vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-17 04:19:40 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'upstream' into concedo_experimental

Browse source 
# Conflicts:
#	README.md
#	docs/build.md
#	docs/development/HOWTO-add-model.md
#	tests/test-backend-ops.cpp
#	tests/test-chat-template.cpp
This commit is contained in:
Concedo 2025-01-10 17:57:38 +08:00
commit b154bd3671
50 changed files with 189817 additions and 187510 deletions
Download patch file Download diff file Expand all files Collapse all files

61
ggml/src/ggml.c
View file

@ -981,6 +981,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"GET_REL_POS",
"ADD_REL_POS",
"RWKV_WKV6",
"GATED_LINEAR_ATTN",
"UNARY",
@ -1000,7 +1001,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
"OPT_STEP_ADAMW",
};
static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
static_assert(GGML_OP_COUNT == 83, "GGML_OP_COUNT != 83");
static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"none",
@ -1077,6 +1078,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"get_rel_pos(x)",
"add_rel_pos(x)",
"rwkv_wkv6(k, v, r, tf, td, s)",
"gated_linear_attn(k, v, q, gate, s)",
"unary(x)",
@ -1096,7 +1098,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"adamw(x)",
};
static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
static_assert(GGML_OP_COUNT == 83, "GGML_OP_COUNT != 83");
static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
@ -4642,15 +4644,13 @@ struct ggml_tensor * ggml_rwkv_wkv6(
GGML_ASSERT(ggml_is_contiguous(state));
const int64_t S = k->ne[0];
const int64_t H = k->ne[2];
const int64_t n_tokens = k->ne[3];
const int64_t H = k->ne[1];
const int64_t n_tokens = k->ne[2];
const int64_t n_seqs = state->ne[1];
{
GGML_ASSERT(k->ne[1] == 1);
GGML_ASSERT(v->ne[0] == 1 && v->ne[1] == S && v->ne[2] == H && v->ne[3] == n_tokens);
GGML_ASSERT(r->ne[0] == 1 && r->ne[1] == S && r->ne[2] == H && r->ne[3] == n_tokens);
// TODO: RWKV v4 and v5
GGML_ASSERT(td->ne[0] == 1 && td->ne[1] == S && td->ne[2] == H && td->ne[3] == n_tokens);
GGML_ASSERT(v->ne[0] == S && v->ne[1] == H && v->ne[2] == n_tokens);
GGML_ASSERT(r->ne[0] == S && r->ne[1] == H && r->ne[2] == n_tokens);
GGML_ASSERT(td->ne[0] == S && td->ne[1] == H && td->ne[2] == n_tokens);
GGML_ASSERT(ggml_nelements(state) == S * S * H * n_seqs);
}
@ -4669,6 +4669,49 @@ struct ggml_tensor * ggml_rwkv_wkv6(
return result;
}
// ggml_gated_linear_attn
struct ggml_tensor * ggml_gated_linear_attn(
struct ggml_context * ctx,
struct ggml_tensor * k,
struct ggml_tensor * v,
struct ggml_tensor * q,
struct ggml_tensor * g,
struct ggml_tensor * state,
float scale) {
GGML_ASSERT(ggml_is_contiguous(k));
GGML_ASSERT(ggml_is_contiguous(v));
GGML_ASSERT(ggml_is_contiguous(q));
GGML_ASSERT(ggml_is_contiguous(g));
GGML_ASSERT(ggml_is_contiguous(state));
const int64_t S = k->ne[0];
const int64_t H = k->ne[1];
const int64_t n_tokens = k->ne[2];
const int64_t n_seqs = state->ne[1];
{
GGML_ASSERT(v->ne[0] == S && v->ne[1] == H && v->ne[2] == n_tokens);
GGML_ASSERT(q->ne[0] == S && q->ne[1] == H && q->ne[2] == n_tokens);
GGML_ASSERT(g->ne[0] == S && g->ne[1] == H && g->ne[2] == n_tokens);
GGML_ASSERT(ggml_nelements(state) == S * S * H * n_seqs);
}
// concat output and new_state
const int64_t ne[4] = { S * H, n_tokens + S * n_seqs, 1, 1 };
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
ggml_set_op_params_f32(result, 0, scale);
result->op = GGML_OP_GATED_LINEAR_ATTN;
result->src[0] = k;
result->src[1] = v;
result->src[2] = q;
result->src[3] = g;
result->src[4] = state;
return result;
}
// ggml_unary
static struct ggml_tensor * ggml_unary_impl(