upstream-archive/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-07-23 11:10:55 -05:00
Code Issues Packages Projects Releases Wiki Activity

DeepseekV4: Add fused hyper-connection ops (#25585)

Browse Source 
* dsv4 hc-ops

* add missing files;

* add cparams

* update rpc version

* address review comments

* address review comments
This commit is contained in:
Aman Gupta
2026-07-17 00:33:33 +08:00
committed by GitHub
parent b2dd28a3b6
commit 0dc74e332e
16 changed files with 1062 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

173
tests/test-backend-ops.cpp
View File

@@ -3751,6 +3751,166 @@ struct test_snake_fuse : public test_case {
}
};
struct test_dsv4_hc : public test_case {
static constexpr int64_t hc = 4;
ggml_tensor * out = nullptr;
static uint32_t tensor_seed(const ggml_tensor * t) {
uint32_t seed = 2166136261u;
for (const char * p = ggml_get_name(t); *p; ++p) {
seed ^= (uint8_t) *p;
seed *= 16777619u;
}
for (int i = 0; i < GGML_MAX_DIMS; ++i) {
seed ^= (uint32_t) t->ne[i];
seed *= 16777619u;
}
return seed;
}
static bool tensor_range(const std::string & name, float & lo, float & hi) {
if (name == "mixes") {
lo = -2.0f; hi = 2.0f; return true;
}
if (name == "scale") {
lo = -0.5f; hi = 0.5f; return true;
}
if (name == "base") {
lo = -0.25f; hi = 0.25f; return true;
}
if (name == "weights" || name == "comb") {
lo = 0.0f; hi = 1.0f; return true;
}
if (name == "post") {
lo = 0.0f; hi = 2.0f; return true;
}
if (name == "x" || name == "residual") {
lo = -1.0f; hi = 1.0f; return true;
}
return false;
}
void initialize_tensors(ggml_context * ctx) override {
for (ggml_tensor * t = ggml_get_first_tensor(ctx); t != nullptr; t = ggml_get_next_tensor(ctx, t)) {
const std::string name = ggml_get_name(t);
float lo;
float hi;
if (!tensor_range(name, lo, hi)) {
continue;
}
GGML_ASSERT(t->type == GGML_TYPE_F32);
std::mt19937 rng(tensor_seed(t));
std::uniform_real_distribution<float> dist(lo, hi);
std::vector<float> data(ggml_nelements(t));
for (float & v : data) {
v = dist(rng);
}
ggml_backend_tensor_set(t, data.data(), 0, data.size()*sizeof(float));
}
}
};
struct test_dsv4_hc_comb : public test_dsv4_hc {
const int64_t n_tokens;
const int32_t n_iter;
const float eps;
std::string op_desc(ggml_tensor * t) override {
GGML_UNUSED(t);
return "DSV4_HC_COMB";
}
std::string vars() override {
return VARS_TO_STR3(n_tokens, n_iter, eps);
}
test_dsv4_hc_comb(int64_t n_tokens = 17, int32_t n_iter = 4, float eps = 1e-6f)
: n_tokens(n_tokens), n_iter(n_iter), eps(eps) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * mixes = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, (2 + hc)*hc, n_tokens);
ggml_set_name(mixes, "mixes");
ggml_tensor * scale = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 3);
ggml_set_name(scale, "scale");
ggml_tensor * base = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, (2 + hc)*hc);
ggml_set_name(base, "base");
out = ggml_dsv4_hc_comb(ctx, mixes, scale, base, eps, n_iter);
ggml_set_name(out, "out");
return out;
}
};
struct test_dsv4_hc_pre : public test_dsv4_hc {
const int64_t n_embd;
const int64_t n_tokens;
std::string op_desc(ggml_tensor * t) override {
GGML_UNUSED(t);
return "DSV4_HC_PRE";
}
std::string vars() override {
return VARS_TO_STR2(n_embd, n_tokens);
}
test_dsv4_hc_pre(int64_t n_embd = 31, int64_t n_tokens = 17)
: n_embd(n_embd), n_tokens(n_tokens) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * x = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, n_embd, hc, n_tokens);
ggml_set_name(x, "x");
ggml_tensor * weights = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, hc, n_tokens);
ggml_set_name(weights, "weights");
out = ggml_dsv4_hc_pre(ctx, x, weights);
ggml_set_name(out, "out");
return out;
}
};
struct test_dsv4_hc_post : public test_dsv4_hc {
const int64_t n_embd;
const int64_t n_tokens;
std::string op_desc(ggml_tensor * t) override {
GGML_UNUSED(t);
return "DSV4_HC_POST";
}
std::string vars() override {
return VARS_TO_STR2(n_embd, n_tokens);
}
test_dsv4_hc_post(int64_t n_embd = 31, int64_t n_tokens = 17)
: n_embd(n_embd), n_tokens(n_tokens) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * x = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_tokens);
ggml_set_name(x, "x");
ggml_tensor * residual = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, n_embd, hc, n_tokens);
ggml_set_name(residual, "residual");
ggml_tensor * post = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, hc, n_tokens);
ggml_set_name(post, "post");
ggml_tensor * comb = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hc, hc, n_tokens);
ggml_set_name(comb, "comb");
out = ggml_dsv4_hc_post(ctx, x, residual, post, comb);
ggml_set_name(out, "out");
return out;
}
};
// GGML_OP_SSM_CONV
struct test_ssm_conv : public test_case {
const ggml_type type;
@@ -7882,6 +8042,19 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_snake_fuse(type, { 64, 32, 2, 3})); // ne[2] > 1 and ne[3] > 1
}
test_cases.emplace_back(new test_dsv4_hc_comb(1, 1));
test_cases.emplace_back(new test_dsv4_hc_comb(17, 4));
test_cases.emplace_back(new test_dsv4_hc_comb(257, 8));
test_cases.emplace_back(new test_dsv4_hc_pre(1, 1));
test_cases.emplace_back(new test_dsv4_hc_pre(31, 17));
test_cases.emplace_back(new test_dsv4_hc_pre(128, 257));
test_cases.emplace_back(new test_dsv4_hc_pre(4096, 21));
test_cases.emplace_back(new test_dsv4_hc_post(1, 1));
test_cases.emplace_back(new test_dsv4_hc_post(31, 17));
test_cases.emplace_back(new test_dsv4_hc_post(128, 257));
// glu ops
for (ggml_type type : {GGML_TYPE_F16, GGML_TYPE_F32}) {
for (int v : {0, 1}) {