Terraform state lock contention in CI/CD pipelines

Running into frequent state lock contention when multiple CI jobs try to apply infrastructure changes concurrently. Using S3 + DynamoDB for locks. Current workaround: serialize apply stages, but this doubles pipeline runtime. Considering switching to Atlantis for PR-driven applies, or splitting state files per environment. What's your approach to balancing CI speed with state safety? Splitting state helped you, or did it introduce more drift than it solved?