enable q8, q5, 5_1, and f32 for linux gpu (#699)

llm/llm.go

@@ -5,6 +5,7 @@ import (
 	"fmt"
 	"log"
 	"os"
+	"runtime"
 
 	"github.com/pbnjay/memory"
 
@@ -37,6 +38,7 @@ func New(workDir, model string, adapters []string, opts api.Options) (LLM, error
 		return nil, err
 	}
 
+	if runtime.GOOS == "darwin" {
 		switch ggml.FileType() {
 		case "Q8_0":
 			if ggml.Name() != "gguf" && opts.NumGPU != 0 {
@@ -53,6 +55,7 @@ func New(workDir, model string, adapters []string, opts api.Options) (LLM, error
 				opts.NumGPU = 0
 			}
 		}
+	}
 
 	totalResidentMemory := memory.TotalMemory()
 	switch ggml.ModelType() {