Changelog¶

All notable changes to TiRex-2 are documented here.

Unreleased¶

Improvements¶

• ⭐️🚀 Faster CUDA inference with CUDA Graphs and improved Triton kernels in #80, contributed by Martin Loretz and Zhihao Dai:
  • Graphs replays accelerate small batches by up to 5×. Enabled by default for up to 32 variate rows on CUDA; disable with load_model(..., cuda_graphs=False).
  • Triton sLSTM kernels replace FlashRNN, with FP32 states and PyTorch fallbacks. Removed the flashrnn dependency.
  • Triton mLSTM kernels replace mlstm_kernels. Removed the mlstm_kernels dependency.
  • CUDA Toolkit is no longer required. On Linux, a C compiler is needed for kernel compilation; on Windows, PyTorch fallbacks are used. See FAQ for details.
  • Fused sLSTM gate and mLSTM input projections; existing checkpoints adapt on load.
  • Added kernel and CUDA Graph tests. Benchmark scores remain unchanged.
• ⭐️🤖 Forecasting skill: Added a forecasting skill with Claude Code and Codex plugins, including dataframe and tensor examples in #84, contributed by Zhihao Dai.
• ⭐️📦 Offline Docker images: Docker images for CPU/GPU inference servers now include the default model weights for offline deployment in #82, contributed by Zhihao Dai.
• ⭐️📦 Smaller GPU Docker image: Switching to python:3.12-slim without the CUDA Toolkit reduces image size by about 60% (26.3 GB → 10.3 GB). CPU/GPU images now share the same virtualenv setup and healthcheck, in #119, contributed by Daniil Yefimov.
• ‼️ GPU Docker user changed: Both images run as appuser (UID 1000). Update GPU volume mounts from /home/ubuntu/... to /home/appuser/..., in #119, contributed by Daniil Yefimov.
• ⭐️🚀 Inference server batching and workers: The inference server collects concurrent requests in a queue and forecasts them in batches, optionally on several model workers (MODEL_WORKERS). Under concurrent load this gives up to 33x higher throughput on CPU and 22x on GPU when a lot of different clients are connected. Overloaded servers reject requests with HTTP 503 instead of letting them time out (MAX_QUEUE_WAIT_S), in #XXX, contributed by Martin Loretz.
• ‼️‼️ Legacy checkpoint support removed: Checkpoints must be in json and safetensors formats. Legacy yaml and ckpt formats are no longer supported onwards. Use tirex-2<=0.4.0 for legacy checkpoints, or convert them locally with this script, in #112, contributed by Martin Loretz and Zhihao Dai.
• Monotonic quantile forecasts: Sort predictions across quantile levels to prevent quantile crossings in #101, contributed by Martin Loretz.
• ⭐️🚀 Faster batched GPU forecasts: Preprocessing now runs on the CPU and handles all target rows at once, with trend detection via a closed-form least-squares fit instead of np.polyfit per row. GPU forecasts of 32–512 series are 1.5–1.9x faster with identical outputs, in #111, contributed by Daniil Yefimov.
• PyYAML is no longer a direct runtime dependency; it remains explicitly required in the Pixi test and example environments, in #112, contributed by Martin Loretz and Zhihao Dai.

Documentation and development¶

• Migrated the documentation site to Zensical, preserving the classic theme, API reference, and existing preview/build commands, and added pixi run docs-build-clean in #123, contributed by Zhihao Dai.
• Simplified Pixi CUDA setup, removed CUDA Toolkit dependencies, added pixi run build, and validated minimum dependency versions in #117, contributed by Zhihao Dai.
• Added Python 3.15 preview testing and refreshed Pixi dependencies in #116, contributed by Zhihao Dai.
• Moved example and benchmark dependencies from pip extras to Pixi environments and updated setup docs and notebooks in #115, contributed by Zhihao Dai.
• Simplified model code and tests in #80, contributed by Martin Loretz:
• Replaced the FlashRNN layer with sLSTMLayer in slstm_block.py.
• Removed xLSTMMixedConfig and xLSTMLargeConfig; BiXLSTM now takes plain arguments.
• Renamed conv_mLSTMLayerConfig to mLSTMLayerConfig.
• Removed unused feed-forward and norm options.
• Switched CPU reference tests to relative norms to tolerate BF16 rounding differences across machines.
• Removed the legacy conversion script in #113. Use the archived version.

Security and CI¶

• Added plugin CI checks for manifest validation, Codex installation, package contents, and version bumps in #100, contributed by Zhihao Dai.
• Updated CA certificates and CI tooling for the eMudhra/emSign security incident in #83, contributed by Zhihao Dai.
• Simplified Docker image builds and consolidated CI to test CPU/GPU images before publishing in #82, contributed by Zhihao Dai.

0.4.0 - 2026-10-20¶

Improvements¶

• ⭐️🏋 Default safetensors weights: All checkpoints ship model configuration and weights in json and safetensors formats. load_model now loads from config.json and model.safetensors, which avoids unpickling and is faster. The legacy yaml and ckpt files are deprecated with a FutureWarning in #55, contributed by Suad0.
  • ‼️ Legacy yaml and ckpt formats will no longer be supported from v0.5.0 onwards.
• Removed the xlstm dependency to reduce the package footprint, in #71, contributed by Martin Loretz.

Documentation and development¶

• Refactored the sLSTM recurrence interface to support future kernel implementations and decoupled its configuration from FlashRNN in #70, contributed by Martin Loretz.
• Added the Intended Use, Limitations and EU AI Act Notice (INTENDED_USE.md), linked from the README and the documentation, in #67, contributed by Lukas Fischer and Yipeng Sun.
• Added a script to convert legacy checkpoints to new formats in #55, contributed by Suad0.

0.3.1 - 2026-10-05¶

Documentation and development¶

• Refactored internal prediction pipeline, and renamed _predict_once(...) to _predict(...) for clarity in #62, contributed by Christian Ganhör.

0.3.0 - 2026-09-28¶

Improvements¶

• ⭐️📊 Use forecast_df() to forecast directly from pandas, Polars, PyArrow, and other supported eager dataframes. It handles multiple series, covariates, and batching, and returns results in the input dataframe library by default, in #18, contributed by Suad0 and Zhihao Dai.
• ⭐️🚀 Added a pad_context=False option to forecast(...) to skip padding to the default context length, enabling much faster inference for short contexts at a small accuracy cost, in #57, contributed by Daniil Yefimov.
• ⭐️🚀 Added opt-in recurrent-layer compilation with load_model(..., compile=True) for faster repeated forecasts (up to 1.82x on CPU), in #47, contributed by Daniil Yefimov.
• Added a preserve_grad=True option to _predict_once(...) and postprocessing for fine-tuning, preserving finite gradients through inverse differencing, in #58, contributed by Zhihao Dai.
• Replaced einops rearrangements with native PyTorch operations and removed the direct einops dependency in #41, contributed by Copilot and Zhihao Dai.
• Removed a redundant torch.no_grad() context from prediction code in #42, contributed by Zhihao Dai.

Documentation and development¶

• Added a dataframe forecasting guide for native dataframe forecasting via forecast_df, contributed by Suad0 and Zhihao Dai.
• Corrected deployment examples and guidance for covariates, model caching, compilation, and the HTTP API, and added an Open Source vs PRO feature comparison table with consistent PRO naming in #21, contributed by Lukas Fischer, Claude Sonnet 5, and Zhihao Dai.
• Made the default Windows CUDA Pixi platform use the cu130 PyTorch index and fail CUDA tests early when CUDA is unavailable, in #51, contributed by Yipeng Sun.
• Added FAQs covering CUDA setup, supported NVIDIA GPU architectures, and Windows MSVC compiler setup in #19, contributed by Yipeng Sun and Daniil Yefimov.
• Added a Pixi documentation environment and preview/build commands, migrated documentation deployment to GitHub Pages artifact actions, and updated contributor instructions in #37, contributed by Zhihao Dai.
• Removed the unused root requirements.txt in #40, contributed by Copilot and Zhihao Dai.
• Added named Pixi platforms for Linux and Windows CPU builds and CUDA 12.6, 13.0, and 13.2 builds, with matching CUDA toolchains and updated development and benchmark commands in #20, contributed by Zhihao Dai.
• Added GitHub issue and pull request templates in #25, contributed by Zhihao Dai.

Security and CI¶

• Migrated the cross-platform CPU test matrix to Pixi for Python 3.11 and 3.14, added lower-bound tests with Python 3.11 and PyTorch 2.8.0 and Python 3.15 preview coverage, and restricted test workflow permissions in #43, contributed by Zhihao Dai.
• Added a CPU test job using the frozen Pixi lockfile in #44, contributed by Zhihao Dai.
• Updated requests from 2.34.0 to 2.34.2 in the inference dependencies in #35, contributed by Dependabot and Zhihao Dai.
• Hardened Pixi dependency resolution with a seven-day package release cooldown (except for xlstm, mlstm_kernels, and flashrnn), disabled dependency wheel builds, and required Pixi 0.67.0 or newer in #24, contributed by Zhihao Dai and Copilot Autofix.
• Enabled weekly Dependabot version updates for the root Python dependencies in #28, contributed by Zhihao Dai.
• Added a security policy documenting supported versions and private vulnerability reporting in #23, contributed by Copilot and Zhihao Dai.
• Pinned GitHub Actions to commit SHAs and updated action versions in #27, contributed by Copilot and Zhihao Dai.
• Added Visual Studio developer-shell initialization for Windows builds in #27, contributed by Copilot and Zhihao Dai.

0.2.1 - 2026-08-05¶

Improvements¶

• Fixed TiRex-2 functionality on macOS for CPU and MPS backends in #10, contributed by Suad0.
• Fixed TiRex-2 functionality on Windows for CPU and CUDA backends in #5, contributed by Daniil Yefimov.
• Relaxed package requirements to make installation via pip more flexible in #12 and #13, contributed by Daniil Yefimov.
• Added an option to use Flex Attention in #6, contributed by danieleb1861.

0.1.1 - 2026-07-02¶

Initial TiRex-2 release, contributed by martinloretzzz in the initial commit.