bias_probs : {false, true}) {