Utilities for Trainer¶
This page lists all the utility functions used by
Most of those are only useful if you are studying the code of the Trainer in the library.
EvalPrediction(predictions: Union[numpy.ndarray, Tuple[numpy.ndarray]], label_ids: numpy.ndarray)¶
Evaluation output (always contains labels), to be used to compute metrics.
np.ndarray) – Predictions of the model.
np.ndarray) – Targets to be matched.
Helper function for reproducible behavior to set the seed in
int) – The seed to set.
Decorator to make all processes in distributed training wait for each local_master to do something.
int) – The rank of the local process.
CallbackHandler(callbacks, model, tokenizer, optimizer, lr_scheduler)¶
Internal class that just calls the list of callbacks in order.
DistributedTensorGatherer(world_size, num_samples, make_multiple_of=None, padding_index=- 100)¶
A class responsible for properly gathering tensors (or nested list/tuple of tensors) on the CPU by chunks.
If our dataset has 16 samples with a batch size of 2 on 3 processes and we gather then transfer on CPU at every step, our sampler will generate the following indices:
[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 0, 1]
to get something of size a multiple of 3 (so that each process gets the same dataset length). Then process 0, 1 and 2 will be responsible of making predictions for the following samples:
[0, 1, 2, 3, 4, 5]
[6, 7, 8, 9, 10, 11]
[12, 13, 14, 15, 0, 1]
The first batch treated on each process will be
So if we gather at the end of the first batch, we will get a tensor (nested list/tuple of tensor) corresponding to the following indices:
[0, 1, 6, 7, 12, 13]
If we directly concatenate our results without taking any precautions, the user will then get the predictions for the indices in this order at the end of the prediction loop:
[0, 1, 6, 7, 12, 13, 2, 3, 8, 9, 14, 15, 4, 5, 10, 11, 0, 1]
For some reason, that’s not going to roll their boat. This class is there to solve that problem.
int) – The number of processes used in the distributed training.
int) – The number of samples in our dataset.
int, optional) – If passed, the class assumes the datasets passed to each process are made to be a multiple of this argument (by adding samples).
int, optional, defaults to -100) – The padding index to use if the arrays don’t all have the same sequence length.
arraysto the internal storage, Will initialize the storage to the full size at the first arrays passed so that if we’re bound to get an OOM, it happens at the beginning.
Return the properly gathered arrays and truncate to the number of samples (since the sampler added some extras to get each process a dataset of the same length).
HfArgumentParser(dataclass_types: Union[NewType.<locals>.new_type, Iterable[NewType.<locals>.new_type]], **kwargs)¶
This subclass of argparse.ArgumentParser uses type hints on dataclasses to generate arguments.
The class is designed to play well with the native argparse. In particular, you can add more (non-dataclass backed) arguments to the parser after initialization and you’ll get the output back after parsing as an additional namespace. Optional: To create sub argument groups use the _argument_group_name attribute in the dataclass.
DebugUnderflowOverflow(model, max_frames_to_save=21, trace_batch_nums=, abort_after_batch_num=None)¶
This debug class helps detect and understand where the model starts getting very large or very small, and more importantly
infweight and activation elements.
There are 2 working modes:
Underflow/overflow detection (default)
Specific batch absolute min/max tracing without detection
Mode 1: Underflow/overflow detection
To activate the underflow/overflow detection, initialize the object with the model
debug_overflow = DebugUnderflowOverflow(model)
then run the training as normal and if
infgets detected in at least one of the weight, input or output elements this module will throw an exception and will print
max_frames_to_saveframes that lead to this event, each frame reporting
the fully qualified module name plus the class name whose
the absolute min and max value of all elements for each module weights, and the inputs and output
For example, here is the header and the last few frames in detection report for
google/mt5-smallrun in fp16 mixed precision
Detected inf/nan during batch_number=0 Last 21 forward frames: abs min abs max metadata [...] encoder.block.2.layer.1.DenseReluDense.wi_0 Linear 2.17e-07 4.50e+00 weight 1.79e-06 4.65e+00 input 2.68e-06 3.70e+01 output encoder.block.2.layer.1.DenseReluDense.wi_1 Linear 8.08e-07 2.66e+01 weight 1.79e-06 4.65e+00 input 1.27e-04 2.37e+02 output encoder.block.2.layer.1.DenseReluDense.wo Linear 1.01e-06 6.44e+00 weight 0.00e+00 9.74e+03 input 3.18e-04 6.27e+04 output encoder.block.2.layer.1.DenseReluDense T5DenseGatedGeluDense 1.79e-06 4.65e+00 input 3.18e-04 6.27e+04 output encoder.block.2.layer.1.dropout Dropout 3.18e-04 6.27e+04 input 0.00e+00 inf output
You can see here, that
T5DenseGatedGeluDense.forwardresulted in output activations, whose absolute max value was around 62.7K, which is very close to fp16’s top limit of 64K. In the next frame we have
Dropoutwhich renormalizes the weights, after it zeroed some of the elements, which pushes the absolute max value to more than 64K, and we get an overlow.
As you can see it’s the previous frames that we need to look into when the numbers start going into very large for fp16 numbers.
The tracking is done in a forward hook, which gets invoked immediately after
By default the last 21 frames are printed. You can change the default to adjust for your needs. For example
debug_overflow = DebugUnderflowOverflow(model, max_frames_to_save=100)
To validate that you have set up this debugging feature correctly, and you intend to use it in a training that may take hours to complete, first run it with normal tracing enabled for one of a few batches as explained in the next section.
Mode 2. Specific batch absolute min/max tracing without detection
The second work mode is per-batch tracing with the underflow/overflow detection feature turned off.
Let’s say you want to watch the absolute min and max values for all the ingredients of each
forwardcall of a given batch, and only do that for batches 1 and 3. Then you instantiate this class as
debug_overflow = DebugUnderflowOverflow(model, trace_batch_nums=[1,3])
And now full batches 1 and 3 will be traced using the same format as explained above. Batches are 0-indexed.
This is helpful if you know that the program starts misbehaving after a certain batch number, so you can fast-forward right to that area.
You can also specify the batch number after which to stop the training, with
debug_overflow = DebugUnderflowOverflow(model, trace_batch_nums=[1,3], abort_after_batch_num=3)
This feature is mainly useful in the tracing mode, but you can use it for any mode.
As this module measures absolute
maxof each weight of the model on every forward it’ll slow the training down. Therefore remember to turn it off once the debugging needs have been met.
nn.Module) – The model to debug.
int, optional, defaults to 21) – How many frames back to record
List[int], optional, defaults to
) – Which batch numbers to trace (turns detection off)
int, optional) – Whether to abort after a certain batch number has finished