Update ollama to version 0.1.26.

Also move the `bench` image to a subdirectory.

PiperOrigin-RevId: 612060420

images/gpu/ollama/Dockerfile

@@ -1,5 +1,5 @@
 # https://hub.docker.com/r/ollama/ollama
-FROM ollama/ollama:0.1.13
+FROM ollama/ollama:0.1.26
 
 ENV PATH=$PATH:/usr/local/nvidia/bin:/bin/nvidia/bin
 ENV OLLAMA_ORIGINS=*

images/gpu/ollama/bench/Dockerfile

@@ -1,5 +1,5 @@
 # https://hub.docker.com/r/ollama/ollama
-FROM ollama/ollama:0.1.18
+FROM ollama/ollama:0.1.26
 
 ENV PATH=$PATH:/usr/local/nvidia/bin:/bin/nvidia/bin
 ENV OLLAMA_ORIGINS=*
@@ -19,6 +19,8 @@ RUN bash -c '                                  \
     ollama pull llama2:70b-chat        &&      \
     ollama pull mistral:7b-instruct    &&      \
     ollama pull mixtral:instruct       &&      \
+    ollama pull gemma:2b-instruct      &&      \
+    ollama pull gemma:7b-instruct      &&      \
     kill "$serverpid" &&                       \
     wait "$serverpid"                          \
 '

test/gpu/ollama/ollama.go

@@ -123,9 +123,9 @@ func New(ctx context.Context, server Server, logger testutil.Logger) (*Ollama, e
 		return nil, fmt.Errorf("could not get logs: %w", err)
 	}
 	switch {
-	case strings.Contains(logs, "check that you have installed GPU drivers"):
+	case strings.Contains(logs, "no GPU detected"):
 		llm.HasGPU = false
-	case strings.Contains(logs, "VRAM available"):
+	case strings.Contains(logs, "Nvidia GPU detected"):
 		llm.HasGPU = true
 	default:
 		return nil, fmt.Errorf("cannot determine whether ollama is using GPU from logs:\n%s", logs)