fix perplexity - it's memory needs dont grow, so we skip it

Author: Green-Sky

llama.cpp

@@ -632,7 +632,7 @@ static bool llama_eval_internal(
     auto & mem_at_token1 = lctx.mem_at_token1;
 
     // TODO: fix this hardcoded size
-    static size_t buf_size = size_t(n_ctx)*1024*1024;
+    static size_t buf_size = size_t(n_ctx)*size_t(N)*128*1024;
     static void * buf = malloc(buf_size);
 
     const size_t C0 = mem_at_token0; // ~base

main.cpp

@@ -219,7 +219,7 @@ int main(int argc, char ** argv) {
     // (fill in mem_at_token0 and mem_at_token1)
     // TODO: better way to do that
     // TODO(Green-Sky): move to internal and detect first time usage
-    {
+    if (!params.perplexity) { // perplexity does not grow over time
         // we make 2 evals, of batchsize to take 2 measurements, to determine base and growth
         std::vector<llama_token> tmp(params.n_batch*2, 2);
         tmp[0] = llama_token_bos();