How abolfazlbyte/llmcompass_mod works
Project type and purpose
• From the README and file tree, this is an open-source hardware-software co-design and simulation framework (“LLMCompass”) for evaluating large language model (LLM) inference hardware.
• It exposes a modular pipeline to explore design spaces (configs → models → metrics → visualization).
File structure
• Top-level directories:
• configs: JSON definitions of hardware templates and experiment scenarios
• hardware_model: Python classes for device, system, compute/memory/io modules, interconnect
• software_model: operator kernels (matmul, layernorm, gelu, transformer, etc.) and communication primitives
• cost_model: performance/area/supply-chain estimation routines
• systolic_array_model: LUT files for array sizes mapping to cycle counts or bandwidth
• design_space_exploration: orchestration entry point (dse.py)
• docs, Dockerfile, environment.yml, utils.py, README.md
• Patterns:
• Pipeline style: config loader → model instantiation → simulation → cost estimation → result output
• Clean separation: hardware vs. software vs. cost vs. exploration logic
README for additional insights
• The user guide (docs/run.md) describes how to run simulations for each figure/experiment.
• Dependencies: Python 3.9, scalesim, PyTorch, matplotlib, seaborn, scipy.
• Dockerfile for reproducibility.