Handle Gemma4 BOS override in llama server

llm/llama_server.go

@@ -241,7 +241,14 @@ func (s *llamaServerRunner) tokenizerAddsBOS() bool {
 		return false
 	}
 
-	return s.ggml.KV().Bool("tokenizer.ggml.add_bos_token")
+	kv := s.ggml.KV()
+	if kv.Bool("tokenizer.ggml.add_bos_token") {
+		return true
+	}
+
+	// llama.cpp forces add_bos on for Gemma4 at load time, even for older GGUFs
+	// whose tokenizer.ggml.add_bos_token metadata is false.
+	return kv.String("tokenizer.ggml.pre") == "gemma4"
 }
 
 func (s *llamaServerRunner) completionPromptForRequest(ctx context.Context, req CompletionRequest) (any, error) {

llm/llama_server_test.go

@@ -981,6 +981,7 @@ func TestLlamaServerCompletionBOSOwnership(t *testing.T) {
 		name             string
 		leadingBOS       string
 		tokenizerAddsBOS bool
+		ggmlKV           ggml.KV
 		prompt           string
 		wantPrompt       string
 	}{
@@ -1020,6 +1021,20 @@ func TestLlamaServerCompletionBOSOwnership(t *testing.T) {
 			prompt:           "<bos>hello<bos>",
 			wantPrompt:       "hello<bos>",
 		},
+		{
+			name:       "gemma4 llama.cpp runtime bos override",
+			leadingBOS: "<bos>",
+			ggmlKV: ggml.KV{
+				"general.architecture":            "gemma4",
+				"tokenizer.ggml.pre":              "gemma4",
+				"tokenizer.ggml.add_bos_token":    false,
+				"tokenizer.ggml.bos_token_id":     uint32(2),
+				"tokenizer.ggml.eos_token_id":     uint32(1),
+				"tokenizer.ggml.unknown_token_id": uint32(0),
+			},
+			prompt:     "<bos><|turn>user\nhello<turn|>\n<|turn>model\n",
+			wantPrompt: "<|turn>user\nhello<turn|>\n<|turn>model\n",
+		},
 	}
 
 	for _, tt := range tests {
@@ -1053,7 +1068,9 @@ func TestLlamaServerCompletionBOSOwnership(t *testing.T) {
 				sem:     semaphore.NewWeighted(1),
 				options: api.Options{Runner: api.Runner{NumCtx: 2048}},
 			}
-			if tt.tokenizerAddsBOS {
+			if tt.ggmlKV != nil {
+				runner.ggml = loadTestGGML(t, tt.ggmlKV)
+			} else if tt.tokenizerAddsBOS {
 				runner.ggml = loadTestGGML(t, ggml.KV{
 					"general.architecture":         "gemma3",
 					"tokenizer.ggml.add_bos_token": true,