From b75babebb23a047063d13236dc714f69dc1d546a Mon Sep 17 00:00:00 2001 From: Diego Devesa Date: Tue, 27 May 2025 04:05:18 -0700 Subject: [PATCH] ggml : allow CUDA graphs when using pipeline parallelism (llama/13814) --- ggml/src/ggml-backend.cpp | 3 +++ 1 file changed, 3 insertions(+) diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp index b30b4cb3..1f40f10e 100644 --- a/ggml/src/ggml-backend.cpp +++ b/ggml/src/ggml-backend.cpp @@ -1598,6 +1598,9 @@ void ggml_backend_sched_synchronize(ggml_backend_sched_t sched) { for (int i = 0; i < sched->n_backends; i++) { ggml_backend_synchronize(sched->backends[i]); } + // reset the current copy to 0 so that the graphs will be similar during generation + // necessary for CUDA graphs + sched->cur_copy = 0; } void ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data) {