From: Diego Devesa <redacted>
Date: Tue, 27 May 2025 11:05:18 +0000 (-0700)
Subject: ggml : allow CUDA graphs when using pipeline parallelism (llama/13814)
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=b75babebb23a047063d13236dc714f69dc1d546a;p=pkg%2Fggml%2Fsources%2Fwhisper.cpp

ggml : allow CUDA graphs when using pipeline parallelism (llama/13814)
---

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) {