SDCGuard

SDCGuard is an LLM testing tool that shows which requests an inference engine's KV-cache checks actually covered.

SDCGuard runs release tests for LLM inference servers to detect wrong answers and silent failures. A vLLM plugin records checksum calls and compared tokens, while the tool matches those records to client requests offline. It passes a request only if the entire prompt was compared. It supports a specific vLLM setup with prefill and decode on an H100 over NIXL using Qwen3-0.6B. The software is built for developers and engineers testing LLM inference engines and GPU math accuracy.

SDCGuard at a glance

Best for: Software engineer
Runs on: CLI, Python, self-hosted, open source
Pricing: Free and open source under the MIT license
Released: November 2026
Category: AI coding tool
Website: github.com/hari-kancharla/sdcguard

Key features

Records checksum calls and compared tokens
Matches records to client requests offline
Exits 2 on HOLD and 0 on a pass
Replays saved test runs locally without a GPU
Compares two builds using transfer-compare
Checks whether every transfer was compared

Use it for

Testing LLM inference engine upgrades
Detecting stale KV block reads in vLLM
Comparing checksum test results across builds
Verifying prompt check coverage

Alternatives to SDCGuard

llm-d — Deploys production-grade LLM serving on existing hardware clusters.
Secret Guard — Prevents sensitive credentials from reaching third-party AI vendor servers.
LMCache — It stores, compresses, searches, and reuses KV caches across clusters.
AgentGuard — Get a security and quality assessment of AI skills in seconds without an account.
Sandkiln — Runs untrusted code in hardware-isolated Firecracker microVMs.
Lcc — Cuts token counts significantly without losing exact session details.
CodeScan Artex — It uses multiple parallel AI agents to read source code and verify vulnerabilities.
llmrun — Check which open models run on your specific GPU or Mac instantly.
GitGuardian — Detects leaked secrets and insecure machine identities across your entire stack.
StackEngine — It automatically redacts and restores sensitive credentials during live LLM workflows.
whichllm — Ranks models using real benchmarks and hardware fit instead of parameter count.
DevSpecs — Keeps a durable thread of AI work locally without cloud sync or LLM calls.

All SDCGuard alternatives