From 9a43994c45f8da1b21fd302d5ef000cee36c4e16 Mon Sep 17 00:00:00 2001
From: Jesse Gross <jesse@ollama.com>
Date: Thu, 10 Jul 2025 16:55:34 -0700
Subject: [PATCH] ggml: Disable unused pipeline parallelism

We're not currently using it, even in cases where we could. Disabling
it improves generation performance by 10-30% with multiple GPUs.
---
 ml/backend/ggml/ggml.go | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/ml/backend/ggml/ggml.go b/ml/backend/ggml/ggml.go
index 680910f8dd..7d6831eed0 100644
--- a/ml/backend/ggml/ggml.go
+++ b/ml/backend/ggml/ggml.go
@@ -418,7 +418,7 @@ func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
 			(*C.ggml_backend_buffer_type_t)(unsafe.Pointer(&schedBufts[0])),
 			C.int(len(schedBackends)),
 			C.size_t(maxGraphNodes),
-			C._Bool(len(gpus) > 1 && slices.Contains(gpus, output.d)),
+			C._Bool(false),
 			C._Bool(false),
 		),
 		schedBackends: schedBackends,