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 : fix memory allocation size (#12649)

Browse Source 
issue:
https://github.com/CodeLinaro/llama.cpp/pull/17#issuecomment-2760611283

This patch fixes the memory allocation size
not exceeding the maximum size of the OpenCL device.
This commit is contained in:
Junil Kim
2025-04-02 01:54:34 +09:00
committed by GitHub
parent e39e727e9a
commit f423981ac8
1 changed files with 26 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
ggml/src/ggml-opencl/ggml-opencl.cpp
View File

@@ -921,10 +921,33 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
backend_ctx->program_CL_gemm = build_program_from_source(context, device, kernel_src_CL_gemm.c_str(), compile_opts);
CL_CHECK((backend_ctx->CL_mul_mat_Ab_Bi_8x4 = clCreateKernel(backend_ctx->program_CL_gemm, "kernel_mul_mat_Ab_Bi_8x4", &err), err));
// 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 max_A_q_d_bytes = 311164928;
size_t max_A_s_d_bytes = 38895616;
size_t max_B_d_bytes = 45088768;
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) {
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) {
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) {
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);
}
CL_CHECK((backend_ctx->A_q_d_max = clCreateBuffer(context, 0, max_A_q_d_bytes, NULL, &err), err));
CL_CHECK((backend_ctx->A_s_d_max = clCreateBuffer(context, 0, max_A_s_d_bytes, NULL, &err), err));