Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

sync : ggml (backend v2) #3912

Merged
merged 19 commits into from
Nov 13, 2023
Merged

sync : ggml (backend v2) #3912

Changes from 1 commit
Commits
Show all changes
19 commits
Select commit Hold shift + click to select a range
aa7a2c4
sync : ggml (backend v2) (wip)
ggerganov Nov 2, 2023
e819070
sync : migrate examples and llama.cpp to dynamic graphs (wip)
ggerganov Nov 2, 2023
4fe646f
sync : update tests + fix max op params to 64
ggerganov Nov 2, 2023
83c96d5
sync : ggml-cuda
ggerganov Nov 2, 2023
8401e3e
llama : fix save/load state context size
ggerganov Nov 2, 2023
815f44e
sync : try to fix build on tvOS
ggerganov Nov 2, 2023
16e819d
sync : pass custom graph sizes in training examples
ggerganov Nov 2, 2023
e2349ec
sync : update graph copies to new ggml API
ggerganov Nov 2, 2023
f3fb45b
Merge branch 'master' into sync
ggerganov Nov 2, 2023
7f8e2a5
sync : update sync-ggml.sh with new files
ggerganov Nov 2, 2023
075ee61
Merge branch 'master' into sync
ggerganov Nov 3, 2023
dc22db7
scripts : fix header in sync script
ggerganov Nov 3, 2023
b1592ea
train : fix context size calculations
ggerganov Nov 3, 2023
e50ab5a
llama : increase inference graph size up to 4096 nodes
ggerganov Nov 3, 2023
081a86d
Merge branch 'master' into sync
ggerganov Nov 6, 2023
aa1f36c
Merge branch 'master' into sync
ggerganov Nov 7, 2023
a4de804
train : allocate grads for backward graphs
ggerganov Nov 7, 2023
548ec46
train : allocate grads for gb_tmp
ggerganov Nov 7, 2023
9efc6b9
Merge branch 'master' into sync
ggerganov Nov 12, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
20 changes: 11 additions & 9 deletions ggml-alloc.c
View file
Open in desktop
Copy link
Owner Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This merge mostly applies 875fb42

Original file line number Diff line number Diff line change
Expand Up @@ -446,12 +446,14 @@ static ggml_tallocr_t node_tallocr(ggml_gallocr_t galloc, struct ggml_tensor * n
return galloc->hash_allocs[ggml_hash_find_or_insert(galloc->hash_set, node)];
}

static void init_view(ggml_gallocr_t galloc, struct ggml_tensor * view) {
static void init_view(ggml_gallocr_t galloc, struct ggml_tensor * view, bool update_backend) {
ggml_tallocr_t alloc = node_tallocr(galloc, view);

//printf("init_view: %s from src %s\n", view->name, view->view_src->name);
GGML_ASSERT(view->view_src != NULL && view->view_src->data != NULL);
view->backend = view->view_src->backend;
if (update_backend) {
view->backend = view->view_src->backend;
}
view->buffer = view->view_src->buffer;
view->data = (char *)view->view_src->data + view->view_offs;

Expand All @@ -469,7 +471,7 @@ static void allocate_node(ggml_gallocr_t galloc, struct ggml_tensor * node) {

if (node->data == NULL) {
if (ggml_is_view(node)) {
init_view(galloc, node);
init_view(galloc, node, true);
} else {
// see if we can reuse a parent's buffer (inplace)
if (ggml_op_can_inplace(node->op)) {
Expand Down Expand Up @@ -499,14 +501,14 @@ static void allocate_node(ggml_gallocr_t galloc, struct ggml_tensor * node) {
AT_PRINTF("reusing view parent %s (%s) for %s\n", parent->name, view_src->name, node->name);
node->view_src = view_src;
view_src_hn->n_views += 1;
init_view(galloc, node);
init_view(galloc, node, false);
return;
}
} else {
AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name);
node->view_src = parent;
p_hn->n_views += 1;
init_view(galloc, node);
init_view(galloc, node, false);
return;
}
}
Expand Down Expand Up @@ -536,7 +538,7 @@ static void ggml_tallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
hash_get(galloc, view_src)->n_views += 1;
if (node->buffer == NULL && node->data != NULL) {
// view of a pre-allocated tensor, didn't call init_view() yet
init_view(galloc, node);
init_view(galloc, node, true);
}
}

Expand All @@ -547,7 +549,7 @@ static void ggml_tallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
}
hash_get(galloc, parent)->n_children += 1;
if (ggml_is_view(parent) && parent->buffer == NULL && parent->data != NULL) {
init_view(galloc, parent);
init_view(galloc, parent, true);
}
}
}
Expand Down Expand Up @@ -662,7 +664,7 @@ size_t ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, ggml_tallocr_t talloc, st
return max_size;
}

void ggml_gallocr_alloc_graph_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, struct ggml_hash_set hash_set, ggml_tallocr_t * hash_node_alloct) {
void ggml_gallocr_alloc_graph_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, struct ggml_hash_set hash_set, ggml_tallocr_t * hash_node_talloc) {
const size_t hash_size = hash_set.size;

GGML_ASSERT(hash_size >= (size_t)(graph->n_nodes + graph->n_leafs));
Expand All @@ -685,7 +687,7 @@ void ggml_gallocr_alloc_graph_n(ggml_gallocr_t galloc, struct ggml_cgraph * grap
// reset hash values
memset(galloc->hash_values, 0, sizeof(struct hash_node) * hash_size);

galloc->hash_allocs = hash_node_alloct;
galloc->hash_allocs = hash_node_talloc;

ggml_tallocr_alloc_graph_impl(galloc, graph);

Expand Down
Loading
You are viewing a condensed version of this merge commit. You can view the full changes here.