mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 30 Packages Projects Releases Wiki Activity

ggml: add GATED_DELTA_NET op (#19504)

Browse Source 
* ggml: add GATED_DELTA_NET op

* remove the transpose

* add KDA

* add qwen35 dense

* llama : check for fused gated delta net backend support

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Aman Gupta
2026-03-07 15:41:10 +08:00
committed by GitHub
parent 6fce5c6a7d
commit c5a778891b
15 changed files with 627 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml.c
+55 -2
View File
@@ -1031,6 +1031,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"GATED_LINEAR_ATTN",
"RWKV_WKV7",
"SOLVE_TRI",
"GATED_DELTA_NET",
"UNARY",
@@ -1048,7 +1049,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"GLU",
};
static_assert(GGML_OP_COUNT == 95, "GGML_OP_COUNT != 95");
static_assert(GGML_OP_COUNT == 96, "GGML_OP_COUNT != 96");
static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"none",
@@ -1140,6 +1141,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"gated_linear_attn(k, v, q, gate, s)",
"rwkv_wkv7(r, w, k, v, a, b, s)",
"A X = B, A triangular, solve X",
"gated_delta_net(q, k, v, g, beta, s)",
"unary(x)",
@@ -1157,7 +1159,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"glu(x)",
};
static_assert(GGML_OP_COUNT == 95, "GGML_OP_COUNT != 95");
static_assert(GGML_OP_COUNT == 96, "GGML_OP_COUNT != 96");
static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
@@ -6124,6 +6126,57 @@ struct ggml_tensor * ggml_solve_tri(
return result;
}
// ggml_gated_delta_net
struct ggml_tensor * ggml_gated_delta_net(
struct ggml_context * ctx,
struct ggml_tensor * q,
struct ggml_tensor * k,
struct ggml_tensor * v,
struct ggml_tensor * g,
struct ggml_tensor * beta,
struct ggml_tensor * state) {
GGML_ASSERT(ggml_is_contiguous_rows(q));
GGML_ASSERT(ggml_is_contiguous_rows(k));
GGML_ASSERT(ggml_is_contiguous_rows(v));
GGML_ASSERT(ggml_is_contiguous(g));
GGML_ASSERT(ggml_is_contiguous(beta));
GGML_ASSERT(ggml_is_contiguous(state));
GGML_ASSERT(q->type == GGML_TYPE_F32);
GGML_ASSERT(k->type == GGML_TYPE_F32);
GGML_ASSERT(v->type == GGML_TYPE_F32);
GGML_ASSERT(g->type == GGML_TYPE_F32);
GGML_ASSERT(beta->type == GGML_TYPE_F32);
GGML_ASSERT(state->type == GGML_TYPE_F32);
const int64_t S_v = v->ne[0];
const int64_t H = v->ne[1];
const int64_t n_tokens = v->ne[2];
const int64_t n_seqs = v->ne[3];
// gate: scalar [1, H, T, B] or vector [S_v, H, T, B] (KDA)
GGML_ASSERT(g->ne[0] == 1 || g->ne[0] == S_v);
GGML_ASSERT(beta->ne[0] == 1);
GGML_ASSERT(ggml_nelements(state) == S_v * S_v * H * n_seqs);
// concat output and new_state into a single tensor
// output: S_v * H * n_tokens * n_seqs, state: S_v * S_v * H * n_seqs
const int64_t ne[4] = { S_v * H, n_tokens * n_seqs + S_v * n_seqs, 1, 1 };
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
result->op = GGML_OP_GATED_DELTA_NET;
result->src[0] = q;
result->src[1] = k;
result->src[2] = v;
result->src[3] = g;
result->src[4] = beta;
result->src[5] = state;
return result;
}
////////////////////////////////////////////////////////////////////////////////
struct ggml_hash_set ggml_hash_set_new(size_t size) {