textgen_test: Use instruct code generation model.

`codellama:7b` is the bare completion model, and works better in code
completion contexts. Here, the test uses it as an instruction-following
model ("Do this."), so the `instruct` variant is the more appropriate one.

PiperOrigin-RevId: 607430293

images/gpu/ollama/Dockerfile

@@ -10,7 +10,7 @@ ENV PATH=$PATH:/usr/local/nvidia/bin:/bin/nvidia/bin
 RUN bash -c '                                  \
     ( ollama serve ) & serverpid="$!";         \
     sleep 5;                                   \
-    ollama pull codellama:7b &&                \
+    ollama pull codellama:7b-instruct &&       \
     ollama pull llama2-chinese:7b-chat &&      \
     kill "$serverpid" &&                       \
     wait "$serverpid"                          \

test/gpu/textgen_test.go

@@ -128,7 +128,7 @@ func TestLLM(t *testing.T) {
 		)
 		promptCtx, promptCancel := context.WithTimeout(ctx, 3*time.Minute)
 		prompt := ollama.Prompt{
-			Model: ollama.ZeroTemperatureModel("codellama:7b"),
+			Model: ollama.ZeroTemperatureModel("codellama:7b-instruct"),
 			Query: fmt.Sprintf(`
 				Generate a Python function that takes a string and verifies that it
 				is a valid Chinese translation of the English phrase "Hello World".