Skip to content

Changelog

Unreleased

[0.20.1] - 2025-09-13

Added

  • Open-Sourced the project under MIT|Apache2 license
  • Official support for tract v0.22.0
  • test coverage of LLM export with various transformers lib version (trying to support last 10ish minor versions with CI/CD)
  • Add context manager to force loading with offloaded tensor
  • Added opt-in support for reification of spda operator when targeting tract export (thanks to @emricksinisonos contribution) this should help further optimization in tract of attention blocks
  • Added support for upsample operator via deconv or debox depending on tract version
  • Added Licenses file
  • ModTensorUpdater is now useful with legacy torch version (bellow 2.0)
  • Add aten::new
  • New logo (thanks to @lizavetaobadava-commits)

Formatting & style

  • All exception now inherit from T2NError (allow easier catch)
  • Stricter line length (even in doc)
  • Stricter doc formatting with ruff
  • Improved prospector strictness
  • isort retired in favor of ruff

Documentation

  • Documentation versioning with mike (allowing to get older version doc)
  • Documentation: fixed typos, rewording (thanks to @thomasnigoghossiansonos for the review)
  • WASM LLM poetry generator example expose the prompt for clarity
  • Nicer WASM example with more loading state infos
  • Fix WASM VAD example handling more audio context (more robust)
  • Added WASM Yolo example with pose-estimation

Fixed

  • transformers regression since 4.56 around cache handling
  • better support for OffloadedTensor with assignations and some in-place operations
  • pylint tweaks

Change

  • Following open-sourcing of the project, packaging is now targeting PyPI.

[0.19.1] - 2025-08-06

Added

  • CI/CD for torch version bellow 2.7: 2.2, 1.13 and 1.10
  • specific checks around dtype for qtensors tests generated assets

Fixes

  • make this package work again for torch version between 1.10 and 2.3.

[0.20.0] - 2025-09-13

Failed release in CI/CD

[0.19.0] - 2025-07-25

Added

  • mkdoc documentation revamp
  • no more approximation of logsofmax for TractNNEF
  • added support for operators: fmod, expm1, atan2, addmm, maximum, minimum, logical_or, logical_and, logical_not, fill_, var, avg_adaptive_pool_nd, max_adaptive_pool_nd, amin, amax, nn.LocalResponseNorm

[0.18.6] - 2025-07-03

Added

  • base param updater

[0.18.5] - 2025-06-13

Added

  • update_values in OffloadedTensor

Fix

  • Parameter addition in OffloadedTensor
  • to_json_file use in config dump in LLM

[0.18.4] - 2025-06-11

Change

  • bunch of cache, ordered conditioning, torch.compile to make export faster

[0.18.3] - 2025-06-05

Fix

  • dtype change (via .to) of OffloadedTensor handled
  • numel call avoid OffloadedTensor reload
  • dtype getter aligned
  • dissociated getter aligned

Added

  • support for python 3.13 landed (some issue still with Quant+Offloaded tensor mem alloc)
  • aten::bitwise_or operator

[0.18.2] - 2025-06-04

Fix

  • safetensors import, only called when needed

[0.18.1] - 2025-06-03

Added

  • Official tract support version: 0.21.13

[0.18.0] - 2025-06-03

Added

  • addition of an OffloadTensor that allow to write on disk the tensor and reload it each time from there (trading memory space for disk usage/reloading speed -> this is not intended to be used beyond compression and export of neural net stage).
  • Plug of a load step by step into OffloadTensor method for tract_llm (as an opt-in via --device-map=t2n_offload_disk option). This option is also compatible with accelerate if installed to spread model partitions load across available hardware devices in an instance.

Change

  • refactor of all custom PyTorch tensors used on torch to NNEF into a unified module
  • [OPTIM] removal of part of redundant inference tracing computation for shape and type

Fix

  • avoid duplicate weights in Numpy data within nnef.Graph until serialization (write) step

[0.17.4] - 2025-05-15

Fix

  • Add eq in TensorVariable to build proper dict keys and in queries from it (without traced data accounted)
  • all tract_core_gather add attrs datum_type
  • Q4 compression_method tag compat with internal llm lib
  • Skip check_io between wrapper_model vs hf_model if wrapped_model

[0.17.3] - 2025-05-09

Added

  • aten::full_like, _softmax, mm, logical_not, scalar_tensor, to_copy
  • forward signature of wrapped llm models is updated live based on model KV cache quantity to help torch.export understand all parameters (args, *kwargs does not work)

Change

  • HFConfigHelper now only need HF conf (no more slug name)

[0.17.2] - 2025-04-10

Added

  • bump tract 0.21.12
  • avoid compress weight if shared with 1 module that is not requested to compress (by example: request nn.Linear only while shared with nn.Embedding)

Fix

  • some ignore-already-exist-dir missing case in llm_tract

[0.17.1] - 2025-04-02

Fix

  • Avoid duplicating weights in case they are shared with assignation post nn.Module load

[0.17.0] - 2025-03-31

Change

  • All parameters variable in graph are be named the same their label if NamedTensor

Fix

  • RNN expansion with multiple call within same graph now refer to same set of weight instead of duplicating them

[0.16.11] - 2025-03-27

Fix

  • set_priority in with sdpa_kernel only appear in torch 2.6

[0.16.10] - 2025-03-24

Fix

  • aten::flatten with partial dimensions
  • aten::remainder force f32 (if other implicit dtype support like ints)
  • aten::pad... now support dynamic dimensions
  • aten::zeros, ... now default to f32 in cases where unspecified in jit graph
  • Merge of subgraph in ir_graph is now done with preserving subgraph output names (needed since some output may be repeated while main graph unaware of it)

Added

  • Conv are now supported for Q40 exports (tract v0.21.12)
  • compress registry min_max_q4_0_all export all supported tensors in Q40 (including Conv1d, Conv2d)

[0.16.9] - 2025-03-20

Fix

  • regression on uint32, uint64 support (pre torch 2.4)

[0.16.8] - 2025-03-20

Fix

  • regression on uint16 support (pre torch 2.4)

[0.16.7] - 2025-03-20

Fix

  • complex slice index gather nd fix

[0.16.6] - 2025-03-20

Added

  • official tract support is now 0.21.11 (new default target)
  • support to device like cuda,mps for our internal QTensor ...
  • support for new operators: aten::empty_like, aten::prod, aten::index_select, aten::scatter, aten::numel

Change

  • additional tracing cues for whole number values that may be used in tensors shaping/construction.
  • disabled support for Python >=3.13 as of now as it leads to unexpected hash/set issues to be investigated

Fix

  • aten::baddbmm extra args handled during tracing
  • better alignment of arity for rnn inputs
  • equality operators (ne, ge, le, gt, eq) now implicit cast to common dtype if heterogeneous
  • to operators with from float to unsigned with negative values was found to have an arch dependant behavior (code now align to the arch used at export with warning for non arm)
  • tolerate export pad operators with dynamic values

[0.16.5] - 2025-03-11

Change

  • test by default on 2.6

Fix

  • SPDA regression if pytorch > 2.3 and usage of specific scale

[0.16.4] - 2025-03-11

Added

  • support new Q40 tract format starting with target tract>=0.21.11

Fix

  • remove useless hard dependencies (regression since 0.15.10 about) and relaxing numpy version

[0.16.3] - 2025-03-07

Fix

  • edge-case in tract_llm export forward_kwargs

[0.16.2] - 2025-03-07

Added

  • better debug dump with shell script to reproduce failing case

Fix

  • export RNN with 2nd or 3rd outputs used only
  • export support tract_llm architecture without num-logits-to-keep
  • explicit peft dependency referenced in pyproject

[0.16.1] - 2025-03-06

Added

  • export with tract_llm merge PEFT option is set
  • CI now fail-fast
  • VERSION is set at project root to help compare with str
  • better test_suite naming for dump and debug

Change

  • export with tract_llm will use num-logits-to-keep avoiding useless compute at inference

[0.16.0] - 2025-03-03

Change

  • Breaking change -f16,--as-float16 removed and replaced by --force-module-dtype, --force-inputs-dtype that re-express this

[0.15.18] - 2025-02-28

Fix

  • PEFT loading from tract llm cli regression
  • using embedding gather with 1d tensor indices input

[0.15.17] - 2025-02-24

Fix

  • correct branching in tract selection cmd llm export

[0.15.16] - 2025-02-24

Fix

  • Avoid auto log settings except in cli's

Added

  • f32 norm options in llm cli

[0.15.15] - 2025-02-19

Fix

  • Format safety in tract_properties (avoid caret return escape and other closing quote)

[0.15.14] - 2025-02-19

Fix

  • another compress import issue

[0.15.13] - 2025-02-19

Fix

  • wrong default for compress registry llm_tract cli

[0.15.12] - 2025-02-19

Change

  • move torch_to_nnef.llm_tract.compress to torch_to_nnef.compress as it is generic

Fix

  • test suite pass again on Darwin OS
  • some remaining trace of flake8,black to ruff

[0.15.11] - 2025-02-17

Added

  • support p norm with p != 1 or 2 (including inf and -inf norms)
  • upcast to f32 norm operations if f16 inputs such as BatchNorm, norm_p2, group_norm, weight_norm
  • more tract default properties among which export command, python version, (and opt-out) username, hostname, OS info (uname -a)

[0.15.10] - 2025-02-14

Change

  • packaging/building project with uv (poetry deprecated since latest uv version are better)

[0.15.9] - 2025-02-10

Added

  • ready to support tract 0.21.9 (once regression tract side solved)

[0.15.8] - 2025-02-07

Added

  • TractNNEF now dump: tract_properties in graph.nnef with metadata infos and possible additional custom infos can be passed with specific_properties
  • TractNNEF: control over check io precision with check_io_tolerance parameters (exposed in llm cli)
  • TractNNEF: has now force_attention_inner_in_f32 that force f32 compute for SDPA in tract
  • TractNNEF: has now force_linear_accumulation_in_f32 that should be active after tract release 0.21.10 and allow accumulation in f32 for linears (opt-in)
  • cli llm: export of specific model like qwen force f32 parameters defined upper by default (for others architectures those are exposed in cli directly)

[0.15.7] - 2025-01-29

Fix

  • LLM cli export: PEFT better support
  • LLM cli export: multiple .safetensors support
  • LLMExporter decoupled and better supported

[0.15.6] - 2025-01-10

Fix

  • unsqueeze on dim -1
  • sum without arguments

Added

  • uint16 support (since PyTorch 2.4)
  • gather, sort, argsort, topk PyTorch operators support

[0.15.5] - 2024-12-13

Change

  • erf, hardswish use tract NNEF core component if inference targeted.

[0.15.4] - 2024-11-04

Fix

  • test suite working again for KhronosNNEF (full test suite green)
  • hide some warning

Change

  • export_tensors_to_nnef, export_tensors_from_disk_to_nnef as root module access
  • allow compression method to use gradients if needed
  • expose ability to manage device in QTensor mechanism with .to_device in QScheme & U8Compressor
  • better collision handling of tensor with different dtype in QTensorTractScaleOnly

Added

  • dump debug bundle with KhronosNNEF inference_target
  • new option in cli --no-verify skip all correctness checks of exported LLM model
  • new option in cli --sample-generation-total-size Number of tokens to generate in total for reference 'modes' samples npz dumped
  • new option in compress quantization min_max_q4_0_with_embeddings

[0.15.3] - 2024-10-16

Fix

  • implicit casting of dtype in mixed tensor math ops (better strategy)

Change

  • API of llm_tract compress registry functions

[0.15.2] - 2024-10-14

Fix

  • bugs with weight_and_biases operators (linear, conv, ...) with new introduced NamedTensor

Added

  • API to export only specific tensors
  • PEFT export cli support
  • maintain order in NNEF custom_extensions (as some tract extensions are order sensitive)

[0.15.1] - 2024-10-10

Fix

  • edge case of interaction between QTensor and NamedTensor
  • f16 mix and allclose check

[0.15.0] - 2024-10-09

Change

  • NNEF variable label values are now same as PyTorch module attributes naming, if Tensor are holded in any (sub-)modules

[0.14.0] - 2024-10-08

Added

  • refactor of llm_tract into sub-modules
  • added support for modes IO dump and checks

Fix

  • intel based mac tract export download correct CLI
  • expand more robust
  • align correctly all dimensional 'int' value as Int64
  • force implicit mixed inputs dtype in PyTorch math operator to add explicit casting in exported graph
  • Phi3 export correctly

[0.13.16] - 2024-10-01

Fix

  • dynamic_axes working for Llama model family

[0.13.15] - 2024-09-24

Fix

  • slice with dyn axis edge case

[0.13.14] - 2024-09-23

Added

  • Official support tract 0.21.7

[0.13.13] - 2024-09-20

Fix

  • Support QTensor for legacy (bellow 2.0), up to 1.12.0 <= torch

[0.13.12] - 2024-09-18

Fix

  • flexible checks

[0.13.11] - 2024-09-18

Fix

  • Split further functionalities & add some arguments as opt-in in LLM cli to add more reusable code

[0.13.10] - 2024-09-18

Fix

  • (missfire) mkdir parents dir if needed while cache tract binary

[0.13.9] - 2024-09-17

Fix

  • mkdir parents dir if needed while cache tract binary

[0.13.8] - 2024-09-17

Fix

  • filter more possible stdout tract (avoid to land in stderr)
  • tract inference target more robust with no subprocess shell=True and no wget needed
  • in case of potential collision while merging graph and sub-graph during torch graph parsing, auto incrementation of variable name is performed

Added

  • aten::linalg_norm basic support for p=1 and p=2 added

[0.13.7] - 2024-09-16

Fix

  • export_llm_to_tract API underlying no more need hugging face slug if only local dir.
  • export_llm_to_tract log error if IO check wrong.

[0.13.6] - 2024-09-11

Fix

  • export_llm_to_tract export cli more modular and reusable fn's

[0.13.5] - 2024-09-11

Fix

  • f16 export of LLM more stable (LayerNorm handling)
  • more robust export_llm_to_tract export cli (+ full tokenizer, config export)

[0.13.4] - 2024-09-09

Fix

  • f16 export of LLM export correctly
  • Q4_0 accurately serialize to tract

[0.13.3] - 2024-09-05

Change

  • QTensor inherit now from torch.Tensor and support any weight sharing

[0.13.2] - 2024-08-27

Fix

  • add missing arm64 in arch64 for tract downloader

[0.13.1] - 2024-08-26

Added

  • tract_llm with various tract target support

Change

  • refactor renaming_scheme -> nnef_variable_naming_scheme

Fix

  • few remaining nnef_spec_strict replaced
  • logger.warning for unofficially supported inference target fixed

[0.13.0] - 2024-08-22

Added

  • Support for explicit InferenceTarget in core function export_model_to_nnef (so far 2 variants: KhronosNNEF and TractNNEF)
  • Added KhronosNNEF test suite based on nnef-tool interpreter
  • In case of TractNNEF binary management is handled internally (no more system wide tract reference)

Change

  • refactor tract within inference_target
  • refactor module "primitives" as "aten"
  • refactor class "NamedItemOrderedSet" as "ReactiveNamedItemDict"
  • updated README in accordance with new exposed API

[0.12.3] - 2024-08-21

Added

  • support for all variants of torch.nn.functional.scaled_dot_product_attention
  • add GELU with tanh approximation option
  • slice with out of bound reformulation, to allow tract to work (ie. [-100:] on a 50 size dim)
  • new LLM pass: Mistral is passing, Gemma2 pass but some IO diff

[0.12.2] - 2024-08-19

Added

  • refactor NNEF variable naming in a ir_naming in module aside
  • new NNEF variable naming scheme natural_verbose_camel
  • added export IO support for dict/list/tuple of torch.Tensor via flattening
  • added export IO support for other object via constantization (not part of graph external)

[0.12.1] - 2024-08-09

Added

  • tract Q4_0 support
  • new llm_tract extension installable with pip install torch_to_nnef[llm_tract]
  • hold cli export_llm_to_tract for direct LLM export from any huggingface model with optional quant
  • replace scripts dir at root of the project
  • added support for Python 3.12

Removed

  • dropped support for Python 3.8
  • dropped support for unused QTensor formats

[0.11.3] - 2024-07-26

Added

  • Tested export for Llama,openELM, Phi LLM family works
  • Added support aten::ops : tril, repeat_interleave, type_as
  • Variable naming scheme: old natural_verbose option renamed raw, new option natural_verbose means 'as close as possible' to torch Python code
  • Protection against variable naming collision with input_names, output_names
  • Updated NNEF extensions to comply to tract expectations

Fix

  • Improved support aten::ops : index_ multi index gathering, masked_fill, ones_like
  • added naming for models unit-tests, 'useful' in case of failures
  • Compliance with tract>0.21.3 (introduced more restrictive definition within NNEF with different notation of scalar between float and TDim/long/int )
  • Substantial performance improvement for internals graph IR (via by example new data-structures: NamedItemOrderedSet)

[0.10.2] - 2024-06-21

Fix

  • squeeze after getting shape slice to get scalar (specific to tract to get rank 0)

[0.10.1] - 2024-04-19

Fix

  • better dynamic shape handling: remove realized shape from IR and adapt translation of slice accordingly

[0.10.0] - 2024-04-17

Removed

  • drop python 3.7.0 support

Added

  • added triu export support
  • script to export Llama2

Fix

  • Support aten::ops : ones_like, zeros_like, arange with dynamic shape
  • Support aten::ops: expand, masked_fill with dynamic shape (but no tract support)
  • more unit test of primitive
  • fix poetry source pypi

[0.9.1] - 2024-04-04

Removed

  • drop python 3.7.0 support
  • updated tract version tested against: 0.19.16, 0.20.22, 0.21.3

Added

  • (alpha) scripts/generate_qtensor_gguf_matmul.py to generate unit tests with GGUF format for tract
  • (alpha) [gguf] feature gate to support export to GGUF format and quantization
  • (alpha) Support 2 new quantization tensor type (implemented as module for now):
  • QTensorGGUF support almost all GGUF data types -> with export prototype working
  • QTensorSepParamsWithPack more flexible than GGUF format, with support of classical per group with different sizes, per channel, per weight quantisation scheme at different bit-width 1, 2, 3, 4, 8 (useful for experimentation/accuracy simulation)
  • move [dev] dependencies as a poetry group, to avoid exposition as packaged optional feature
  • new torch_version() and tract_version() utility functions now allows for direct comparison to string version "X.Y.Z"
  • Updated all tests packages torch/torch_audio/..., to torch 2.2 compatible 🎉
  • added weight_norm export support

Fix

  • support for latest scaled_dot_product_attention aten version (last PyTorch version)
  • quantization of bias as i32 at export for better support in tract (checked accuracy no-regression on bigger model)
  • additional test for quantization with PyTorch different inputs q params activated (since last tract version merged related PR)
  • custom_extractors have been refactored into sub-modules

[0.8.11] - 2024-03-04

Fix

  • linear, conv, quantized operators accurately export bias to tract
  • activations, quantized operators export output tensor scale / offset

[0.8.10] - 2024-02-23

Added

  • add, mul, div element wise operators for quantized elements

Fix

  • deconv with group now export correctly to tract

[0.8.9] - 2024-01-16

Added

  • tract_core_external in case of graph input being not i64, nor f32

[0.8.8] - 2023-11-29

Fix

  • rnn states can now be manipulated in graph even in multi-layers

[0.8.7] - 2023-11-29

Fix

  • rnn states can now be manipulated in graph
  • dynamic_axes with tensor construction such as zeros, ones (and all related variants) now produce correct dynamic matrix

[0.8.6] - 2023-10-27

Fix

  • tract_core NNEF extension added when using slice with dynamic_axes (to use tract_core_shape_of)
  • python 3.7 is now authorized again for the package even if no more supported

Added

  • tract_extra is added parameters to tract when running check_same_io_as_tract starting at tract 0.20.20

[0.8.5] - 2023-09-12

Added

  • PyTorch v2 support
  • Python 3.7 no more tested/supported as it is deprecated
  • Support Python 3.8 to 3.11 tested/supported

[0.8.4] - 2023-08-28

Fix

  • In case of RNN,GRU,LSTM we expand explicitly state initializers to batch dimensions (helping tract in case of some dynamic_axes graph formulation)
  • Refactor of torch_graph module in sub-modules

Added

  • hstack and vstack support
  • unflatten support
  • einsum support

[0.8.2] - 2023-08-02

Fix

  • slice with end being dynamic (akka max dimension size) given tract export target and dynamic_axes enabled

[0.8.1] - 2023-08-01

Fix

  • fail if tract binary not found but check_same_io_as_tract=True
  • better tract handling when check_same_io_as_tract
  • disable fft's tests for now

Added

  • Llama partial export
  • _convolution_mode aten operator (padding same and valid)

[0.8.0] - 2023-05-01

Added

  • Refactored internals in primitive/quantized with submodule and registries
  • relu6, hardswish activations

Fix

  • Support tract 0.19.15
  • Support tract 0.20.4

Removed

  • deprecated support tract 0.17 (we support only last 3 majors)
  • deprecated support of fft's ops prior to tract 0.20.0

[0.7.7] - 2023-02-20

Added

  • add narrow support
  • fix copy should not be used for tract
  • tile akka expand allow dynamic dimension as repeat

[0.7.6] - 2023-01-25

Added

  • complex support for abs
  • log10 ops supported
  • torchaudio.transform.MelSpectrogram supported out of the box

[0.7.5] - 2023-01-23

Added

  • stft, fft, ifft and basic complex number manipulations, torch now export to nnef with tract core experimental implementation in 0.19.0

[0.7.4] - 2023-01-18

Fix

  • Avoid global log config setting in export module (restrict it to test)

[0.7.3] - 2023-01-12

Fix

  • aten:Int catched even if not part of a list
  • In case a float or an int is too big it use exponential notation and may trunk part of the number at serialization by example: torch.finfo(self.dtype).min (from huggingface transformers lib).

Added

  • embedding operator
  • Albert model is passing

[0.7.2] - 2023-01-11

Fix

  • dynamic_axes generated stream variables should be better casted to NNEF tensor ref

[0.7.1] - 2023-01-10

Added

  • roll, new_zeros, zeros operators
  • pow operator now support negative and scalars as exponent

Fix

  • rsub & remainder operator with constant should be precomputed output constants
  • avg_pool1d, avg_pool2d operators now work as expected

[0.6.10] - 2022-11-07

Fix

  • aten:floor_divide new op from torch 1.13 (torch 1.13 is passing)

[0.6.9] - 2022-11-04

Fix

  • aten:size fix lost context for dyn shapes

[0.6.8] - 2022-10-31

Fix

  • aten:size expand is now consistant in nameing pattern and should be more robust

[0.6.7] - 2022-10-31

Fix

  • aten:size case with negative index is now translated correctly
  • ...-pre tract version are now handled correctly

[0.6.6] - 2022-10-21

Fix

  • Handle case with no tract binary found ( thanks to Theo 🎉 )

[0.6.5] - 2022-10-20

Fix

  • Missing use of SONOS infra

[0.6.4] - 2022-10-20

Fix

  • Push to SONOS repo as well

[0.6.3] - 2022-10-19

Fix

  • round operator is now following tract core IEE implementation and warn if vanilla NNEF version is used
  • ipdb is no more a dependency of this package
  • bump to black formatter v22 (to avoid click raising errors)
  • support tract > v0.18.0 (changed Conv1d bias expected shapes)

[0.6.1] - 2022-09-27

  • baddbmm operator is supported

Fix

  • all small fixes to have torch_to_nnef works with torch 1.12.0 and beyond (keeping backward compatibility)

[0.6.0] - 2022-09-27

Added

  • nnef_spec_strict option in export allows to export strict the NNEF spec compliant model.
  • select, group_norm, erf operators are supported.
  • gelu was rewritten with erf fragment for precision.
  • ConvTasNet is supported.
  • Wav2Vec2 encoder is supported.
  • VisionTransformer (ViT) is supported.

Fix

  • negative index in slice are now handled for fixed dimensions

Change

  • Exceptions are now unified under T2NError

[0.5.3] - 2022-09-08

Change

  • naming exported file with .nnef is no more required

[0.5.2] - 2022-09-06

Change

  • update nnef deps with real original dir since poetry now support subdirectory
  • tract v0.17.7 should make the CI tests pass again

[0.5.1] - 2022-08-17

Change

  • update nnef deps

[0.5.0] - 2022-08-16

Change

  • aten:size is now transformed in tract_core_shape_of which is against NNEF protocol specification but allow 'more' dynamic network to be expressed
  • aten:reshape allow symbolic dims as parameters

[0.4.0] - 2022-07-20

Added

  • tensor.norm with p 1 or 2
  • tensor.clamp_min(float) and tensor.clamp_max(float)

Fix

  • fix nn.MultiHeadAttention case (not self attention) allow to export Transpose

Change

  • torch quantize op lead to explicit tract_core_cast now

[0.3.4] - 2022-05-06

Fix

  • expand can be expressed with negative values and repeat within rank dim
  • Conformer Architecture now export correctly regardless the number of Attention Head

[0.3.3] - 2022-05-02

Fix

  • Quantization info are passed correctly in case of type neutral information like ((un)squeeze, transpose, split).
  • Dequantize is applied as a forced cast

[0.3.2] - 2022-04-29

Fix

  • Arity was not properly tracked in some Subgraph expansion when parameter where flipped during torch optimization process (that modified ordering), this lead to wrong matching between io of graph and subgraph during recursive process.

  • Div with an int type was not possible to cast implicitly to float by tract, to avoid rounding behavior missmatch we did had casting wrapper to handle such usecase properly.

Added

  • Better collected environment with OS, GCC, python and more package info
  • Export Q8 Conv{1,2}d and Linear
  • In Quantized network use scale/zero point of weight & input for bias export

[0.3.1] - 2022-04-22

Fix

  • LogSofmax with negative value #9
  • switch-on cast test

Added

  • dynamic_axes in export API allowing to handle streaming dimensions
  • Added aten::ops : stack, unbind,
  • Filter slice if applied without effect (slice on full range)

[0.3.0] - 2022-04-13

Fix

  • Rank expansion done right (TRUnet normalisations options works)
  • TorchTrace optimization may from time to time change signature of nn.Module so we needed to take it into account in torch_to_nnef.torch_graph module.
  • NNEF fragments file now express with their own extensions, this allows for finer grain export notation
  • macos-latest OS removed from matrix test in CI since we have limited use ( we will re-add it once tract latest version will be out )

Added

  • Added aten::ops : zeros_like, ones, expand, GLU, split, arange, chunk, layer_norm, trunc, masked_fill, clamp, to
  • Ability to export and unit-tested: Deepspeech, Conformer
  • Ability to export Wavenet, TDNN-ECAPA
  • Added LSTM with projection

[0.2.2] - 2022-04-04

Fix

  • Fix base TRUNet
  • Expose renaming scheme
  • Add id to unittest for easier debug

[0.2.1] - 2022-03-31

Fix

  • try correct parse with release workflow

[0.2.0] - 2022-03-30

Added

  • Hook system on modules (allowing to avoid jit.trace expansion replaced by custom code )
  • py.test Tract IO complaints added to errors
  • better test representation
  • LSTM/GRU/RNN handled (excepted LSTM with projection)
  • Hard tanh
  • ISO with tract check
  • Logging with log level exposed
  • TRUNet export
  • debug bundling opt-in
  • Numerous operators
  • Q8 errors explorations

[0.1.4] - 2022-03-17

Fixed

  • CI calibration finished

[0.1.1] - 2022-03-17

Added

  • Support basic models conversion (if there is not quantized layers nor LSTM in it)
  • CI is working with appropriate test suite (PyTorch->Tract ISO IO checked for ~80 cases)
  • variable renaming scheme to keep NNEF generated files short

[0.1.0] - 2022-02-28

  • First release on Sonos Org.