Limitations and Roadmap

Limitations and Roadmap

Current limitations

• Local batch job only; no server/API or autoscaling scheduler.
• Entire model replicated on every GPU; no cross-GPU weight sharding.
• Simple sentiment classes only; strict postprocessing may return null for verbose completions.
• No GPU out-of-memory recovery, batching backpressure metrics, persistent job queue or checkpoint/resume.
• No baseline-driven multi-GPU scalability evidence is included in the uploaded project.
• Synthetic data and mock backend are not accuracy benchmarks.

Potential enhancements (not implemented)

1. Reliable benchmark automation with warm-up and memory telemetry.
2. Better label constrained decoding and structured outputs.
3. Dataset chunking, streaming, resume/retry and failure isolation.
4. Quantization and batch-adaptive scheduling.
5. Optional vLLM or Triton serving implementation (would be a distinct architecture).
6. CI tests for JSONL validation, mock multiprocessing and diagram-link checking.