sdgoij/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-05-14 13:04:08 +00:00
Code Issues Packages Projects Releases Wiki Activity

state

Browse Source
This commit is contained in:
Ruben Ortlam
2026-04-09 11:11:17 +02:00
parent e5e0be0add
commit ddaafa3dc1
1 changed files with 26 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@@ -13903,16 +13903,23 @@ static vk_peer_copy_buf * ggml_vk_get_peer_copy_buf(vk_device& src_dev, vk_devic
bool use_sync_fd = src_dev->external_semaphore_sync_fd && dst_dev->external_semaphore_sync_fd;
staging.use_sync_fd = use_sync_fd;
GGML_LOG_DEBUG("ggml_vulkan: peer_staging init: src=%s dst=%s use_sync_fd=%d\n",
src_dev->name.c_str(), dst_dev->name.c_str(), (int)use_sync_fd);
if (use_sync_fd) {
vk::ExportSemaphoreCreateInfo export_ci{ vk::ExternalSemaphoreHandleTypeFlagBits::eSyncFd };
vk::SemaphoreCreateInfo sci{};
sci.setPNext(&export_ci);
staging.src_sem = src_dev->device.createSemaphore(sci);
GGML_LOG_DEBUG("ggml_vulkan: created exportable src_sem=%p on %s\n",
(void *)(VkSemaphore)staging.src_sem, src_dev->name.c_str());
} else {
vk::SemaphoreTypeCreateInfo tci{ vk::SemaphoreType::eTimeline, 0 };
vk::SemaphoreCreateInfo sci{};
sci.setPNext(&tci);
staging.tl_sem = src_dev->device.createSemaphore(sci);
GGML_LOG_DEBUG("ggml_vulkan: created timeline tl_sem=%p on %s\n",
(void *)(VkSemaphore)staging.tl_sem, src_dev->name.c_str());
}
}
@@ -13971,12 +13978,18 @@ static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend_src, ggml_ba
size_t src_offset = vk_tensor_offset(src) + src->view_offs;
size_t dst_offset = vk_tensor_offset(dst) + dst->view_offs;
GGML_LOG_DEBUG("ggml_vulkan: cross-device copy %s -> %s, nbytes=%zu\n",
src_dev->name.c_str(), dst_dev->name.c_str(), nbytes);
vk_peer_copy_buf * copy_buf = ggml_vk_get_peer_copy_buf(src_dev, dst_dev, nbytes);
if (!copy_buf) {
GGML_LOG_DEBUG("ggml_vulkan: cross-device copy: failed to get peer copy buf\n");
return false;
}
auto& staging = src_dev->peer_staging[dst_dev.get()];
GGML_LOG_DEBUG("ggml_vulkan: cross-device copy: buf_idx=%zu/%zu use_sync_fd=%d\n",
staging.buf_idx, staging.bufs.size(), (int)staging.use_sync_fd);
// HOP 1: src VRAM → staging (on source compute queue)
// Implicit queue submission ordering guarantees this
@@ -14012,9 +14025,20 @@ static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend_src, ggml_ba
};
int sync_fd = src_dev->device.getSemaphoreFdKHR(get_fd_info);
GGML_LOG_DEBUG("ggml_vulkan: cross-device sync_fd: src_dev=%s dst_dev=%s "
"nbytes=%zu sync_fd=%d src_sem=%p dst_compute_ctx=%p\n",
src_dev->name.c_str(), dst_dev->name.c_str(),
nbytes, sync_fd, (void *)(VkSemaphore)staging.src_sem,
(void *)dst_compute_ctx.get());
// Per-copy dest semaphore (temporary import is consumed per wait)
vk_semaphore * dst_sem = ggml_vk_create_binary_semaphore(ctx);
GGML_LOG_DEBUG("ggml_vulkan: cross-device importing sync_fd=%d into dst_sem=%p "
"(binary_semaphore_idx=%zu)\n",
sync_fd, (void *)(VkSemaphore)dst_sem->s,
ctx->binary_semaphore_idx);
vk::ImportSemaphoreFdInfoKHR import_info{
dst_sem->s,
vk::SemaphoreImportFlagBits::eTemporary,
@@ -14023,6 +14047,8 @@ static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend_src, ggml_ba
};
dst_dev->device.importSemaphoreFdKHR(import_info);
GGML_LOG_DEBUG("ggml_vulkan: cross-device import succeeded\n");
dst_compute_ctx->s->wait_semaphores.push_back({ dst_sem->s, 0 });
} else {
// Tier 2: timeline semaphore with CPU wait