whisper : fix support for larger graphs

whisper.cpp

@@ -120,6 +120,7 @@ static void byteswap_tensor(ggml_tensor * tensor) {
 //#define WHISPER_USE_FLASH_ATTN
 //#define WHISPER_USE_FLASH_FF
 #define WHISPER_MAX_DECODERS 16
+#define WHISPER_MAX_NODES 4096
 
 //
 // ggml helpers
@@ -663,7 +664,7 @@ static void whisper_allocr_graph_init(struct whisper_allocr & allocr, std::funct
     auto & meta  = allocr.meta;
     auto & data  = allocr.data;
 
-    meta.resize(ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead());
+    meta.resize(ggml_tensor_overhead()*WHISPER_MAX_NODES + ggml_graph_overhead());
 
     alloc = ggml_allocr_new_measure(tensor_alignment);
 
@@ -1616,7 +1617,7 @@ static struct ggml_cgraph * whisper_build_graph_encoder(
 
     struct ggml_context * ctx0 = ggml_init(params);
 
-    ggml_cgraph * gf = ggml_new_graph(ctx0);
+    ggml_cgraph * gf = ggml_new_graph_custom(ctx0, WHISPER_MAX_NODES, false);
 
     ggml_allocr * alloc = wstate.alloc_encode.alloc;
 
@@ -2034,7 +2035,7 @@ static struct ggml_cgraph * whisper_build_graph_decoder(
 
     struct ggml_context * ctx0 = ggml_init(params);
 
-    ggml_cgraph * gf = ggml_new_graph(ctx0);
+    ggml_cgraph * gf = ggml_new_graph_custom(ctx0, WHISPER_MAX_NODES, false);
 
     ggml_allocr * alloc = wstate.alloc_decode.alloc;