llm: add minimum based on layer size

gpu/gpu.go

@@ -31,8 +31,8 @@ type handles struct {
 }
 
 const (
-	cudaMinimumMemory = 457 * format.MebiByte
-	rocmMinimumMemory = 457 * format.MebiByte
+	cudaMinimumMemory = 256 * format.MebiByte
+	rocmMinimumMemory = 256 * format.MebiByte
 )
 
 var gpuMutex sync.Mutex

gpu/gpu_darwin.go

@@ -15,7 +15,7 @@ import (
 )
 
 const (
-	metalMinimumMemory = 512 * format.MebiByte
+	metalMinimumMemory = 384 * format.MebiByte
 )
 
 func GetGPUInfo() GpuInfoList {