Hybrid approach based on H100 measurements: - aux_child: grouped GEMM for IQL high+low, 2 streams for IQN+Attention - forward_child: backward branch parallelism (forward already multi-stream) - Target: 4.0s/step → 2.0s/step Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>