All notable changes to this project will be documented in this file.
The format is based on Keep a Changelog and this project adheres to Semantic Versioning.
- Adds option --add-lsh to marian-conv which allows the LSH to be memory-mapped.
- Early stopping based on first, all, or any validation metrics via
--early-stopping-on
- Compute 8.6 support if using CUDA>=11.1
- Support for RMSNorm as drop-in replace for LayerNorm from
Biao Zhang; Rico Sennrich (2019). Root Mean Square Layer Normalization
. Enabled in Transformer model via--transformer-postprocess dar
instead ofdan
. - Extend suppression of unwanted output symbols, specifically "\n" from default vocabulary if generated by SentencePiece with byte-fallback. Deactivates with --allow-special
- Allow for fine-grained CPU intrinsics overrides when BUILD_ARCH != native e.g. -DBUILD_ARCH=x86-64 -DCOMPILE_AVX512=off
- Adds custom bias epilogue kernel.
- Adds support for fusing relu and bias addition into gemms when using cuda 11.
- Better suppression of unwanted output symbols, specifically "\n" from SentencePiece with byte-fallback. Can be deactivated with --allow-special
- Display decoder time statistics with marian-decoder --stat-freq 10 ...
- Support for MS-internal binary shortlist
- Local/global sharding with MPI training via
--sharding local
- fp16 support for factors.
- Correct training with fp16 via
--fp16
. - Dynamic cost-scaling with
--cost-scaling
. - Dynamic gradient-scaling with
--dynamic-gradient-scaling
. - Add unit tests for binary files.
- Fix compilation with OMP
- Compute aligned memory sizes using exact sizing
- Support for loading lexical shortlist from a binary blob
- Integrate a shortlist converter (which can convert a text lexical shortlist to a binary shortlist) into marian-conv with --shortlist option
- Do not set guided alignments for case augmented data if vocab is not factored
- Various fixes to enable LSH in Quicksand
- Added support to MPIWrappest::bcast (and similar) for count of type size_t
- Adding new validation metrics when training is restarted and --reset-valid-stalled is used
- Missing depth-scaling in transformer FFN
- Fixed an issue when loading intgemm16 models from unaligned memory.
- Fix building marian with gcc 9.3+ and FBGEMM
- Find MKL installed under Ubuntu 20.04 via apt-get
- Support for CUDA 11.
- General improvements and fixes for MPI handling, was essentially non-functional before (syncing, random seeds, deadlocks during saving, validation etc.)
- Allow to compile -DUSE_MPI=on with -DUSE_STATIC_LIBS=on although MPI gets still linked dynamically since it has so many dependencies.
- Fix building server with Boost 1.75
- Missing implementation for cos/tan expression operator
- Fixed loading binary models on architectures where
size_t
!=uint64_t
. - Missing float template specialisation for elem::Plus
- Broken links to MNIST data sets
- Optimize LSH for speed by treating is as a shortlist generator. No option changes in decoder
- Set REQUIRED_BIAS_ALIGNMENT = 16 in tensors/gpu/prod.cpp to avoid memory-misalignment on certain Ampere GPUs.
- For BUILD_ARCH != native enable all intrinsics types by default, can be disabled like this: -DCOMPILE_AVX512=off
- Moved FBGEMM pointer to commit c258054 for gcc 9.3+ fix
- Change compile options a la -DCOMPILE_CUDA_SM35 to -DCOMPILE_KEPLER, -DCOMPILE_MAXWELL, -DCOMPILE_PASCAL, -DCOMPILE_VOLTA, -DCOMPILE_TURING and -DCOMPILE_AMPERE
- Disable -DCOMPILE_KEPLER, -DCOMPILE_MAXWELL by default.
- Dropped support for legacy graph groups.
- Developer documentation framework based on Sphinx+Doxygen+Breathe+Exhale
- Expresion graph documentation (#788)
- Graph operators documentation (#801)
- Remove unused variable from expression graph
- Added
intgemm8(ssse3|avx|avx512)?
,intgemm16(sse2|avx|avx512)?
types to marian-conv with uses intgemm backend. Types intgemm8 and intgemm16 are hardware-agnostic, the other ones hardware-specific. - Shortlist is now always multiple-of-eight.
- Added intgemm 8/16bit integer binary architecture agnostic format.
- Add --train-embedder-rank for fine-tuning any encoder(-decoder) model for multi-lingual similarity via softmax-margin loss
- Add --logical-epoch that allows to redefine the displayed epoch counter as a multiple of n data epochs, updates or labels. Also allows to define width of fractional part with second argument.
- Add --metrics chrf for computing ChrF according to https://www.aclweb.org/anthology/W15-3049/ and SacreBLEU reference implementation
- Add --after option which is meant to replace --after-batches and --after-epochs and can take label based criteria
- Add --transformer-postprocess-top option to enable correctly normalized prenorm behavior
- Add --task transformer-base-prenorm and --task transformer-big-prenorm
- Turing and Ampere GPU optimisation support, if the CUDA version supports it.
- Printing word-level scores in marian-scorer
- Optimize LayerNormalization on CPU by 6x through vectorization (ffast-math) and fixing performance regression introduced with strides in 77a420
- Decoding multi-source models in marian-server with --tsv
- GitHub workflows on Ubuntu, Windows, and MacOS
- LSH indexing to replace short list
- ONNX support for transformer models (very experimental)
- Add topk operator like PyTorch's topk
- Use cblas_sgemm_batch instead of a for loop of cblas_sgemm on CPU as the batched_gemm implementation
- Supporting relative paths in shortlist and sqlite options
- Training and scoring from STDIN
- Support for reading from TSV files from STDIN and other sources during training and translation with options --tsv and --tsv-fields n.
- Internal optional parameter in n-best list generation that skips empty hypotheses.
- Quantized training (fixed point or log-based quantization) with --quantize-bits N command
- Support for using Apple Accelerate as the BLAS library
- Segfault of spm_train when compiled with -DUSE_STATIC_LIBS=ON seems to have gone away with update to newer SentencePiece version.
- Fix bug causing certain reductions into scalars to be 0 on the GPU backend. Removed unnecessary warp shuffle instructions.
- Do not apply dropout in embeddings layers during inference with dropout-src/trg
- Print "server is listening on port" message after it is accepting connections
- Fix compilation without BLAS installed
- Providing a single value to vector-like options using the equals sign, e.g. --models=model.npz
- Fix quiet-translation in marian-server
- CMake-based compilation on Windows
- Fix minor issues with compilation on MacOS
- Fix warnings in Windows MSVC builds using CMake
- Fix building server with Boost 1.72
- Make mini-batch scaling depend on mini-batch-words and not on mini-batch-words-ref
- In concatenation make sure that we do not multiply 0 with nan (which results in nan)
- Change Approx.epsilon(0.01) to Approx.margin(0.001) in unit tests. Tolerance is now absolute and not relative. We assumed incorrectly that epsilon is absolute tolerance.
- Fixed bug in finding .git/logs/HEAD when Marian is a submodule in another project.
- Properly record cmake variables in the cmake build directory instead of the source tree.
- Added default "none" for option shuffle in BatchGenerator, so that it works in executables where shuffle is not an option.
- Added a few missing header files in shortlist.h and beam_search.h.
- Improved handling for receiving SIGTERM during training. By default, SIGTERM triggers 'save (now) and exit'. Prior to this fix, batch pre-fetching did not check for this sigal, potentially delaying exit considerably. It now pays attention to that. Also, the default behaviour of save-and-exit can now be disabled on the command line with --sigterm exit-immediately.
- Fix the runtime failures for FASTOPT on 32-bit builds (wasm just happens to be 32-bit) because it uses hashing with an inconsistent mix of uint64_t and size_t.
- fix beam_search ABORT_IF(beamHypIdx >= beam.size(), "Out of bounds beamHypIdx??"); when enable openmp and OMP_NUM_THREADS > 1
- Remove
--clip-gemm
which is obsolete and was never used anyway - Removed
--optimize
switch, instead we now determine compute type based on binary model. - Updated SentencePiece repository to version 8336bbd0c1cfba02a879afe625bf1ddaf7cd93c5 from https://github.com/google/sentencepiece.
- Enabled compilation of SentencePiece by default since no dependency on protobuf anymore.
- Changed default value of --sentencepiece-max-lines from 10000000 to 2000000 since apparently the new version doesn't sample automatically anymore (Not quite clear how that affects quality of the vocabulary).
- Change mini-batch-fit search stopping criterion to stop at ideal binary search threshold.
- --metric bleu now always detokenizes SacreBLEU-style if a vocabulary knows how to, use bleu-segmented to compute BLEU on word ids. bleu-detok is now a synonym for bleu.
- Move label-smoothing computation into Cross-entropy node
- Move Simple-WebSocket-Server to submodule
- Python scripts start with #!/usr/bin/env python3 instead of python
- Changed compile flags -Ofast to -O3 and remove --ffinite-math
- Moved old graph groups to depracated folder
- Make cublas and cusparse handle inits lazy to save memory when unused
- Replaced exception-based implementation for type determination in FastOpt::makeScalar
- An option to print cached variables from CMake
- Add support for compiling on Mac (and clang)
- An option for resetting stalled validation metrics
- Add CMAKE options to disable compilation for specific GPU SM types
- An option to print word-level translation scores
- An option to turn off automatic detokenization from SentencePiece
- Separate quantization types for 8-bit FBGEMM for AVX2 and AVX512
- Sequence-level unliklihood training
- Allow file name templated valid-translation-output files
- Support for lexical shortlists in marian-server
- Support for 8-bit matrix multiplication with FBGEMM
- CMakeLists.txt now looks for SSE 4.2
- Purging of finished hypotheses during beam-search. A lot faster for large batches.
- Faster option look-up, up to 20-30% faster translation
- Added --cite and --authors flag
- Added optional support for ccache
- Switch to change abort to exception, only to be used in library mode
- Support for 16-bit packed models with FBGEMM
- Multiple separated parameter types in ExpressionGraph, currently inference-only
- Safe handling of sigterm signal
- Automatic vectorization of elementwise operations on CPU for tensors dims that are divisible by 4 (AVX) and 8 (AVX2)
- Replacing std::shared_ptr with custom IntrusivePtr for small objects like Tensors, Hypotheses and Expressions.
- Fp16 inference working for translation
- Gradient-checkpointing
- Replace value for INVALID_PATH_SCORE with std::numer_limits::lowest() to avoid overflow with long sequences
- Break up potential circular references for GraphGroup*
- Fix empty source batch entries with batch purging
- Clear RNN chache in transformer model, add correct hash functions to nodes
- Gather-operation for all index sizes
- Fix word weighting with max length cropping
- Fixed compilation on CPUs without support for AVX
- FastOpt now reads "n" and "y" values as strings, not as boolean values
- Fixed multiple reduction kernels on GPU
- Fixed guided-alignment training with cross-entropy
- Replace IntrusivePtr with std::uniq_ptr in FastOpt, fixes random segfaults due to thread-non-safty of reference counting.
- Make sure that items are 256-byte aligned during saving
- Make explicit matmul functions respect setting of cublasMathMode
- Fix memory mapping for mixed paramter models
- Removed naked pointer and potential memory-leak from file_stream.{cpp,h}
- Compilation for GCC >= 7 due to exception thrown in destructor
- Sort parameters by lexicographical order during allocation to ensure consistent memory-layout during allocation, loading, saving.
- Output empty line when input is empty line. Previous behavior might result in hallucinated outputs.
- Compilation with CUDA 10.1
- Combine two for-loops in nth_element.cpp on CPU
- Revert LayerNorm eps to old position, i.e. sigma' = sqrt(sigma^2 + eps)
- Downgrade NCCL to 2.3.7 as 2.4.2 is buggy (hangs with larger models)
- Return error signal on SIGTERM
- Dropped support for CUDA 8.0, CUDA 9.0 is now minimal requirement
- Removed autotuner for now, will be switched back on later
- Boost depdendency is now optional and only required for marian_server
- Dropped support for g++-4.9
- Simplified file stream and temporary file handling
- Unified node intializers, same function API.
- Remove overstuff/understuff code
- Alias options and new --task option
- Automatic detection of CPU intrisics when building with -arch=native
- First version of BERT-training and BERT-classifier, currently not compatible with TF models
- New reduction operators
- Use Cmake's ExternalProject to build NCCL and potentially other external libs
- Code for Factored Vocabulary, currently not usable yet without outside tools
- Issue with relative paths in automatically generated decoder config files
- Bug with overlapping CXX flags and building spm_train executable
- Compilation with gcc 8
- Overwriting and unsetting vector options
- Windows build with recent changes
- Bug with read-ahead buffer
- Handling of "dump-config: false" in YAML config
- Errors due to warnings
- Issue concerning failed saving with single GPU training and --sync-sgd option.
- NaN problem when training with Tensor Cores on Volta GPUs
- Fix pipe-handling
- Fix compilation with GCC 9.1
- Fix CMake build types
- Error message when using left-to-right and right-to-left models together in ensembles
- Regression tests included as a submodule
- Update NCCL to 2.4.2
- Add zlib source to Marian's source tree, builds now as object lib
- -DUSE_STATIC_LIBS=on now also looks for static versions of CUDA libraries
- Include NCCL build from github.com/marian-nmt/nccl and compile within source tree
- Set nearly all warnings as errors for Marian's own targets. Disable warnings for 3rd party
- Refactored beam search
- Word alignment generation in scorer
- Attention output generation in decoder and scorer with
--alignment soft
- Support for SentencePiece vocabularies and run-time segmentation/desegmentation
- Support for SentencePiece vocabulary training during model training
- Group training files by filename when creating vocabularies for joint vocabularies
- Updated examples
- Synchronous multi-node training (early version)
- Delayed output in line-by-line translation
- Generated word alignments include alignments for target EOS tokens
- Boost::program_options has been replaced by another CLI library
- Replace boost::file_system with Pathie
- Expansion of unambiguous command-line arguments is no longer supported
- Faster training (20-30%) by optimizing gradient popagation of biases
- Returning Moses-style hard alignments during decoding single models, ensembles and n-best lists
- Hard alignment extraction strategy taking source words that have the attention value greater than the threshold
- Refactored sync sgd for easier communication and integration with NCCL
- Smaller memory-overhead for sync-sgd
- NCCL integration (version 2.2.13)
- New binary format for saving/load of models, can be used with *.bin extension (can be memory mapped)
- Memory-mapping of graphs for inferece with
ExpressionGraph::mmap(const void* ptr)
function. (assumes *.bin model is mapped or in buffer) - Added SRU (--dec-cell sru) and ReLU (--dec-cell relu) cells to inventory of RNN cells
- RNN auto-regression layers in transformer (
--transformer-decoder-autreg rnn
), work with gru, lstm, tanh, relu, sru cells - Recurrently stacked layers in transformer (
--transformer-tied-layers 1 1 1 2 2 2
means 6 layers with 1-3 and 4-6 tied parameters, two groups of parameters) - Seamless training continuation with exponential smoothing
- A couple of bugs in "selection" (transpose, shift, cols, rows) operators during back-prob for a very specific case: one of the operators is the first operator after a branch, in that case gradient propgation might be interrupted. This did not affect any of the existing models as such a case was not present, but might have caused future models to not train properly
- Bug in mini-batch-fit, tied embeddings would result in identical embeddings in fake source and target batch. Caused under-estimation of memory usage and re-allocation
- Average Attention Networks for Transformer model
- 16-bit matrix multiplication on CPU
- Memoization for constant nodes for decoding
- Autotuning for decoding
- GPU decoding optimizations, about 2x faster decoding of transformer models
- Multi-node MPI-based training on GPUs
- Data weighting with
--data-weighting
at sentence or word level - Persistent SQLite3 corpus storage with
--sqlite file.db
- Experimental multi-node asynchronous training
- Restoring optimizer and training parameters such as learning rate, validation results, etc.
- Experimental multi-CPU training/translation/scoring with
--cpu-threads=N
- Restoring corpus iteration after training is restarted
- N-best-list scoring in marian-scorer
- Deterministic data shuffling with specific seed for SQLite3 corpus storage
- Mini-batch fitting with binary search for faster fitting
- Better batch packing due to sorting
- Missing final validation when done with training
- Differing summaries for marian-scorer when used with multiple GPUs
- SQLite3 based corpus storage for on-disk shuffling etc. with
--sqlite
- Asynchronous maxi-batch preloading
- Using transpose in SGEMM to tie embeddings in output layer
- Use valid-mini-batch size during validation with "translation" instead of mini-batch
- Normalize gradients with multi-gpu synchronous SGD
- Fix divergence between saved models and validated models in asynchronous SGD
- Option
--pretrained-model
to be used for network weights initialization with a pretrained model - Version number saved in the model file
- CMake option
-DCOMPILE_SERVER=ON
- Right-to-left training, scoring, decoding with
--right-left
- Fixed marian-server compilation with Boost 1.66
- Fixed compilation on g++-4.8.4
- Fixed compilation without marian-server if openssl is not available
- Added back gradient-dropping
- Fixed parameters initialization for
--tied-embeddings
during translation
- Fixed ensembling with language model and batched decoding
- Fixed attention reduction kernel with large matrices (added missing
syncthreads()
), which should fix stability with large batches and beam-size during batched decoding
- Option
--max-length-crop
to be used together with--max-length N
to crop sentences to length N rather than omitting them. - Experimental model with convolution over input characters
- Fixed a number of bugs for vocabulary and directory handling
- Batched translation for all model types, significant translation speed-up
- Batched translation during validation with translation
--maxi-batch-sort
option formarian-decoder
- Support for CUBLAS_TENSOR_OP_MATH mode for cublas in cuda 9.0
- The "marian-vocab" tool to create vocabularies
- Multi-gpu validation, scorer and in-training translation
- summary-mode for scorer
- New "transformer" model based on Attention is all you need
- Options specific for the transformer model
- Linear learning rate warmup with and without initial value
- Cyclic learning rate warmup
- More options for learning rate decay, including: optimizer history reset, repeated warmup
- Continuous inverted square root decay of learning (
--lr-decay-inv-sqrt
) rate based on number of updates - Exposed optimizer parameters (e.g. momentum etc. for Adam)
- Version of deep RNN-based models compatible with Nematus (
--type nematus
) - Synchronous SGD training for multi-gpu (enable with
--sync-sgd
) - Dynamic construction of complex models with different encoders and decoders, currently only available through the C++ API
- Option
--quiet
to suppress output to stderr - Option to choose different variants of optimization criterion: mean cross-entropy, perplexity, cross-entropy sum
- In-process translation for validation, uses the same memory as training
- Label Smoothing
- CHANGELOG.md
- CONTRIBUTING.md
- Swish activation function default for Transformer (https://arxiv.org/pdf/1710.05941.pdf)
- Changed shape organization to follow numpy.
- Changed option
--moving-average
to--exponential-smoothing
and inverted formula tos_t = (1 - \alpha) * s_{t-1} + \alpha * x_t
,\alpha
is now1-e4
by default - Got rid of thrust for compile-time mathematical expressions
- Changed boolean option
--normalize
to--normalize [arg=1] (=0)
. New behaviour is backwards-compatible and can also be specified as--normalize=0.6
- Renamed "s2s" binary to "marian-decoder"
- Renamed "rescorer" binary to "marian-scorer"
- Renamed "server" binary to "marian-server"
- Renamed option name
--dynamic-batching
to--mini-batch-fit
- Unified cross-entropy-based validation, supports now perplexity and other CE
- Changed
--normalize (bool)
to--normalize (float)arg
, allow to change length normalization weight asscore / pow(length, arg)
- Temporarily removed gradient dropping (
--drop-rate X
) until refactoring.