forked from wylab/llama.cpp
state
This commit is contained in:
@@ -13399,6 +13399,11 @@ static void ggml_vk_graph_cleanup(ggml_backend_vk_context * ctx) {
|
||||
ggml_vk_command_pool_cleanup(ctx->device, ctx->transfer_cmd_pool);
|
||||
}
|
||||
|
||||
// Also reset device-level command pools used by cross-device hop1 temporary contexts
|
||||
if (!ctx->device->peer_staging.empty()) {
|
||||
ggml_vk_queue_command_pools_cleanup(ctx->device);
|
||||
}
|
||||
|
||||
GGML_LOG_DEBUG("ggml_vulkan: graph_cleanup(%s) binary_semaphore_idx=%zu\n",
|
||||
ctx->name.c_str(), ctx->binary_semaphore_idx);
|
||||
ctx->binary_semaphore_idx = 0;
|
||||
|
||||
Reference in New Issue
Block a user