happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

ggml-blas: fix memleak 

Signed-off-by: Aaron Teo <aaron.teo1@ibm.com>
This commit is contained in:
Aaron Teo 2025-12-20 13:37:26 +08:00
parent 04ed19bbc0
commit 623e7135c2
No known key found for this signature in database
1 changed files with 30 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

38
ggml/src/ggml-blas/ggml-blas.cpp
View File

@ -28,6 +28,20 @@ struct ggml_backend_blas_buffer {
size_t size; // ggml_nelements * sizeof(float) size_t size; // ggml_nelements * sizeof(float)
}; };
struct ggml_backend_blas_buffer_context {
void * data;
size_t size;
std::vector<ggml_backend_blas_buffer *> buffers;
~ggml_backend_blas_buffer_context() {
ggml_aligned_free(data, size);
for (auto * extra : buffers) {
ggml_aligned_free(extra->data, extra->size);
delete extra;
}
}
};
struct ggml_backend_blas_buffer_type_context { struct ggml_backend_blas_buffer_type_context {
int n_threads; int n_threads;
@ -41,14 +55,15 @@ struct ggml_backend_blas_buffer_type_context {
static void ggml_backend_blas_buffer_free_buffer(ggml_backend_buffer_t buffer) { static void ggml_backend_blas_buffer_free_buffer(ggml_backend_buffer_t buffer) {
GGML_ASSERT(buffer); GGML_ASSERT(buffer);
ggml_backend_blas_buffer * buf_ctx = (ggml_backend_blas_buffer *)buffer->context; ggml_backend_blas_buffer_context * ctx = (ggml_backend_blas_buffer_context *)buffer->context;
ggml_aligned_free(buf_ctx->data, buf_ctx->size); delete ctx;
ggml_aligned_free(buffer->context, buffer->size);
} }
static void * ggml_backend_blas_buffer_get_base(ggml_backend_buffer_t buffer) { static void * ggml_backend_blas_buffer_get_base(ggml_backend_buffer_t buffer) {
GGML_ASSERT(buffer); GGML_ASSERT(buffer);
uintptr_t data = (uintptr_t)buffer->context;
ggml_backend_blas_buffer_context * ctx = (ggml_backend_blas_buffer_context *)buffer->context;
uintptr_t data = (uintptr_t)ctx->data;
// align the buffer // align the buffer
if (data % TENSOR_ALIGNMENT != 0) { if (data % TENSOR_ALIGNMENT != 0) {
@ -68,16 +83,17 @@ static enum ggml_status ggml_backend_blas_buffer_init_tensor(
} }
if (tensor->type != GGML_TYPE_F32) { if (tensor->type != GGML_TYPE_F32) {
ggml_backend_blas_buffer_context * ctx = (ggml_backend_blas_buffer_context *)buffer->context;
ggml_backend_blas_buffer * extra = new ggml_backend_blas_buffer; ggml_backend_blas_buffer * extra = new ggml_backend_blas_buffer;
extra->data = ggml_aligned_malloc(ggml_nelements(tensor) * sizeof(float)); // sizeof(float) because dequantized extra->data = ggml_aligned_malloc(ggml_nelements(tensor) * sizeof(float)); // sizeof(float) because dequantized
extra->size = ggml_nelements(tensor) * sizeof(float); extra->size = ggml_nelements(tensor) * sizeof(float);
tensor->extra = extra; tensor->extra = extra;
ctx->buffers.push_back(extra);
} }
return GGML_STATUS_SUCCESS; return GGML_STATUS_SUCCESS;
GGML_UNUSED(buffer);
} }
static void ggml_backend_blas_buffer_memset_tensor( static void ggml_backend_blas_buffer_memset_tensor(
@ -194,7 +210,9 @@ static void ggml_backend_blas_buffer_clear(
uint8_t value) { uint8_t value) {
GGML_ASSERT(buffer); GGML_ASSERT(buffer);
memset(buffer->context, value, buffer->size);
ggml_backend_blas_buffer_context * ctx = (ggml_backend_blas_buffer_context *)buffer->context;
memset(ctx->data, value, ctx->size);
} }
static const ggml_backend_buffer_i ggml_backend_blas_buffer_i = { static const ggml_backend_buffer_i ggml_backend_blas_buffer_i = {
@ -226,7 +244,11 @@ static ggml_backend_buffer_t ggml_backend_blas_buffer_type_alloc_buffer(
return NULL; return NULL;
} }
return ggml_backend_buffer_init(buft, ggml_backend_blas_buffer_i, data, size); ggml_backend_blas_buffer_context * ctx = new ggml_backend_blas_buffer_context;
ctx->data = data;
ctx->size = size;
return ggml_backend_buffer_init(buft, ggml_backend_blas_buffer_i, ctx, size);
} }
static size_t ggml_backend_blas_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { static size_t ggml_backend_blas_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {