Skip to content

fixed bad memory access exception on ios 17 #3527

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

Closed
wants to merge 9 commits into from
Closed

fixed bad memory access exception on ios 17 #3527

Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
4c3a7cd
fixed bad memory access exception on ios 17
l3utterfly Oct 7, 2023
cd0022f
added GGML_ALLOCATOR_DEBUG code back
l3utterfly Oct 7, 2023
88a14fc
fixed logic error, should not allocate more memory, and instead updat…
l3utterfly Oct 7, 2023
9ee8aec
fixed memory leak by freeing temporary graph during session load
l3utterfly Oct 7, 2023
cbf6d07
Merge branch 'master' into mem-fix-ios17
l3utterfly Oct 12, 2023
cb58690
reverted changes in ggml.c
l3utterfly Oct 12, 2023
6f553d9
added trailing newlines
l3utterfly Oct 12, 2023
c8159ea
removed trailing whitespaces
l3utterfly Oct 12, 2023
0ae5b1a
removed trailing whitespace
l3utterfly Oct 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
52 changes: 44 additions & 8 deletions llama.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9001,7 +9001,21 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat
const size_t elt_size = ggml_element_size(kv_self.k);

ggml_context * cpy_ctx = ggml_init({ 4096, NULL, /* no_alloc */ true });
ggml_cgraph gf{};

// create a temporary cgraph without initialising ggml objects, code inspired from `ggml.c:ggml_new_graph`
struct ggml_cgraph * gf = (struct ggml_cgraph *) (malloc(sizeof(ggml_cgraph)));

(*gf).n_nodes = 0;
(*gf).n_leafs = 0;
(*gf).order = GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT;
(*gf).perf_runs = 0;
(*gf).perf_cycles = 0;
(*gf).perf_time_us = 0;

memset((*gf).nodes, 0, sizeof((*gf).nodes));
memset((*gf).grads, 0, sizeof((*gf).grads));
memset((*gf).leafs, 0, sizeof((*gf).leafs));
memset((*gf).visited_hash_table, 0, sizeof((*gf).visited_hash_table));

ggml_tensor * kout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_head, n_layer);
std::vector<uint8_t> kout3d_data(ggml_nbytes(kout3d), 0);
Expand All @@ -9019,16 +9033,20 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat
kv_head, n_embd, n_layer,
elt_size*n_ctx, elt_size*n_ctx*n_embd, 0);

ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, k3d, kout3d));
ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, v3d, vout3d));
ggml_graph_compute_helper(ctx->work_buffer, &gf, /*n_threads*/ 1);
ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, k3d, kout3d));
ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, v3d, vout3d));
ggml_graph_compute_helper(ctx->work_buffer, gf, /*n_threads*/ 1);

ggml_free(cpy_ctx);

// our data is now in the kout3d_data and vout3d_data buffers
// write them to file
data_ctx->write(kout3d_data.data(), kout3d_data.size());
data_ctx->write(vout3d_data.data(), vout3d_data.size());

// free our allocated graph
free(gf);
gf = NULL;
}

for (uint32_t i = 0; i < kv_size; ++i) {
Expand Down Expand Up @@ -9129,7 +9147,21 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
const size_t elt_size = ggml_element_size(kv_self.k);

ggml_context * cpy_ctx = ggml_init({ 4096, NULL, /* no_alloc */ true });
ggml_cgraph gf{};

// create a temporary cgraph without initialising ggml objects, code inspired from `ggml.c:ggml_new_graph`
struct ggml_cgraph * gf = (struct ggml_cgraph *) (malloc(sizeof(ggml_cgraph)));

(*gf).n_nodes = 0;
(*gf).n_leafs = 0;
(*gf).order = GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT;
(*gf).perf_runs = 0;
(*gf).perf_cycles = 0;
(*gf).perf_time_us = 0;

memset((*gf).nodes, 0, sizeof((*gf).nodes));
memset((*gf).grads, 0, sizeof((*gf).grads));
memset((*gf).leafs, 0, sizeof((*gf).leafs));
memset((*gf).visited_hash_table, 0, sizeof((*gf).visited_hash_table));

ggml_tensor * kin3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_head, n_layer);
kin3d->data = (void *) inp;
Expand All @@ -9147,11 +9179,15 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
kv_head, n_embd, n_layer,
elt_size*n_ctx, elt_size*n_ctx*n_embd, 0);

ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, kin3d, k3d));
ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, vin3d, v3d));
ggml_graph_compute_helper(ctx->work_buffer, &gf, /*n_threads*/ 1);
ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, kin3d, k3d));
ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, vin3d, v3d));
ggml_graph_compute_helper(ctx->work_buffer, gf, /*n_threads*/ 1);

ggml_free(cpy_ctx);

// free our allocated graph
free(gf);
gf = NULL;
}

ctx->kv_self.head = kv_head;
Expand Down