nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 34 Packages Projects Releases Wiki Activity

ggml : add ggml_build_forward_select (#18550)

Browse Source 
* ggml : add ggml_build_forward_select

* cuda : adapt CUDA graph compat to new feature

* vulkan : update logic to handle command buffer closing

* ggml : check compute for fusion

* ggml : add comment
This commit is contained in:
Georgi Gerganov
2026-01-19 20:03:19 +02:00
committed by GitHub
parent 3d55846a5c
commit 365a3e8c31
17 changed files with 148 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml.c
+52 -18
View File
@@ -3441,7 +3441,8 @@ struct ggml_tensor * ggml_cast(
result->op = GGML_OP_CPY;
result->src[0] = a;
result->src[1] = result;
result->src[1] = result; // note: this self-reference might seem redundant, but it's actually needed by some
// backends for consistency with ggml_cpy_impl() above
return result;
}
@@ -6725,20 +6726,35 @@ static void ggml_compute_backward(
GGML_ASSERT(!src2_needs_grads || ggml_are_same_shape(src2, cgraph->grads[isrc2]));
}
static size_t ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * node) {
// check if already visited
size_t node_hash_pos = ggml_hash_find(&cgraph->visited_hash_set, node);
static size_t ggml_visit_parents_graph(struct ggml_cgraph * cgraph, struct ggml_tensor * node, bool compute) {
if (node->op != GGML_OP_NONE && compute) {
node->flags |= GGML_TENSOR_FLAG_COMPUTE;
}
const size_t node_hash_pos = ggml_hash_find(&cgraph->visited_hash_set, node);
GGML_ASSERT(node_hash_pos != GGML_HASHSET_FULL);
if (!ggml_bitset_get(cgraph->visited_hash_set.used, node_hash_pos)) {
// This is the first time we see this node in the current graph.
cgraph->visited_hash_set.keys[node_hash_pos] = node;
ggml_bitset_set(cgraph->visited_hash_set.used, node_hash_pos);
cgraph->use_counts[node_hash_pos] = 0;
} else {
if (ggml_bitset_get(cgraph->visited_hash_set.used, node_hash_pos)) {
// already visited
if (compute) {
// update the compute flag regardless
for (int i = 0; i < GGML_MAX_SRC; ++i) {
struct ggml_tensor * src = node->src[i];
if (src && ((src->flags & GGML_TENSOR_FLAG_COMPUTE) == 0)) {
ggml_visit_parents_graph(cgraph, src, true);
}
}
}
return node_hash_pos;
}
// This is the first time we see this node in the current graph.
cgraph->visited_hash_set.keys[node_hash_pos] = node;
ggml_bitset_set(cgraph->visited_hash_set.used, node_hash_pos);
cgraph->use_counts[node_hash_pos] = 0;
for (int i = 0; i < GGML_MAX_SRC; ++i) {
const int k =
(cgraph->order == GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT) ? i :
@@ -6747,7 +6763,7 @@ static size_t ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor
struct ggml_tensor * src = node->src[k];
if (src) {
size_t src_hash_pos = ggml_visit_parents(cgraph, src);
const size_t src_hash_pos = ggml_visit_parents_graph(cgraph, src, compute);
// Update the use count for this operand.
cgraph->use_counts[src_hash_pos]++;
@@ -6778,17 +6794,17 @@ static size_t ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor
return node_hash_pos;
}
static void ggml_build_forward_impl(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor, bool expand) {
static void ggml_build_forward_impl(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor, bool expand, bool compute) {
if (!expand) {
// TODO: this branch isn't accessible anymore, maybe move this to ggml_build_forward_expand
ggml_graph_clear(cgraph);
}
const int n0 = cgraph->n_nodes;
const int n_old = cgraph->n_nodes;
ggml_visit_parents(cgraph, tensor);
ggml_visit_parents_graph(cgraph, tensor, compute);
const int n_new = cgraph->n_nodes - n0;
const int n_new = cgraph->n_nodes - n_old;
GGML_PRINT_DEBUG("%s: visited %d new nodes\n", __func__, n_new);
if (n_new > 0) {
@@ -6797,8 +6813,22 @@ static void ggml_build_forward_impl(struct ggml_cgraph * cgraph, struct ggml_ten
}
}
struct ggml_tensor * ggml_build_forward_select(
struct ggml_cgraph * cgraph,
struct ggml_tensor ** tensors,
int n_tensors,
int idx) {
GGML_ASSERT(idx >= 0 && idx < n_tensors);
for (int i = 0; i < n_tensors; i++) {
ggml_build_forward_impl(cgraph, tensors[i], true, i == idx ? true : false);
}
return tensors[idx];
}
void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor) {
ggml_build_forward_impl(cgraph, tensor, true);
ggml_build_forward_impl(cgraph, tensor, true, true);
}
void ggml_build_backward_expand(
@@ -7229,6 +7259,10 @@ bool ggml_can_fuse_subgraph_ext(const struct ggml_cgraph * cgraph,
return false;
}
if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
return false;
}
if (ggml_node_list_find_tensor(cgraph, outputs, num_outputs, node) != -1) {
continue;
}
@@ -7310,7 +7344,7 @@ static void ggml_graph_dump_dot_leaf_edge(FILE * fp, struct ggml_tensor * node,
label);
}
void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename) {
void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * cgraph, const char * filename) {
char color[16];
FILE * fp = ggml_fopen(filename, "w");
@@ -7331,7 +7365,7 @@ void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph
if (node->flags & GGML_TENSOR_FLAG_PARAM) {
snprintf(color, sizeof(color), "yellow");
} else if (grad) {
if (ggml_graph_find(gf, node)) {
if (ggml_graph_find(cgraph, node)) {
snprintf(color, sizeof(color), "green");
} else {
snprintf(color, sizeof(color), "lightblue");