infer-check 0.2.4

Correctness and reliability testing for LLM inference engines

infer-check

Catches the correctness bugs that benchmarks miss in LLM inference engines.

Quantization silently breaks arithmetic. Serving layers silently alter output. KV caches silently corrupt under load. Benchmarks like lm-evaluation-harness test whether models are smart — infer-check tests whether engines are correct.

The problem

Every LLM inference engine has correctness bugs that benchmarks don't catch:

• KV cache NaN pollution in vLLM-Ascend permanently corrupts all subsequent requests
• FP8 KV quantization in vLLM causes repeated garbage output
• 32.5% element mismatches in SGLang's FP8 DeepGEMM kernels on Blackwell GPUs
• Batch-size-dependent output where tokens change depending on concurrent request count

These aren't model quality problems — they're engine correctness failures. infer-check is a CLI tool that runs differential tests across backends, quantization levels, and concurrency conditions to surface them automatically.

Example results

Results from running infer-check on Llama-3.1-8B-Instruct and Qwen3.5-4B (MoE) on Apple Silicon using mlx-lm and vllm-mlx. These demonstrate what the tool catches — not a comprehensive benchmark.

Quantization sweep

4-bit quantization on Llama-3.1-8B showed clear task-dependent degradation. Numerical tasks broke worst:

                       Llama-3.1-8B: bf16 vs 4bit
┏━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━━━━━━━┓
┃ prompt suite          ┃ identical ┃ severe   ┃ mean_similarity ┃
┡━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━━━━━━━┩
│ adversarial-numerics  │     0/30  │   23/30  │          0.311  │
│ reasoning             │     1/50  │   35/50  │          0.384  │
│ code                  │     0/49  │   30/49  │          0.452  │
└───────────────────────┴───────────┴──────────┴─────────────────┘

A "severe" divergence means the quantized output is functionally wrong — not just worded differently, but giving incorrect answers to questions the bf16 baseline handles correctly. This pattern is consistent with published research on quantization-induced degradation, reproduced here on MLX's native quantization scheme.

Dense vs. MoE comparison

Qwen3.5-4B (Gated Delta Networks + sparse MoE) showed similar degradation rates to dense Llama-3.1-8B in our testing — 35/50 severe on reasoning at 4-bit. Small sample, but the tool picks up the signal clearly on both architectures.

Cross-backend diff

mlx-lm vs vllm-mlx at temperature=0 on Llama-3.1-8B-4bit: 50/50 identical (reasoning) and 30/30 identical (numerics). In this test, the vllm-mlx serving layer introduced zero divergence — output differences in production would come from quantization, not from the serving layer itself.

Determinism

Llama-3.1-8B-4bit and Qwen3.5-4B both scored 50/50 identical across 20 runs per prompt on single-request mlx-lm inference at temperature=0.

Stress test

vllm-mlx at concurrency 1/2/4/8: zero errors, 100% output consistency at all levels. No KV cache corruption or batch-dependent divergence detected.

Installation

pip install infer-check

# With MLX backend support (Apple Silicon)
pip install "infer-check[mlx]"

Usage

Quantization sweep

Compare pre-quantized models against a baseline. Each model is a separate HuggingFace repo. Use --max-tokens to control generation length (defaults to 1024) and --num-prompts to limit the number of prompts used.

infer-check sweep \
  --models "bf16=mlx-community/Meta-Llama-3.1-8B-Instruct-bf16,\
            8bit=mlx-community/Meta-Llama-3.1-8B-Instruct-8bit,\
            4bit=mlx-community/Meta-Llama-3.1-8B-Instruct-4bit" \
  --backend mlx-lm \
  --prompts reasoning \
  --max-tokens 512 \
  --num-prompts 10 \
  --output ./results/sweep/

--prompts accepts either a bundled suite name (reasoning, code, adversarial-numerics, determinism, long-context, quant-sensitive) or a path to any .jsonl file.

The baseline is automatically run twice as a self-check — if it's not 50/50 identical, your comparison data is unreliable.

                                 Sweep Summary
┏━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━┳━━━━━━━━━━┳━━━━━━━━┳━━━━━━━━━━━━━━━━━┓
┃ quant_level         ┃ identical ┃ minor ┃ moderate ┃ severe ┃ mean_similarity ┃
┡━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━╇━━━━━━━━━━╇━━━━━━━━╇━━━━━━━━━━━━━━━━━┩
│ bf16 (self-check)   │     50/50 │  0/50 │     0/50 │   0/50 │          1.0000 │
│ 8bit                │     20/50 │  9/50 │    12/50 │   9/50 │          0.8067 │
│ 4bit                │      1/50 │  3/50 │    11/50 │  35/50 │          0.3837 │
└─────────────────────┴───────────┴───────┴──────────┴────────┴─────────────────┘

Cross-backend diff

Same model, same quant, different inference paths. Catches serving-layer bugs.

# Start vllm-mlx in another terminal:
# vllm-mlx serve mlx-community/Meta-Llama-3.1-8B-Instruct-4bit --port 8000

infer-check diff \
  --model mlx-community/Meta-Llama-3.1-8B-Instruct-4bit \
  --backends "mlx-lm,openai-compat" \
  --base-urls ",http://127.0.0.1:8000" \
  --prompts reasoning \
  --output ./results/diff/

Uses /v1/chat/completions by default (--chat) so server-side chat templates match the local backend. Pass --no-chat for raw /v1/completions.

Determinism

Same prompt N times at temperature=0. Output should be bit-identical every run.

infer-check determinism \
  --model mlx-community/Meta-Llama-3.1-8B-Instruct-4bit \
  --backend mlx-lm \
  --prompts determinism \
  --runs 20 \
  --output ./results/determinism/

Stress test

Concurrent requests through a serving backend. Tests KV cache correctness under load.

infer-check stress \
  --model mlx-community/Meta-Llama-3.1-8B-Instruct-4bit \
  --backend openai-compat \
  --base-url http://127.0.0.1:8000 \
  --prompts reasoning \
  --concurrency 1,2,4,8 \
  --output ./results/stress/

Report

Generate an HTML report from all saved results.

infer-check report ./results/ --format html

Prompt suites

Curated prompts targeting known quantization failure modes:

Suite	Count	Purpose
reasoning.jsonl	50	Multi-step math and logic
code.jsonl	49	Python, JSON, SQL generation
adversarial-numerics.jsonl	30	IEEE 754 edge cases, overflow, precision
long-context.jsonl	10	Tables and transcripts with recall questions
quant-sensitive.jsonl	20	Multi-digit arithmetic, long CoT, precise syntax
determinism.jsonl	50	High-entropy continuations for determinism testing

All suites ship with the package — no need to clone the repo. Custom suites are JSONL files with one object per line (default max_tokens is 1024):

{"id": "custom-001", "text": "Your prompt here", "category": "math", "max_tokens": 512}

Supported backends

Backend	Type	Use case
mlx-lm	In-process	Local Apple Silicon inference with logprobs
llama.cpp	HTTP	llama-server via /completion endpoint
vllm-mlx	HTTP	Continuous batching on Apple Silicon
openai-compat	HTTP	Any OpenAI-compatible server (vLLM, SGLang, Ollama)

Roadmap

• GGUF backend (direct llama.cpp integration without HTTP)
• CUDA vLLM backend for GPU-based differential testing
• Logprobs-based divergence scoring where backends support it
• Automated regression CI mode (infer-check ci with pass/fail exit codes)
• Expanded prompt suites for tool use and multi-turn conversations

Requirements

• Python >= 3.11
• macOS with Apple Silicon (for mlx-lm backend) or Linux
• At least one backend installed

License

Apache 2.0