pypi / litert-torch

litert-torch

What changed in litert-torch 0.9, from its official release notes, upgrade guides and advisories. Sources checked 2026-10-18.

1 breaking22 new APIs and options5 behavior and defaults2 tooling10 fixes and other

See which of these apply to your code, ranked, in your agent:

npx @unstale/cli init

@unstale/cli on npm

litert-torch 0.940 changes

Breaking changes 1

Stop token prefixes are expanded automatically, and additional stop tokens are inferred from the chat template.
0.9.4release notes

New APIs and options 22

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

Behavior and default changes 5

Converter legalization and lowering support was added for 1D transposed convolution.
0.9.4release notes
The export_hf pipeline now supports ASR models.
0.9.4release notes
ExecutorMetadata improves handling of sliding window attention and non-SDPA attentions.
0.9.4release notes
The converter now supports the torch.uint8 dtype.
0.9.4release notes
LoRA FlatBuffer export now packs rank metadata.
0.9.3release notes

Tooling 2

An experimental bring-your-own-model NPU export workflow was added to the export_hf pipeline.
0.9.4release notes
A tool was added to compile LiteRT-LM to a chosen NPU backend.
0.9.4release notes

Fixes and other changes 10

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