experiments/M31-Python-Agent-220M-v5/runtime/README_RUNTIME.md

M31Tesla — M31-Python-Agent-220M-v5 Runtime

Standalone runtime for the custom M31-Python-Agent-220M-v5 model.

Architecture: 228,953,984 parameters; vocab 32,768; hidden 896; 20 layers; 14 query heads; 7 KV heads; intermediate 2,816; GQA; RMSNorm; RoPE theta 10,000; tied embeddings.

The v5 trainer used a 512-token training sequence. The runtime therefore defaults to a 512-token working context, although the architecture supports positions up to 2,048.

Use m31_runtime.load_stage("repair") to load a checkpoint and generation_m31.generate(...) to generate text.