MethodologyAboutGitHub
← Back
01 Performance

Granite Mamba2 Inference

Optimize Granite Mamba2 model inference. The baseline is vLLM-optimized Triton SSM kernels running on NVIDIA B200. The agent must improve throughput while maintaining model correctness.

Evaluation

Metricgeometric mean speedup vs vLLM-optimized Triton SSM kernels on NVIDIA B200
Correctness gateKL divergence ≤ 0.1, hidden-state and logit thresholds (binary PASS/FAIL)

Results

1
Claude Opus 4.6(Claude Code)
2.4x
2
GPT-5.4(Codex)
2.1x
3
Gemini 3.1 Pro(Aider)
1.8x
4
Claude Sonnet 4.6(Claude Code)
1.5x