SneppX-ALG
← Back to Blog

"SneppX-ALG v1.1.0 — Serving, Fine-Tuning & Evaluation"

SneppX-ALG v1.1.0 is now released, introducing production model serving with continuous batching, quantized inference, LoRA/QLoRA fine-tuning, DPO/GRPO alignment trainers, and evaluation harnesses.


What's New

1. Production Model Serving

  • **Continuous Batching**: `ContinuousBatchScheduler` with dynamic request queuing and FCFS/priority scheduling.
  • **Quantized Inference**: Weight quantization/dequantization and memory size estimation.
  • **Inference Server**: FastAPI endpoints (`/v1/generate/continuous-batch`, `/v1/models/quantize`) compatible with OpenAI specifications.
  • 2. Fine-Tuning & Alignment

  • **LoRA / QLoRA**: Low-Rank Adaptation layers with rank/alpha configuration and merged weight export.
  • **DPO / GRPO**: Direct Preference Optimization and Group Relative Policy Optimization alignment trainers.
  • 3. Evaluation Harness

  • Synthetic and standardized benchmark suites for MMLU, GSM8K, and HumanEval evaluation.
  • 4. CLI Tools

  • `sneppx-eval`, `sneppx-quantize`, `sneppx-rlhf`, `sneppx-serve`.