From 6df83e6daac96e4fdfb04bec8d36840f5b660a0c Mon Sep 17 00:00:00 2001
From: Jeffrey Morgan <jmorganca@gmail.com>
Date: Tue, 9 Jan 2024 11:47:30 -0500
Subject: [PATCH] update rough cuda overhead estimate to 15% + 384MiB

---
 gpu/gpu.go | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/gpu/gpu.go b/gpu/gpu.go
index 5bc1810e..cbe0ed94 100644
--- a/gpu/gpu.go
+++ b/gpu/gpu.go
@@ -131,8 +131,12 @@ func getCPUMem() (memInfo, error) {
 func CheckVRAM() (int64, error) {
 	gpuInfo := GetGPUInfo()
 	if gpuInfo.FreeMemory > 0 && (gpuInfo.Library == "cuda" || gpuInfo.Library == "rocm") {
-		// leave 20% of VRAM free for unaccounted for overhead
-		return int64(gpuInfo.FreeMemory * 4 / 5), nil
+		// leave 10% or 384Mi of VRAM free for unaccounted for overhead
+		overhead := gpuInfo.FreeMemory / 10
+		if overhead < 384*1024*1024 {
+			overhead = 384 * 1024 * 1024
+		}
+		return int64(gpuInfo.FreeMemory - overhead), nil
 	}
 
 	return 0, fmt.Errorf("no GPU detected") // TODO - better handling of CPU based memory determiniation