From ba04902670cd5945ded682c1c9de2220475b9c38 Mon Sep 17 00:00:00 2001
From: Jeffrey Morgan <jmorganca@gmail.com>
Date: Thu, 26 Jun 2025 00:19:44 -0700
Subject: [PATCH] fs/ggml: add multiplier in graph estimates (#11208)

---
 fs/ggml/ggml.go | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/fs/ggml/ggml.go b/fs/ggml/ggml.go
index c11b3385a..32f459a3a 100644
--- a/fs/ggml/ggml.go
+++ b/fs/ggml/ggml.go
@@ -555,7 +555,7 @@ func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType stri
 			// vocab graph
 			4*batch*(embedding+vocab)+embedding*vocab*105/128,
 		)
-	case "gemma", "gemma2", "gemma3":
+	case "gemma", "gemma2", "gemma3", "gemma3n":
 		fullOffload = max(
 			4*batch*(embedding+vocab),
 			4*batch*(2+context+context*heads+2*embedding+2*embeddingHeadsK*heads),
@@ -568,6 +568,11 @@ func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType stri
 				embedding*embeddingHeadsK*heads*9/16,
 		)
 
+		if f.KV().Architecture() == "gemma3n" {
+			fullOffload *= 4
+			partialOffload *= 4
+		}
+
 		// Gemma2 also has sliding window attention but we only have an optimized implementation in the Ollama
 		// engine. Gemma3 always uses the Ollama engine.
 		if f.KV().Architecture() == "gemma3" {