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

fixed bad memory access exception on ios 17

Browse source
This commit is contained in:
l3utterfly 2023-10-07 22:34:33 +08:00
parent c47066d833
commit 4c3a7cdd46
3 changed files with 90 additions and 61 deletions
Download patch file Download diff file Expand all files Collapse all files

52
ggml-alloc.c
View file

@ -291,21 +291,17 @@ void ggml_allocr_reset(struct ggml_allocr * alloc) {
struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment) { struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment) {
struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */); struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
*alloc = (struct ggml_allocr){ (*alloc).data = data;
/*.data = */ data, (*alloc).size = size;
/*.size = */ size, (*alloc).alignment = alignment;
/*.alignment = */ alignment, (*alloc).n_free_blocks = 0;
/*.n_free_blocks = */ 0, (*alloc).max_size = 0;
/*.free_blocks = */ {{0}}, (*alloc).measure = false;
/*.hash_table = */ {{0}}, (*alloc).parse_seq_len = 0;
/*.max_size = */ 0,
/*.measure = */ false, memset((*alloc).free_blocks, 0, sizeof((*alloc).free_blocks));
/*.parse_seq = */ {0}, memset((*alloc).hash_table, 0, sizeof((*alloc).hash_table));
/*.parse_seq_len = */ 0, memset((*alloc).parse_seq, 0, sizeof((*alloc).parse_seq));
#ifdef GGML_ALLOCATOR_DEBUG
/*.allocated_tensors = */ {0},
#endif
};
ggml_allocr_reset(alloc); ggml_allocr_reset(alloc);
@ -371,21 +367,21 @@ struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
alloc_measure_vmem(&base_addr, &size); alloc_measure_vmem(&base_addr, &size);
*alloc = (struct ggml_allocr){ (*alloc).data = base_addr;
/*.data = */ base_addr, (*alloc).size = size;
/*.size = */ size, (*alloc).alignment = alignment;
/*.alignment = */ alignment, (*alloc).n_free_blocks = 0;
/*.n_free_blocks = */ 0, (*alloc).max_size = 0;
/*.free_blocks = */ {{0}}, (*alloc).measure = true;
/*.hash_table = */ {{0}}, (*alloc).parse_seq_len = 0;
/*.max_size = */ 0,
/*.measure = */ true, memset((*alloc).free_blocks, 0, sizeof((*alloc).free_blocks));
/*.parse_seq = */ {0}, memset((*alloc).hash_table, 0, sizeof((*alloc).hash_table));
/*.parse_seq_len = */ 0, memset((*alloc).parse_seq, 0, sizeof((*alloc).parse_seq));
#ifdef GGML_ALLOCATOR_DEBUG #ifdef GGML_ALLOCATOR_DEBUG
/*.allocated_tensors = */ {0}, (*alloc).allocated_tensors = {0};
#endif #endif
};
ggml_allocr_reset(alloc); ggml_allocr_reset(alloc);

49
ggml.c
View file

@ -4723,18 +4723,20 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
const size_t mem_size = params.mem_buffer ? params.mem_size : GGML_PAD(params.mem_size, GGML_MEM_ALIGN); const size_t mem_size = params.mem_buffer ? params.mem_size : GGML_PAD(params.mem_size, GGML_MEM_ALIGN);
*ctx = (struct ggml_context) { ctx = (struct ggml_context *)malloc(sizeof(struct ggml_context));
/*.mem_size =*/ mem_size,
/*.mem_buffer =*/ params.mem_buffer ? params.mem_buffer : GGML_ALIGNED_MALLOC(mem_size), struct ggml_scratch empty_scratch = { 0, 0, NULL };
/*.mem_buffer_owned =*/ params.mem_buffer ? false : true,
/*.no_alloc =*/ params.no_alloc, (*ctx).mem_size = mem_size;
/*.no_alloc_save =*/ params.no_alloc, (*ctx).mem_buffer = params.mem_buffer ? params.mem_buffer : GGML_ALIGNED_MALLOC(mem_size);
/*.n_objects =*/ 0, (*ctx).mem_buffer_owned = params.mem_buffer ? false : true;
/*.objects_begin =*/ NULL, (*ctx).no_alloc = params.no_alloc;
/*.objects_end =*/ NULL, (*ctx).no_alloc_save = params.no_alloc;
/*.scratch =*/ { 0, 0, NULL, }, (*ctx).n_objects = 0;
/*.scratch_save =*/ { 0, 0, NULL, }, (*ctx).objects_begin = NULL;
}; (*ctx).objects_end = NULL;
(*ctx).scratch = empty_scratch;
(*ctx).scratch_save = empty_scratch;
GGML_ASSERT(ctx->mem_buffer != NULL); GGML_ASSERT(ctx->mem_buffer != NULL);
@ -18079,18 +18081,17 @@ struct ggml_cgraph * ggml_new_graph(struct ggml_context * ctx) {
struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_GRAPH, GGML_GRAPH_SIZE); struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_GRAPH, GGML_GRAPH_SIZE);
struct ggml_cgraph * cgraph = (struct ggml_cgraph *) ((char *) ctx->mem_buffer + obj->offs); struct ggml_cgraph * cgraph = (struct ggml_cgraph *) ((char *) ctx->mem_buffer + obj->offs);
*cgraph = (struct ggml_cgraph) { (*cgraph).n_nodes = 0;
/*.n_nodes =*/ 0, (*cgraph).n_leafs = 0;
/*.n_leafs =*/ 0, (*cgraph).order = GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT;
/*.nodes =*/ { NULL }, (*cgraph).perf_runs = 0;
/*.grads =*/ { NULL }, (*cgraph).perf_cycles = 0;
/*.leafs =*/ { NULL }, (*cgraph).perf_time_us = 0;
/*.hash_table =*/ { NULL },
/*.order =*/ GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT, memset((*cgraph).nodes, 0, sizeof((*cgraph).nodes));
/*.perf_runs =*/ 0, memset((*cgraph).grads, 0, sizeof((*cgraph).grads));
/*.perf_cycles =*/ 0, memset((*cgraph).leafs, 0, sizeof((*cgraph).leafs));
/*.perf_time_us =*/ 0, memset((*cgraph).visited_hash_table, 0, sizeof((*cgraph).visited_hash_table));
};
return cgraph; return cgraph;
} }

48
llama.cpp
View file

@ -8222,7 +8222,21 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat
const size_t elt_size = ggml_element_size(kv_self.k); const size_t elt_size = ggml_element_size(kv_self.k);
ggml_context * cpy_ctx = ggml_init({ 4096, NULL, /* no_alloc */ true }); 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); 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); std::vector<uint8_t> kout3d_data(ggml_nbytes(kout3d), 0);
@ -8240,9 +8254,9 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat
kv_head, n_embd, n_layer, kv_head, n_embd, n_layer,
elt_size*n_ctx, elt_size*n_ctx*n_embd, 0); 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, k3d, kout3d));
ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, v3d, vout3d)); ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, v3d, vout3d));
ggml_graph_compute_helper(ctx->work_buffer, &gf, /*n_threads*/ 1); ggml_graph_compute_helper(ctx->work_buffer, gf, /*n_threads*/ 1);
ggml_free(cpy_ctx); ggml_free(cpy_ctx);
@ -8250,6 +8264,10 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat
// write them to file // write them to file
data_ctx->write(kout3d_data.data(), kout3d_data.size()); data_ctx->write(kout3d_data.data(), kout3d_data.size());
data_ctx->write(vout3d_data.data(), vout3d_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) { for (uint32_t i = 0; i < kv_size; ++i) {
@ -8350,7 +8368,21 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
const size_t elt_size = ggml_element_size(kv_self.k); const size_t elt_size = ggml_element_size(kv_self.k);
ggml_context * cpy_ctx = ggml_init({ 4096, NULL, /* no_alloc */ true }); 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); ggml_tensor * kin3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_head, n_layer);
kin3d->data = (void *) inp; kin3d->data = (void *) inp;
@ -8368,9 +8400,9 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
kv_head, n_embd, n_layer, kv_head, n_embd, n_layer,
elt_size*n_ctx, elt_size*n_ctx*n_embd, 0); 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, kin3d, k3d));
ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, vin3d, v3d)); ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, vin3d, v3d));
ggml_graph_compute_helper(ctx->work_buffer, &gf, /*n_threads*/ 1); ggml_graph_compute_helper(ctx->work_buffer, gf, /*n_threads*/ 1);
ggml_free(cpy_ctx); ggml_free(cpy_ctx);
} }