diff --git a/ggml-cuda.cu b/ggml-cuda.cu
index 7be63925f..c0c9edd56 100644
--- a/ggml-cuda.cu
+++ b/ggml-cuda.cu
@@ -88,6 +88,8 @@
 #define CC_OFFSET_AMD 1000000
 #define CC_RDNA2      (CC_OFFSET_AMD + 1030)
 
+#define GGML_CUDA_MAX_NODES 8192
+
 // define this if you want to always fallback to MMQ kernels and not use cuBLAS for matrix multiplication
 // on modern hardware, using cuBLAS is recommended as it utilizes F16 tensor cores which are very performant
 // for large computational tasks. the drawback is that this requires some extra amount of VRAM:
@@ -7727,7 +7729,7 @@ static void ggml_cuda_alibi(const ggml_tensor * src0, const ggml_tensor * src1,
     ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_alibi);
 }
 
-void ggml_cuda_im2col(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+static void ggml_cuda_im2col(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_im2col);
 }
 
@@ -7842,11 +7844,11 @@ static size_t g_temp_tensor_extra_index = 0;
 
 static ggml_tensor_extra_gpu * ggml_cuda_alloc_temp_tensor_extra() {
     if (g_temp_tensor_extras == nullptr) {
-        g_temp_tensor_extras = new ggml_tensor_extra_gpu[GGML_DEFAULT_GRAPH_SIZE];
+        g_temp_tensor_extras = new ggml_tensor_extra_gpu[GGML_CUDA_MAX_NODES];
     }
 
     size_t alloc_index = g_temp_tensor_extra_index;
-    g_temp_tensor_extra_index = (g_temp_tensor_extra_index + 1) % GGML_DEFAULT_GRAPH_SIZE;
+    g_temp_tensor_extra_index = (g_temp_tensor_extra_index + 1) % GGML_CUDA_MAX_NODES;
     ggml_tensor_extra_gpu * extra = &g_temp_tensor_extras[alloc_index];
     memset(extra, 0, sizeof(*extra));
 
@@ -8173,11 +8175,11 @@ struct ggml_backend_buffer_context_cuda {
 
     ggml_tensor_extra_gpu * ggml_cuda_alloc_temp_tensor_extra() {
         if (temp_tensor_extras == nullptr) {
-            temp_tensor_extras = new ggml_tensor_extra_gpu[GGML_DEFAULT_GRAPH_SIZE];
+            temp_tensor_extras = new ggml_tensor_extra_gpu[GGML_CUDA_MAX_NODES];
         }
 
         size_t alloc_index = temp_tensor_extra_index;
-        temp_tensor_extra_index = (temp_tensor_extra_index + 1) % GGML_DEFAULT_GRAPH_SIZE;
+        temp_tensor_extra_index = (temp_tensor_extra_index + 1) % GGML_CUDA_MAX_NODES;
         ggml_tensor_extra_gpu * extra = &temp_tensor_extras[alloc_index];
         memset(extra, 0, sizeof(*extra));
 
diff --git a/llama.cpp b/llama.cpp
index 2bd5a4973..c4c18c3b8 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -6248,7 +6248,10 @@ static std::vector<llama_vocab::id> llama_tokenize_internal(const llama_vocab &
                         //  by modifying llm_tokenizer_x to operate with string offsets like pre-tokenizer
                         //  and passing 'add space prefix' as bool argument
                         //
-                        auto raw_text = (special ? "" : " ") + fragment.raw_text.substr(fragment.offset, fragment.length);
+                        auto raw_text = fragment.raw_text.substr(fragment.offset, fragment.length);
+                        if (&fragment == &fragment_buffer.front()) {
+                            raw_text = " " + raw_text; // prefix with space if the first token is not special
+                        }
 
 #ifdef PRETOKENIZERDEBUG
                         fprintf(stderr,"TT: (%ld %ld %ld) '%s'\n", raw_text.length(), fragment.offset, fragment.length, raw_text.c_str());