From 3a95ed36f62cd596f9c4923f497689381a17fa06 Mon Sep 17 00:00:00 2001
From: Aman Gupta <amangupta052@gmail.com>
Date: Wed, 28 Oct 2026 01:03:43 +0800
Subject: [PATCH] cont : release loader staging buffer on shrink

---
 src/llama-model-loader.cpp | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/src/llama-model-loader.cpp b/src/llama-model-loader.cpp
index d0e0a8c10e..a9a9cb3649 100644
--- a/src/llama-model-loader.cpp
+++ b/src/llama-model-loader.cpp
@@ -1820,7 +1820,8 @@ bool llama_model_loader::load_all_data(
 
                     // tensors come biggest-first: shrink the buffer when it gets too large, so peak memory stays close to one tensor
                     if (read_buf.size() < buf_size || read_buf.size() > 2*buf_size) {
-                        read_buf = {};
+                        // free the old buffer first, `read_buf = {}` would keep its capacity
+                        read_buf = std::vector<no_init<uint8_t>>();
                         read_buf.resize(buf_size);
                     }
                     uint8_t * read_dst = (uint8_t *) GGML_PAD((uintptr_t) read_buf.data(), align);