cli99/llm-analysis

Provides theoretical performance estimation for LLMs across diverse parallelism schemes (tensor, pipeline, sequence, expert, and data parallelism) and optimization techniques like activation recomputation and quantization. Integrates with Hugging Face model configs for automatic parameter extraction, while supporting custom JSON-based model/GPU/dtype specifications via CLI or Python API. Enables rapid what-if analysis of training/inference setups to identify feasible configurations and optimal throughput-latency tradeoffs without running actual experiments.

479 stars. No commits in the last 6 months.