mirror of
https://github.com/ggml-org/llama.cpp.git
synced 2025-10-27 08:21:30 +00:00
ggml : fix graph reallocation with multiple chunks (#16396)
reallocation is needed if a single chunk grows in size, even if total allocation size stays the same or is lower
This commit is contained in:
@@ -392,12 +392,8 @@ static void ggml_dyn_tallocr_free(struct ggml_dyn_tallocr * alloc) {
|
|||||||
free(alloc);
|
free(alloc);
|
||||||
}
|
}
|
||||||
|
|
||||||
static size_t ggml_dyn_tallocr_max_size(struct ggml_dyn_tallocr * alloc) {
|
static size_t ggml_dyn_tallocr_max_size(struct ggml_dyn_tallocr * alloc, int chunk) {
|
||||||
size_t max_size = 0;
|
return chunk < alloc->n_chunks ? alloc->chunks[chunk]->max_size : 0;
|
||||||
for (int i = 0; i < alloc->n_chunks; i++) {
|
|
||||||
max_size += alloc->chunks[i]->max_size;
|
|
||||||
}
|
|
||||||
return max_size;
|
|
||||||
}
|
}
|
||||||
|
|
||||||
|
|
||||||
@@ -417,10 +413,8 @@ static void ggml_vbuffer_free(struct vbuffer * buf) {
|
|||||||
free(buf);
|
free(buf);
|
||||||
}
|
}
|
||||||
|
|
||||||
static int ggml_vbuffer_n_chunks(struct vbuffer * buf) {
|
static size_t ggml_vbuffer_chunk_size(struct vbuffer * buf, int chunk) {
|
||||||
int n = 0;
|
return buf->chunks[chunk] ? ggml_backend_buffer_get_size(buf->chunks[chunk]) : 0;
|
||||||
while (n < GGML_VBUFFER_MAX_CHUNKS && buf->chunks[n]) n++;
|
|
||||||
return n;
|
|
||||||
}
|
}
|
||||||
|
|
||||||
static size_t ggml_vbuffer_size(struct vbuffer * buf) {
|
static size_t ggml_vbuffer_size(struct vbuffer * buf) {
|
||||||
@@ -885,12 +879,20 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
|
|||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
size_t cur_size = galloc->buffers[i] ? ggml_vbuffer_size(galloc->buffers[i]) : 0;
|
|
||||||
size_t new_size = ggml_dyn_tallocr_max_size(galloc->buf_tallocs[i]);
|
|
||||||
|
|
||||||
// even if there are no tensors allocated in this buffer, we still need to allocate it to initialize views
|
// even if there are no tensors allocated in this buffer, we still need to allocate it to initialize views
|
||||||
if (new_size > cur_size || galloc->buffers[i] == NULL) {
|
bool realloc = galloc->buffers[i] == NULL;
|
||||||
|
size_t new_size = 0;
|
||||||
|
for (int c = 0; c < galloc->buf_tallocs[i]->n_chunks; c++) {
|
||||||
|
size_t cur_chunk_size = galloc->buffers[i] ? ggml_vbuffer_chunk_size(galloc->buffers[i], c) : 0;
|
||||||
|
size_t new_chunk_size = ggml_dyn_tallocr_max_size(galloc->buf_tallocs[i], c);
|
||||||
|
new_size += new_chunk_size;
|
||||||
|
if (new_chunk_size > cur_chunk_size) {
|
||||||
|
realloc = true;
|
||||||
|
}
|
||||||
|
}
|
||||||
|
if (realloc) {
|
||||||
#ifndef NDEBUG
|
#ifndef NDEBUG
|
||||||
|
size_t cur_size = galloc->buffers[i] ? ggml_vbuffer_size(galloc->buffers[i]) : 0;
|
||||||
GGML_LOG_DEBUG("%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
|
GGML_LOG_DEBUG("%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
|
|||||||
@@ -548,6 +548,41 @@ static void test_buffer_size_zero() {
|
|||||||
GGML_ASSERT(backend_b.context->allocated_total() == 0);
|
GGML_ASSERT(backend_b.context->allocated_total() == 0);
|
||||||
}
|
}
|
||||||
|
|
||||||
|
// Test re-using gallocr for a different graph. The new graph has the same
|
||||||
|
// total size, but one of the chunks is larger, so reallocation is required.
|
||||||
|
static void test_reallocation() {
|
||||||
|
dummy_backend backend = dummy_backend_init(32, /*align*/ 4);
|
||||||
|
ggml_gallocr_ptr galloc;
|
||||||
|
{
|
||||||
|
auto [ctx, graph, ctx_ptr] = make_context();
|
||||||
|
ggml_tensor * x[4];
|
||||||
|
x[0] = make_input_with_size(ctx, 24);
|
||||||
|
x[1] = make_input_with_size(ctx, 16);
|
||||||
|
x[2] = ggml_view_1d(ctx, x[0], 4, 0);
|
||||||
|
x[3] = ggml_add(ctx, x[2], x[1]);
|
||||||
|
assign_names(ctx);
|
||||||
|
|
||||||
|
galloc = allocate_graph(graph, x[3], &backend.buffer_type);
|
||||||
|
check_all_allocated(graph);
|
||||||
|
GGML_ASSERT(backend.context->allocated_total() == 40);
|
||||||
|
}
|
||||||
|
{
|
||||||
|
auto [ctx, graph, ctx_ptr] = make_context();
|
||||||
|
ggml_tensor * x[3];
|
||||||
|
x[0] = make_input_with_size(ctx, 20);
|
||||||
|
x[1] = make_input_with_size(ctx, 20);
|
||||||
|
x[2] = ggml_add(ctx, x[0], x[1]);
|
||||||
|
assign_names(ctx);
|
||||||
|
ggml_set_output(x[2]);
|
||||||
|
ggml_build_forward_expand(graph, x[2]);
|
||||||
|
|
||||||
|
bool result = ggml_gallocr_alloc_graph(galloc.get(), graph);
|
||||||
|
GGML_ASSERT(result);
|
||||||
|
check_all_allocated(graph);
|
||||||
|
GGML_ASSERT(backend.context->allocated_total() == 40);
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
static void run(const char * name, void (*f)()) {
|
static void run(const char * name, void (*f)()) {
|
||||||
printf("%s ", name);
|
printf("%s ", name);
|
||||||
fflush(stdout);
|
fflush(stdout);
|
||||||
@@ -568,5 +603,6 @@ int main() {
|
|||||||
run("test_prefer_already_allocated_memory", test_prefer_already_allocated_memory);
|
run("test_prefer_already_allocated_memory", test_prefer_already_allocated_memory);
|
||||||
run("test_multiple_buffer_types", test_multiple_buffer_types);
|
run("test_multiple_buffer_types", test_multiple_buffer_types);
|
||||||
run("test_buffer_size_zero", test_buffer_size_zero);
|
run("test_buffer_size_zero", test_buffer_size_zero);
|
||||||
|
run("test_reallocation", test_reallocation);
|
||||||
return 0;
|
return 0;
|
||||||
}
|
}
|
||||||
|
|||||||
Reference in New Issue
Block a user