nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 18 Packages Projects Releases Wiki Activity

sycl: add GGML_SYCL_USE_ASYNC_MEM_OP env toggle (#22153)

Browse Source 
* sycl: add GGML_SYCL_USE_ASYNC_MEM_OP env toggle

Signed-off-by: Chun Tao <chun.tao@intel.com>

* Use async mem ops for correctness when SYCL graphs are explicitly on.

Signed-off-by: Tao, Chun <chun.tao@intel.com>

---------

Signed-off-by: Chun Tao <chun.tao@intel.com>
Signed-off-by: Tao, Chun <chun.tao@intel.com>
Co-authored-by: Chun Tao <chun.tao@intel.com>
This commit is contained in:
Intel AI Get-to Market Customer Success and Solutions
2026-05-19 09:44:02 +03:00
committed by GitHub
co-authored by Chun Tao
parent c3e9ade6dd
commit 439f1b193d
1 changed files with 7 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-sycl/ggml-sycl.cpp
+7 -4
View File
@@ -72,6 +72,7 @@ int g_ggml_sycl_disable_graph = 0;
int g_ggml_sycl_disable_dnn = 0; int g_ggml_sycl_disable_dnn = 0;
int g_ggml_sycl_prioritize_dmmv = 0; int g_ggml_sycl_prioritize_dmmv = 0;
int g_ggml_sycl_use_async_mem_op = 0; int g_ggml_sycl_use_async_mem_op = 0;
int g_ggml_sycl_use_async_mem_op_requested = 1;
int g_ggml_sycl_enable_level_zero = 0; int g_ggml_sycl_enable_level_zero = 0;
int g_ggml_sycl_enable_flash_attention = 1; int g_ggml_sycl_enable_flash_attention = 1;
@@ -304,6 +305,8 @@ static void ggml_check_sycl() try {
GGML_LOG_INFO(" GGML_SYCL_DISABLE_DNN: DNN disabled by compile flag\n"); GGML_LOG_INFO(" GGML_SYCL_DISABLE_DNN: DNN disabled by compile flag\n");
#endif #endif
GGML_LOG_INFO(" GGML_SYCL_PRIORITIZE_DMMV: %d\n", g_ggml_sycl_prioritize_dmmv); GGML_LOG_INFO(" GGML_SYCL_PRIORITIZE_DMMV: %d\n", g_ggml_sycl_prioritize_dmmv);
g_ggml_sycl_use_async_mem_op_requested = get_sycl_env("GGML_SYCL_USE_ASYNC_MEM_OP", 1);
GGML_LOG_INFO(" GGML_SYCL_USE_ASYNC_MEM_OP: %d\n", g_ggml_sycl_use_async_mem_op_requested);
#ifdef SYCL_FLASH_ATTN #ifdef SYCL_FLASH_ATTN
GGML_LOG_INFO(" GGML_SYCL_ENABLE_FLASH_ATTN: %d\n", g_ggml_sycl_enable_flash_attention); GGML_LOG_INFO(" GGML_SYCL_ENABLE_FLASH_ATTN: %d\n", g_ggml_sycl_enable_flash_attention);
@@ -319,11 +322,11 @@ static void ggml_check_sycl() try {
fprintf(stderr, "%s: SYCL_USE_XMX: no\n", __func__); fprintf(stderr, "%s: SYCL_USE_XMX: no\n", __func__);
#endif #endif
*/ */
// Currently, we only use async malloc / free when graphs are enabled as it is required for the calls to be // Async USM allocation/free is also useful outside the graph path: it avoids the host waits in the reorder
// properly recorded. As this SYCL extension matures it may be beneficial to enable as the default path and in // staging path while preserving queue ordering semantics. Graph support still depends on the extension being
// other places. // available, but it no longer needs to control the non-graph fast path.
#if defined(GGML_SYCL_GRAPH) && SYCL_EXT_ONEAPI_ASYNC_MEMORY_ALLOC #if defined(GGML_SYCL_GRAPH) && SYCL_EXT_ONEAPI_ASYNC_MEMORY_ALLOC
g_ggml_sycl_use_async_mem_op = !g_ggml_sycl_disable_graph; g_ggml_sycl_use_async_mem_op = g_ggml_sycl_use_async_mem_op_requested || !g_ggml_sycl_disable_graph;
if (g_ggml_sycl_use_async_mem_op) { if (g_ggml_sycl_use_async_mem_op) {
for (unsigned int i = 0; i < dpct::dev_mgr::instance().device_count(); ++i) { for (unsigned int i = 0; i < dpct::dev_mgr::instance().device_count(); ++i) {
if (!dpct::dev_mgr::instance().get_device(i).has(sycl::aspect::ext_oneapi_async_memory_alloc)) { if (!dpct::dev_mgr::instance().get_device(i).has(sycl::aspect::ext_oneapi_async_memory_alloc)) {