From 43bc23ef05177bebc2d6b97abce240c76c5f6d45 Mon Sep 17 00:00:00 2001
From: David Korczynski <david@adalogics.com>
Date: Thu, 25 Jul 2024 13:02:26 -0700
Subject: [PATCH] llamacpp: add load model fuzzer

Signed-off-by: David Korczynski <david@adalogics.com>
---
 projects/llamacpp/build.sh                    | 15 ++++++
 projects/llamacpp/fuzzers/fuzz_load_model.cpp | 53 +++++++++++++++++++
 .../llamacpp/fuzzers/fuzz_load_model.dict     |  1 +
 3 files changed, 69 insertions(+)
 create mode 100644 projects/llamacpp/fuzzers/fuzz_load_model.cpp
 create mode 100644 projects/llamacpp/fuzzers/fuzz_load_model.dict

diff --git a/projects/llamacpp/build.sh b/projects/llamacpp/build.sh
index e462ddb25dab..3af95082abb6 100755
--- a/projects/llamacpp/build.sh
+++ b/projects/llamacpp/build.sh
@@ -18,6 +18,11 @@
 export GGML_NO_OPENMP=1
 sed -i 's/:= c++/:= ${CXX}/g' ./Makefile
 sed -i 's/:= cc/:= ${CC}/g' ./Makefile
+# Avoid function that forks + starts instance of gdb.
+sed -i 's/ggml_print_backtrace();//g' ./ggml/include/ggml.h
+
+# Remove statefulness during fuzzing.
+sed -i 's/static bool is_first_call/bool is_first_call/g' ./ggml/src/ggml.c
 
 UNAME_M=amd642 UNAME_p=amd642 LLAMA_NO_METAL=1 make -j$(nproc)
 
@@ -42,6 +47,16 @@ cp fuzzers/*.dict $OUT/
 $CXX $LIB_FUZZING_ENGINE $CXXFLAGS ${FLAGS} ${OBJ_FILES} fuzzers/fuzz_json_to_grammar.cpp -o $OUT/fuzz_json_to_grammar
 $CXX $LIB_FUZZING_ENGINE $CXXFLAGS ${FLAGS} ${OBJ_FILES} fuzzers/fuzz_grammar.cpp -o $OUT/fuzz_grammar
 
+# Create a corpus for load_model_fuzzer
+./llama-gguf dummy.gguf w
+mkdir $SRC/load-model-corpus
+mv dummy.gguf $SRC/load-model-corpus/
+zip -j $OUT/fuzz_load_model_seed_corpus.zip $SRC/load-model-corpus/*
+$CXX $LIB_FUZZING_ENGINE $CXXFLAGS ${FLAGS} ${OBJ_FILES} \
+    -Wl,--wrap,abort fuzzers/fuzz_load_model.cpp -o $OUT/fuzz_load_model
+echo "[libfuzzer]" > $OUT/fuzz_load_model.options
+echo "detect_leaks=0" >> $OUT/fuzz_load_model.options
+
 if [ "$FUZZING_ENGINE" != "afl" ]
 then
     $CXX $LIB_FUZZING_ENGINE $CXXFLAGS ${FLAGS} ${OBJ_FILES} -DFUZZ_BGE fuzzers/fuzz_tokenizer.cpp -o $OUT/fuzz_tokenizer_bge
diff --git a/projects/llamacpp/fuzzers/fuzz_load_model.cpp b/projects/llamacpp/fuzzers/fuzz_load_model.cpp
new file mode 100644
index 000000000000..d8e9f3c2c2d5
--- /dev/null
+++ b/projects/llamacpp/fuzzers/fuzz_load_model.cpp
@@ -0,0 +1,53 @@
+/* Copyright 2024 Google LLC
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+      http://www.apache.org/licenses/LICENSE-2.0
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+*/
+
+#include "llama.h"
+#include <iostream>
+#include <setjmp.h>
+#include <unistd.h>
+
+jmp_buf fuzzing_jmp_buf;
+
+extern "C" void __wrap_abort(void) { longjmp(fuzzing_jmp_buf, 1); }
+
+extern "C" int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  llama_backend_init();
+
+  char filename[256];
+  sprintf(filename, "/tmp/libfuzzer.%d", getpid());
+
+  FILE *fp = fopen(filename, "wb");
+  if (!fp) {
+    return 0;
+  }
+  fwrite(data, size, 1, fp);
+  fclose(fp);
+
+  auto params = llama_model_params{};
+  memset(&params, 0x0, sizeof(struct llama_model_params));
+  params.use_mmap = false;
+  params.progress_callback = [](float progress, void *ctx) {
+    (void)ctx;
+    return progress > 0.50;
+  };
+
+  if (setjmp(fuzzing_jmp_buf) == 0) {
+    auto *model = llama_load_model_from_file(filename, params);
+    if (model != nullptr) {
+      llama_free_model(model);
+    }
+  }
+  llama_backend_free();
+
+  unlink(filename);
+  return 0;
+}
diff --git a/projects/llamacpp/fuzzers/fuzz_load_model.dict b/projects/llamacpp/fuzzers/fuzz_load_model.dict
new file mode 100644
index 000000000000..463bd4d1c8f5
--- /dev/null
+++ b/projects/llamacpp/fuzzers/fuzz_load_model.dict
@@ -0,0 +1 @@
+"GUFF"