CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-30 08:42:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

opencl: use max_alloc_size in backend ctx instead of querying again (#12705)

Browse Source
This commit is contained in:
lhez
2025-04-02 17:01:42 -07:00
committed by GitHub
parent f01bd02376
commit 97a20c012b
1 changed files with 6 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
ggml/src/ggml-opencl/ggml-opencl.cpp
View File

@@ -924,27 +924,24 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
// TODO: fixme: these sizes are hardcoded for now.
// they should be allocated based on the model's size
// and the device's max alloc size
size_t max_alloc_size;
CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_MAX_MEM_ALLOC_SIZE, sizeof(size_t), &max_alloc_size, NULL));
// Allocate intermediate buffers and images
size_t required_A_q_d_bytes = 311164928;
size_t required_A_s_d_bytes = 38895616;
size_t required_B_d_bytes = 45088768;
// Ensure buffer sizes do not exceed the maximum allocation size
size_t max_A_q_d_bytes = MIN(required_A_q_d_bytes, max_alloc_size);
size_t max_A_s_d_bytes = MIN(required_A_s_d_bytes, max_alloc_size);
size_t max_B_d_bytes = MIN(required_B_d_bytes, max_alloc_size);
if (required_A_q_d_bytes > max_alloc_size) {
size_t max_A_q_d_bytes = MIN(required_A_q_d_bytes, backend_ctx->max_alloc_size);
size_t max_A_s_d_bytes = MIN(required_A_s_d_bytes, backend_ctx->max_alloc_size);
size_t max_B_d_bytes = MIN(required_B_d_bytes, backend_ctx->max_alloc_size);
if (required_A_q_d_bytes > backend_ctx->max_alloc_size) {
GGML_LOG_WARN("ggml_opencl: A_q_d buffer size reduced from %zu to %zu due to device limitations.\n",
required_A_q_d_bytes, max_A_q_d_bytes);
}
if (required_A_s_d_bytes > max_alloc_size) {
if (required_A_s_d_bytes > backend_ctx->max_alloc_size) {
GGML_LOG_WARN("ggml_opencl: A_s_d buffer size reduced from %zu to %zu due to device limitations.\n",
required_A_s_d_bytes, max_A_s_d_bytes);
}
if (required_B_d_bytes > max_alloc_size) {
if (required_B_d_bytes > backend_ctx->max_alloc_size) {
GGML_LOG_WARN("ggml_opencl: B_d buffer size reduced from %zu to %zu due to device limitations.\n",
required_B_d_bytes, max_B_d_bytes);
}