tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-27 20:05:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

sycl : Overcoming workaround for mmap() allocation on Windows (#13482)

Browse Source 
* Remove mmap workaround on windows

After some testing I found that mmap is supported on windows and for
many GPUs on Linux. Therefore I remove the workaround for windows since
it is not necessary.

* Update llama-bench README

SYCL backend introduced a workaround that allows execution of
llama-bench also without specifying `--mmp 0` flag
This commit is contained in:
Nicolò Scipione
2025-05-20 02:54:43 +02:00
committed by GitHub
parent 1dfbf2cf3a
commit f7c9429c85
2 changed files with 7 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
ggml/src/ggml-sycl/ggml-sycl.cpp
View File

@ -385,16 +385,17 @@ static void ggml_backend_sycl_buffer_set_tensor(ggml_backend_buffer_t buffer,
ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context; ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
ggml_sycl_set_device(ctx->device); ggml_sycl_set_device(ctx->device);
auto stream = &(dpct::dev_mgr::instance().get_device(ctx->device).default_queue()); auto stream = &(dpct::dev_mgr::instance().get_device(ctx->device).default_queue());
SYCL_CHECK( SYCL_CHECK(CHECK_TRY_ERROR(dpct::dev_mgr::instance().get_device(ctx->device).queues_wait_and_throw()));
CHECK_TRY_ERROR(dpct::dev_mgr::instance().get_device(ctx->device).queues_wait_and_throw())); #ifndef _WIN32
// Note: Use host buffer to save the data from mmap(), then copy to device. It's workaround for mmap() issue on PVC GPU. // Note: Use host buffer to save the data from mmap(), then copy to device. It's workaround for mmap() issue on PVC GPU.
// This function will be called during load model from disk. Use memory buffer replace dynamic won't save more time and brings potential memory leak risk here. // This function will be called during load model from disk. Use memory buffer replace dynamic won't save more time and brings potential memory leak risk here.
char* host_buf = (char*)malloc(size); char * host_buf = (char *) malloc(size);
memcpy(host_buf, data, size); memcpy(host_buf, data, size);
SYCL_CHECK( SYCL_CHECK(CHECK_TRY_ERROR((*stream).memcpy((char *) tensor->data + offset, host_buf, size).wait()));
CHECK_TRY_ERROR((*stream).memcpy((char *)tensor->data + offset, host_buf, size)
.wait()));
free(host_buf); free(host_buf);
#else
SYCL_CHECK(CHECK_TRY_ERROR((*stream).memcpy((char *) tensor->data + offset, data, size).wait()));
#endif
} }
catch (sycl::exception const &exc) { catch (sycl::exception const &exc) {
std::cerr << exc.what() << "Exception caught at file:" << __FILE__ std::cerr << exc.what() << "Exception caught at file:" << __FILE__

4
tools/llama-bench/README.md
View File

@ -80,10 +80,6 @@ Using the `-d <n>` option, each test can be run at a specified context depth, pr
For a description of the other options, see the [main example](../main/README.md). For a description of the other options, see the [main example](../main/README.md).
Note:
- When using SYCL backend, there would be hang issue in some cases. Please set `--mmp 0`.
## Examples ## Examples
### Text generation with different models ### Text generation with different models