litert-torch

Converts PyTorch models (e.g. ResNet, timm, HuggingFace transformers) directly to LiteRT (.tflite) flatbuffer format. Use when converting PyTorch models to TFLite, setting up export environments, or troubleshooting torch-to-litert conversion bugs. Don't use for ONNX exports or converting existing TensorFlow models.

Onboards and exports new or custom Hugging Face (or Diffusers/PyTorch) model architectures in `litert_torch.generative.export_hf` using pluggable `ModelExtension` classes (built-in under `model_ext/<arch>/` or out-of-tree via `--extension_modules`). Use when supporting a new `model_type` in `export_hf` or `litert_lm`, inspecting a checkpoint with `model_ext/inspector.py`, patching RMSNorm/SwiGLU/ShortConv/Attention/RoPE for `torch.export`, configuring fused weights or `LlmMetadata`, or verifying extensions with `model_ext/testing.py`. Don't use for generic non-generative PyTorch-to-TFLite conversion or standalone post-export equivalence evaluation.

Validates equivalence between LiteRT models (litert_lm) and PyTorch models (transformers). Use when you need to verify that an exported LiteRT model produces the same outputs as the original Hugging Face model. Supports multi-turn conversations and custom prompts.

Assists the user to calibrate, merge, and statically quantize litert LLM models (such as Gemma 3) in standard open-source (OSS) environments. Use when the user wants to run LLM calibration, merge task JSON results, align KV cache parameters across models, protect sensitive layers in Float32, or run quantized inference testing. Don't use for JAX/PyTorch custom quantization configurations or non-litert models.