* This adds GPU-side hashing to tensors (a hash based on mumurhash3)
* The hash is used to check if parameters across nodes have diverged, if yes, resync all parameters and optimizer shards. Before it would resync every N (100 or 200) updates. Now this can be skipped if nothing diverged.
Enables loading of model checkpoints from main node only via MPI.
Until now the checkpoint needed to present in the same location on all nodes. That could be done either via writing to a shared filesystem (problematic due to bad syncing) or by manual copying to the same local location, e.g. /tmp on each node (while writing only happened to one main location).
Now, marian can resume training from only one location on the main node. The remaining nodes do not need to have access. E.g. local /tmp on the main node can be used, or race conditons on shared storage are avoided.
Also avoids creating files for logging on more than one node. This is a bit wonky, done via environment variable lookup.
This PR introduces small fixes around fp16 training and batch fitting:
* Multi-loss casts type to first loss-type before accumulation (aborted before due to missing cast)
* Throw `ShapeSizeException` if total expanded shape size exceeds numeric capacity of the maximum int value (2^31-1)
* During mini-batch-fitting, catch `ShapeSizeException` and use another sizing hint. Aborts outside mini-batch-fitting.
* Negative `--workspace -N` value allocates workspace as total available GPU memory minus N megabytes.
This PR sets default parameters for cost-scaling to 8.f 10000 1.f 8.f, i.e. when scaling scale by 8 and do not try to automatically scale up or down. This seems most stable than variable cost-scaling with larger numbers that was the default before.
This PR fixes case augmentation with multi-threaded reading. The solution is to not look at iterator::pos_ in lazy processing, rather pass it as an argument to the lazy function.
This replaces dense alignment storage and training with a sparse representation. Training speed with guided alignment matches now nearly normal training speed, regaining about 25% speed.
This is no. 1 of 2 PRs. The next one will introduce a new guided-alignment training scheme with better alignment accuracy.
* Add GCC 11 support
Some C++ Standard Library headers have been changed to no longer include other headers that they do need to depend on. As such, C++ programs that used standard library components without including the right headers will no longer compile.
The following headers are used less widely in libstdc++ and may need to be included explicitly when compiled with GCC 11:
<limits> (for std::numeric_limits)
<memory> (for std::unique_ptr, std::shared_ptr etc.)
<utility> (for std::pair, std::tuple_size, std::index_sequence etc.)
<thread> (for members of namespace std::this_thread.)
Co-authored-by: Roman Grundkiewicz <rgrundkiewicz@gmail.com>
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 of `dan`.
This PR adds cleaner suppression of unwanted output words. We identified a situation where SPM with byte-fallback can generate random bytes with output-sampling.
That is particularly harmful when that random bytes happens to be a newline symbol. Here we suppress newline in output unless explicitly wanted.
