# What changed in litert-torch

> 40 changes in litert-torch 0.9 from its official release notes, upgrade guides and advisories: 1 breaking, 22 new APIs and options. Each one cited. Sources checked 2026-10-18. HTML: https://unstale.sh/pypi/litert-torch/

To see which of these apply to a repository, run `npx @unstale/cli init` in it.

## litert-torch 0.9

### Breaking changes

- Stop token prefixes are expanded automatically, and additional stop tokens are inferred from the chat template. (0.9.4) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.4)

### New APIs and options

- A tool was added for post-training/static range quantization. (0.9.4) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.4)
- This flag enables dynamic context length support for GPU cache. (0.9.4) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.4)
- This flag enables dynamic context length support for GPU prefill. (0.9.4) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.4)
- This experimental flag enables Jinja template transpilation for better LiteRT-LM support. (0.9.4) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.4)
- This experimental flag enables fp16/fp32 mixed precision support for GPU. (0.9.4) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.4)
- This experimental flag accompanies fp16/fp32 mixed precision support for GPU. (0.9.4) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.4)
- FFT lowering support was added to the converter. (0.9.4) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.4)
- Export support was added for the Gemma 4 architecture. (0.9.4) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.4)
- Experimental support was added for exporting models with an optimized GPU attention custom op. (0.9.4) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.4)
- A sampling executor was added to support inference with exported models. (0.9.4) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.4)
- A tokenizer patch adds Metaspace pre-tokenizer and BPE character support. (0.9.3) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.3)
- Constant-fusing patch updates are supported. (0.9.3) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.3)
- Lowering now supports 1D transposed convolutions, including `output_padding`. (0.9.3) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.3)
- Lowering now supports 1D Real-to-Complex FFT. (0.9.3) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.3)
- An experimental feature transpiles Jinja2 chat templates into Minijinja-compliant templates. (0.9.3) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.3)
- An experimental LiteRT-LM NPU compiler tool is available. (0.9.3) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.3)
- A package provides object-oriented management of `.litertlm` containers. (0.9.3) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.3)
- The Generative API adds a Mixture-of-Experts layer with FP32 and INT8 lowerings. (0.9.3) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.3)
- An OS NPU quantization calibration pipeline is available with sampling executors. (0.9.3) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.3)
- Split-cache models are now exportable. (0.9.3) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.3)
- The `torch.uint8` tensor data type is now supported. (0.9.3) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.3)
- Unified split-cache export now uses a single `valid_mask` input tensor. (0.9.3) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.3)

### Behavior and default changes

- Converter legalization and lowering support was added for 1D transposed convolution. (0.9.4) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.4)
- The export_hf pipeline now supports ASR models. (0.9.4) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.4)
- ExecutorMetadata improves handling of sliding window attention and non-SDPA attentions. (0.9.4) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.4)
- The converter now supports the torch.uint8 dtype. (0.9.4) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.4)
- LoRA FlatBuffer export now packs rank metadata. (0.9.3) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.3)

### Tooling

- An experimental bring-your-own-model NPU export workflow was added to the export_hf pipeline. (0.9.4) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.4)
- A tool was added to compile LiteRT-LM to a chosen NPU backend. (0.9.4) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.4)

### Fixes and other changes

- Tokenizers using a Metaspace pre-tokenizer and BPE chars are fixed to correctly de-tokenize BPE tokens. (0.9.4) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.4)
- Constant cache lookup now distinguishes PyTorch tensors that share a data_ptr but have different offsets. (0.9.4) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.4)
- Argument flattening preserves dictionary keys and supports the updated PyTree API. (0.9.3) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.3)
- Gemma 4 unified models are now supported. (0.9.3) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.3)
- Key lookup collisions are fixed for PyTorch tensors sharing `data_ptr` but having different offsets. (0.9.3) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.3)
- Static shape export no longer has state contamination for LFM2 ShortConv. (0.9.3) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.3)
- Input tensor shape handling is fixed for Parakeet speech models. (0.9.3) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.3)
- Conversion now has an explicit exception safeguard for already-quantized checkpoints. (0.9.3) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.3)
- The RMSNorm patch now supports FP16 mixed-precision export. (0.9.3) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.3)
- Export is supported for Qwen3-ASR with transformers v5, Whisper, Parakeet CTC/TDT, and Moonshine. (0.9.3) — [release notes](https://github.com/google-ai-edge/litert-torch/releases/tag/v0.9.3)