提交 #828086: SGLang Project (sgl-project) SGLang 0.5.10.post1 Reachable Assertion信息

标题SGLang Project (sgl-project) SGLang 0.5.10.post1 Reachable Assertion
描述SGLang through 0.5.10.post1, when launched with --enable-lora, does not enforce the --max-loras-per-batch N limit at scheduler batch-construction time. When concurrent inference requests in a single scheduling round target more than N distinct LoRA adapter UIDs (including base-model requests with no lora_path, which count as a distinct UID), the scheduler assembles a batch with N+1 adapters. This triggers a reachable assertion (assert len(cur_uids) <= self.max_loras_per_batch) in python/sglang/srt/lora/lora_manager.py that is evaluated by the model runner. The unhandled AssertionError terminates the scheduler process, SIGQUIT propagates to the server, and the server becomes permanently unresponsive until restarted. A remote attacker with network access to the inference HTTP endpoint can cause denial of service by sending a small burst of completion requests across more adapters than the configured cap. No authentication is required in default deployments.
来源⚠️ https://github.com/sgl-project/sglang/issues/23141
用户
 Zyz3366 (UID 97230)
提交2026-05-13 02時33分 (23 日前)
管理2026-05-31 20時11分 (19 days later)
状态已接受
VulDB条目367593 [SGLang 0.5.10.post1 Inference HTTP Endpoint lora_manager.py lora_path 拒绝服务]
积分20

Want to know what is going to be exploited?

We predict KEV entries!