diff --git a/otherarch/gpt2_v2.cpp b/otherarch/gpt2_v2.cpp
index f8ae9a582..151e9dd4a 100644
--- a/otherarch/gpt2_v2.cpp
+++ b/otherarch/gpt2_v2.cpp
@@ -372,10 +372,11 @@ bool gpt2_eval(
 
     //todo: there is a bug that causes the buffer to oom and I cannot figure it out, hack to increase size for now  
     static size_t buf_size = 256u*1024*1024;
+    const size_t extra_buf = 64u*1024*1024;
     static void * buf = malloc(buf_size);
 
     if (mem_per_token > 0 && mem_per_token*N*1.6 > buf_size) {
-        const size_t buf_size_new = 2*(mem_per_token*N); // add 10% to account for ggml object overhead
+        const size_t buf_size_new = extra_buf + 2*(mem_per_token*N); // add 10% to account for ggml object overhead
         //printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
 
         // reallocate
diff --git a/otherarch/gptj_v2.cpp b/otherarch/gptj_v2.cpp
index 307ee5a4e..7d548081c 100644
--- a/otherarch/gptj_v2.cpp
+++ b/otherarch/gptj_v2.cpp
@@ -383,10 +383,11 @@ bool gptj_eval(
 
     //todo: there is a bug that causes the buffer to oom and I cannot figure it out, hack to increase size for now  
     static size_t buf_size = 256u*1024*1024;
+    const size_t extra_buf = 64u*1024*1024;
     static void * buf = malloc(buf_size);
 
-    if (mem_per_token > 0 && mem_per_token*N*1.4 > buf_size) {
-        const size_t buf_size_new = 1.6*(mem_per_token*N); // add 10% to account for ggml object overhead
+    if (mem_per_token > 0 && mem_per_token*N*1.5 > buf_size) {
+        const size_t buf_size_new = extra_buf + 1.6*(mem_per_token*N); // add 10% to account for ggml object overhead
         //printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
 
         // reallocate
diff --git a/otherarch/neox.cpp b/otherarch/neox.cpp
index 14c9f5b74..5b9656aa5 100644
--- a/otherarch/neox.cpp
+++ b/otherarch/neox.cpp
@@ -365,10 +365,11 @@ bool stablelm_eval(
     const int n_rot   = hparams.n_rot;
 
     static size_t buf_size = 256u*1024*1024;
+    const size_t extra_buf = 64u*1024*1024;
     static void * buf = malloc(buf_size);
 
-    if (mem_per_token > 0 && mem_per_token*N*1.2 > buf_size) {
-        const size_t buf_size_new = 1.4*(mem_per_token*N); // add 10% to account for ggml object overhead
+    if (mem_per_token > 0 && mem_per_token*N*1.5 > buf_size) {
+        const size_t buf_size_new = extra_buf + 1.6*(mem_per_token*N); // add 10% to account for ggml object overhead
         //printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
 
         // reallocate