ExternalVendorCode/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-04-20 17:11:17 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

opencl : fix buffer alignment (llama/12197)

Browse Source 
Fix the following error:

```
ggml-alloc.c:99: not enough space in the buffer
ggml_tallocr_alloc: not enough space in the buffer to allocate blk.17.ffn_down.weight (needed 27525120, available 27521024)
```

which occurs when `ggml_backend_opencl_context::alignment` is larger
than `cl_ptr_base` (hard-coded to `0x1000`).

Also, fix `ggml_backend_opencl_context::alignment` was set to
`CL_DEVICE_MEM_BASE_ADDR_ALIGN` which was treated as bytes but the
value is reported in bits.
This commit is contained in:
Henry Linjamäki 2025-03-06 03:33:40 +02:00 committed by Georgi Gerganov
parent 442cd1d2e7
commit 76385c8311
1 changed files with 7 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
ggml/src/ggml-opencl/ggml-opencl.cpp
View File

@ -524,7 +524,10 @@ static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
return backend_ctx;
}
CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_MEM_BASE_ADDR_ALIGN, sizeof(cl_uint), &backend_ctx->alignment, NULL));
cl_uint base_align_in_bits;
CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_MEM_BASE_ADDR_ALIGN, sizeof(cl_uint), &base_align_in_bits, NULL));
GGML_ASSERT(base_align_in_bits % 8u == 0);
backend_ctx->alignment = base_align_in_bits / 8u;
GGML_LOG_INFO("ggml_opencl: mem base addr align: %u\n", backend_ctx->alignment);
clGetDeviceInfo(device, CL_DEVICE_MAX_MEM_ALLOC_SIZE, sizeof(size_t), &backend_ctx->max_alloc_size, NULL);
@ -1198,17 +1201,14 @@ struct ggml_backend_opencl_buffer_context {
std::string name;
};
static void * const cl_ptr_base = (void *)(uintptr_t) 0x1000;
static void ggml_backend_opencl_buffer_free_buffer(ggml_backend_buffer_t buffer) {
ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
delete ctx;
}
static void * ggml_backend_opencl_buffer_get_base(ggml_backend_buffer_t buffer) {
return cl_ptr_base;
GGML_UNUSED(buffer);
ggml_backend_opencl_context * backend_ctx = ggml_cl2_init(buffer->buft->device);
return (void *) (uintptr_t) backend_ctx->alignment;
}
static enum ggml_status ggml_backend_opencl_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
@ -1241,7 +1241,7 @@ static enum ggml_status ggml_backend_opencl_buffer_init_tensor(ggml_backend_buff
tensor->extra = view_extra;
} else {
{
size_t offset = (char *)tensor->data - (char *)cl_ptr_base;
size_t offset = (char *) tensor->data - (char *) ggml_backend_opencl_buffer_get_base(buffer);
ggml_tensor_extra_cl * extra = ctx->ggml_opencl_alloc_temp_tensor_extra();
extra->offset = offset;