vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

ggml : remove n_dims from ggml_tensor (#4469)

Browse source 
ggml-ci
This commit is contained in:
slaren 2023-12-14 16:52:08 +01:00 • committed by GitHub
parent c50e400163
commit cafcd4f895
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
9 changed files with 81 additions and 73 deletions
Download patch file Download diff file Expand all files Collapse all files

94
ggml.c
View file

@ -2054,24 +2054,37 @@ size_t ggml_element_size(const struct ggml_tensor * tensor) {
return ggml_type_size(tensor->type);
}
static inline bool ggml_is_scalar(const struct ggml_tensor * tensor) {
bool ggml_is_scalar(const struct ggml_tensor * tensor) {
static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
return tensor->ne[0] == 1 && tensor->ne[1] == 1 && tensor->ne[2] == 1 && tensor->ne[3] == 1;
}
static inline bool ggml_is_vector(const struct ggml_tensor * tensor) {
bool ggml_is_vector(const struct ggml_tensor * tensor) {
static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
return tensor->ne[1] == 1 && tensor->ne[2] == 1 && tensor->ne[3] == 1;
}
static inline bool ggml_is_matrix(const struct ggml_tensor * tensor) {
bool ggml_is_matrix(const struct ggml_tensor * tensor) {
static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
return tensor->ne[2] == 1 && tensor->ne[3] == 1;
}
bool ggml_is_3d(const struct ggml_tensor * tensor) {
return tensor->ne[3] == 1;
}
int ggml_n_dims(const struct ggml_tensor * tensor) {
for (int i = GGML_MAX_DIMS - 1; i >= 1; --i) {
if (tensor->ne[i] > 1) {
return i + 1;
}
}
return 1;
}
static inline bool ggml_can_mul_mat(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
@ -2521,7 +2534,6 @@ static struct ggml_tensor * ggml_new_tensor_impl(
/*.type =*/ type,
/*.backend =*/ GGML_BACKEND_CPU,
/*.buffer =*/ NULL,
/*.n_dims =*/ n_dims,
/*.ne =*/ { 1, 1, 1, 1 },
/*.nb =*/ { 0, 0, 0, 0 },
/*.op =*/ GGML_OP_NONE,
@ -2628,7 +2640,7 @@ struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value) {
}
struct ggml_tensor * ggml_dup_tensor(struct ggml_context * ctx, const struct ggml_tensor * src) {
return ggml_new_tensor(ctx, src->type, src->n_dims, src->ne);
return ggml_new_tensor(ctx, src->type, GGML_MAX_DIMS, src->ne);
}
static void ggml_set_op_params(struct ggml_tensor * tensor, const void * params, size_t params_size) {
@ -3077,7 +3089,7 @@ struct ggml_tensor * ggml_format_name(struct ggml_tensor * tensor, const char *
struct ggml_tensor * ggml_view_tensor(
struct ggml_context * ctx,
struct ggml_tensor * src) {
struct ggml_tensor * result = ggml_new_tensor_impl(ctx, src->type, src->n_dims, src->ne, src, 0);
struct ggml_tensor * result = ggml_new_tensor_impl(ctx, src->type, GGML_MAX_DIMS, src->ne, src, 0);
ggml_format_name(result, "%s (view)", src->name);
for (int i = 0; i < GGML_MAX_DIMS; i++) {
@ -3235,10 +3247,10 @@ static struct ggml_tensor * ggml_add_cast_impl(
is_node = true;
}
struct ggml_tensor * result = ggml_new_tensor(ctx, type, a->n_dims, a->ne);
struct ggml_tensor * result = ggml_new_tensor(ctx, type, GGML_MAX_DIMS, a->ne);
result->op = GGML_OP_ADD;
result->grad = is_node ? ggml_new_tensor(ctx, GGML_TYPE_F32, a->n_dims, a->ne) : NULL;
result->grad = is_node ? ggml_new_tensor(ctx, GGML_TYPE_F32, GGML_MAX_DIMS, a->ne) : NULL;
result->src[0] = a;
result->src[1] = b;
@ -3607,12 +3619,12 @@ struct ggml_tensor * ggml_sum_rows(
is_node = true;
}
int64_t ne[4] = {1,1,1,1};
for (int i=1; i<a->n_dims; ++i) {
int64_t ne[GGML_MAX_DIMS] = { 1 };
for (int i = 1; i < GGML_MAX_DIMS; ++i) {
ne[i] = a->ne[i];
}
struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, a->n_dims, ne);
struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, GGML_MAX_DIMS, ne);
result->op = GGML_OP_SUM_ROWS;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@ -3633,8 +3645,8 @@ struct ggml_tensor * ggml_mean(
is_node = true;
}
int64_t ne[GGML_MAX_DIMS] = { 1, a->ne[1], a->ne[2], a->ne[3] };
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, a->n_dims, ne);
int64_t ne[4] = { 1, a->ne[1], a->ne[2], a->ne[3] };
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
result->op = GGML_OP_MEAN;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@ -3656,8 +3668,7 @@ struct ggml_tensor * ggml_argmax(
is_node = true;
}
int64_t ne[GGML_MAX_DIMS] = { a->ne[1], 1, 1, 1 };
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_I32, a->n_dims, ne);
struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, a->ne[1]);
result->op = GGML_OP_ARGMAX;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@ -3680,7 +3691,7 @@ struct ggml_tensor * ggml_repeat(
is_node = true;
}
struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, b->n_dims, b->ne);
struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, GGML_MAX_DIMS, b->ne);
result->op = GGML_OP_REPEAT;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@ -3707,7 +3718,7 @@ struct ggml_tensor * ggml_repeat_back(
return a;
}
struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, b->n_dims, b->ne);
struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, GGML_MAX_DIMS, b->ne);
result->op = GGML_OP_REPEAT_BACK;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@ -4083,7 +4094,7 @@ struct ggml_tensor * ggml_mul_mat(
}
const int64_t ne[4] = { a->ne[1], b->ne[1], b->ne[2], b->ne[3] };
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, MAX(a->n_dims, b->n_dims), ne);
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
result->op = GGML_OP_MUL_MAT;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@ -4117,7 +4128,7 @@ struct ggml_tensor * ggml_mul_mat_id(
}
const int64_t ne[4] = { as[0]->ne[1], b->ne[1], b->ne[2], b->ne[3] };
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, MAX(as[0]->n_dims, b->n_dims), ne);
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
ggml_set_op_params_i32(result, 0, id);
ggml_set_op_params_i32(result, 1, n_as);
@ -4155,7 +4166,7 @@ struct ggml_tensor * ggml_out_prod(
// a is broadcastable to b for ne[2] and ne[3] -> use b->ne[2] and b->ne[3]
const int64_t ne[4] = { a->ne[0], b->ne[0], b->ne[2], b->ne[3] };
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, MAX(a->n_dims, b->n_dims), ne);
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
result->op = GGML_OP_OUT_PROD;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@ -4440,7 +4451,7 @@ struct ggml_tensor * ggml_reshape(
//GGML_ASSERT(false);
}
struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, b->n_dims, b->ne, a, 0);
struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, GGML_MAX_DIMS, b->ne, a, 0);
ggml_format_name(result, "%s (reshaped)", a->name);
result->op = GGML_OP_RESHAPE;
@ -4818,7 +4829,7 @@ struct ggml_tensor * ggml_diag(
}
const int64_t ne[4] = { a->ne[0], a->ne[0], a->ne[2], a->ne[3] };
struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, MAX(a->n_dims, 2), ne);
struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, 4, ne);
result->op = GGML_OP_DIAG;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@ -5465,7 +5476,7 @@ struct ggml_tensor * ggml_pool_1d(
is_node = true;
}
const int64_t ne[3] = {
const int64_t ne[2] = {
ggml_calc_pool_output_size(a->ne[0], k0, s0, p0),
a->ne[1],
};
@ -5584,7 +5595,7 @@ struct ggml_tensor * ggml_argsort(
enum ggml_sort_order order) {
bool is_node = false;
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_I32, a->n_dims, a->ne);
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_I32, GGML_MAX_DIMS, a->ne);
ggml_set_op_params_i32(result, 0, (int32_t) order);
@ -5631,7 +5642,7 @@ struct ggml_tensor * ggml_flash_attn(
}
//struct ggml_tensor * result = ggml_dup_tensor(ctx, q);
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, q->n_dims, q->ne);
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, GGML_MAX_DIMS, q->ne);
int32_t t = masked ? 1 : 0;
ggml_set_op_params(result, &t, sizeof(t));
@ -5664,7 +5675,7 @@ struct ggml_tensor * ggml_flash_ff(
}
//struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, a->n_dims, a->ne);
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, GGML_MAX_DIMS, a->ne);
result->op = GGML_OP_FLASH_FF;
result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@ -5780,7 +5791,6 @@ struct ggml_tensor * ggml_win_part(
const int np = npx*npy;
const int64_t ne[4] = { a->ne[0], w, w, np, };
struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
int32_t params[] = { npx, npy, w };
@ -14563,7 +14573,7 @@ static struct ggml_tensor * ggml_recompute_graph_node(
return replacements->vals[i];
}
struct ggml_tensor * clone = ggml_new_tensor(ctx, node->type, node->n_dims, node->ne);
struct ggml_tensor * clone = ggml_new_tensor(ctx, node->type, GGML_MAX_DIMS, node->ne);
// insert clone into replacements
GGML_ASSERT(replacements->set.keys[i] == NULL); // assert that we don't overwrite
@ -16564,7 +16574,7 @@ static void ggml_graph_export_leaf(const struct ggml_tensor * tensor, FILE * fou
fprintf(fout, "%-6s %-12s %8d %" PRId64 " %" PRId64 " %" PRId64 " %" PRId64 " %16zu %16zu %16zu %16zu %16p %32s\n",
ggml_type_name(tensor->type),
ggml_op_name (tensor->op),
tensor->n_dims,
ggml_n_dims(tensor),
ne[0], ne[1], ne[2], ne[3],
nb[0], nb[1], nb[2], nb[3],
tensor->data,
@ -16579,7 +16589,7 @@ static void ggml_graph_export_node(const struct ggml_tensor * tensor, const char
arg,
ggml_type_name(tensor->type),
ggml_op_name (tensor->op),
tensor->n_dims,
ggml_n_dims(tensor),
ne[0], ne[1], ne[2], ne[3],
nb[0], nb[1], nb[2], nb[3],
tensor->data,
@ -16669,11 +16679,9 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) {
const uint32_t type = tensor->type;
const uint32_t op = tensor->op;
const uint32_t n_dims = tensor->n_dims;
fwrite(&type, sizeof(uint32_t), 1, fout);
fwrite(&op, sizeof(uint32_t), 1, fout);
fwrite(&n_dims, sizeof(uint32_t), 1, fout);
for (int j = 0; j < GGML_MAX_DIMS; ++j) {
const uint64_t ne = tensor->ne[j];
@ -16703,11 +16711,9 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) {
const uint32_t type = tensor->type;
const uint32_t op = tensor->op;
const uint32_t n_dims = tensor->n_dims;
fwrite(&type, sizeof(uint32_t), 1, fout);
fwrite(&op, sizeof(uint32_t), 1, fout);
fwrite(&n_dims, sizeof(uint32_t), 1, fout);
for (int j = 0; j < GGML_MAX_DIMS; ++j) {
const uint64_t ne = tensor->ne[j];
@ -16879,12 +16885,10 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
{
uint32_t type;
uint32_t op;
uint32_t n_dims;
for (uint32_t i = 0; i < n_leafs; ++i) {
type = *(const uint32_t *) ptr; ptr += sizeof(type);
op = *(const uint32_t *) ptr; ptr += sizeof(op);
n_dims = *(const uint32_t *) ptr; ptr += sizeof(n_dims);
int64_t ne[GGML_MAX_DIMS];
size_t nb[GGML_MAX_DIMS];
@ -16900,7 +16904,7 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
nb[j] = nb_cur;
}
struct ggml_tensor * tensor = ggml_new_tensor(*ctx_eval, (enum ggml_type) type, n_dims, ne);
struct ggml_tensor * tensor = ggml_new_tensor(*ctx_eval, (enum ggml_type) type, GGML_MAX_DIMS, ne);
tensor->op = (enum ggml_op) op;
@ -16917,7 +16921,7 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
ptr += ggml_nbytes(tensor);
fprintf(stderr, "%s: loaded leaf %d: '%16s', %3d dims, %9zu bytes\n", __func__, i, tensor->name, n_dims, ggml_nbytes(tensor));
fprintf(stderr, "%s: loaded leaf %d: '%16s', %9zu bytes\n", __func__, i, tensor->name, ggml_nbytes(tensor));
}
}
@ -16927,12 +16931,10 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
{
uint32_t type;
uint32_t op;
uint32_t n_dims;
for (uint32_t i = 0; i < n_nodes; ++i) {
type = *(const uint32_t *) ptr; ptr += sizeof(type);
op = *(const uint32_t *) ptr; ptr += sizeof(op);
n_dims = *(const uint32_t *) ptr; ptr += sizeof(n_dims);
enum ggml_op eop = (enum ggml_op) op;
@ -17003,7 +17005,7 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
} break;
default:
{
tensor = ggml_new_tensor(*ctx_eval, (enum ggml_type) type, n_dims, ne);
tensor = ggml_new_tensor(*ctx_eval, (enum ggml_type) type, GGML_MAX_DIMS, ne);
tensor->op = eop;
} break;
@ -17022,7 +17024,7 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
result->nodes[i] = tensor;
fprintf(stderr, "%s: loaded node %d: '%16s', %3d dims, %9zu bytes\n", __func__, i, tensor->name, n_dims, ggml_nbytes(tensor));
fprintf(stderr, "%s: loaded node %d: '%16s', %9zu bytes\n", __func__, i, tensor->name, ggml_nbytes(tensor));
}
}
}
@ -17160,7 +17162,7 @@ void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph
fprintf(fp, "(%s)|", ggml_type_name(node->type));
}
if (node->n_dims == 2) {
if (ggml_is_matrix(node)) {
fprintf(fp, "%d [%" PRId64 ", %" PRId64 "] | <x>%s", i, node->ne[0], node->ne[1], ggml_op_symbol(node->op));
} else {
fprintf(fp, "%d [%" PRId64 ", %" PRId64 ", %" PRId64 "] | <x>%s", i, node->ne[0], node->ne[1], node->ne[2], ggml_op_symbol(node->op));
@ -17427,7 +17429,7 @@ static enum ggml_opt_result ggml_opt_adam(
int64_t i = 0;
for (int p = 0; p < np; ++p) {
const int64_t ne = ggml_nelements(ps[p]);
const float p_decay = ((ps[p]->n_dims >= decay_min_ndim) ? decay : 0.0f) * sched;
const float p_decay = ((ggml_n_dims(ps[p]) >= decay_min_ndim) ? decay : 0.0f) * sched;
for (int64_t j = 0; j < ne; ++j) {
float x = ggml_get_f32_1d(ps[p], j);
float g_ = g[i]*gnorm;
@ -19205,8 +19207,8 @@ void gguf_add_tensor(
ctx->infos[idx].ne[i] = 1;
}
ctx->infos[idx].n_dims = tensor->n_dims;
for (int i = 0; i < tensor->n_dims; i++) {
ctx->infos[idx].n_dims = ggml_n_dims(tensor);
for (uint32_t i = 0; i < ctx->infos[idx].n_dims; i++) {
ctx->infos[idx].ne[i] = tensor->ne[i];
}