Add support for new model architectures
To contribute and add support for a model architecture that is not currently supported by the optimum.graphcore library, you will have to:
- Make sure the original model implementation inherits from
transformers.PreTrainedModel. This is not 100% needed, but it is highly recommended to have access to all the features. - Create a “pipelined” version of the original class. To do that:
- Inherit from both the original class and
modeling_utils.PipelineMixin - Implement the
~modeling_utils.PipelineMixin.parallelizemethod, which specifies how each part of the model should be placed on the hardware. - Implement the
~modeling_utils.PipelineMixin.deparallelizemethod, which takes a parallelized instance of the model back to its original version. This is needed to make sure that both the original and pipelined versions of the model can share the same state dictionary.
- Register the pipelined version of the class. This will enable the
IPUTrainerclass to automatically convert the original instance of a model to its pipelined counterpart.
Example: transformers.ViTForImageClassification to PipelinedViTForImageClassification
import poptorch
import transformers
from optimum.utils import logging
from optimum.graphcore.modeling_utils import PipelineMixin, get_layer_ipu, recomputation_checkpoint, register
logger = logging.get_logger(__name__)
@register(transformers.ViTForImageClassification)
class PipelinedViTForImageClassification(transformers.ViTForImageClassification, PipelineMixin):
def parallelize(self):
super().parallelize()
logger.info("---------- Device Allocation -----------")
logger.info("Embedding --> IPU 0")
self.vit.embeddings = poptorch.BeginBlock(self.vit.embeddings, "Embedding", ipu_id=0)
layer_ipu = get_layer_ipu(self.ipu_config.layers_per_ipu, self.vit.encoder.layer)
for index, layer in enumerate(self.vit.encoder.layer):
if self.ipu_config.recompute_checkpoint_every_layer:
# Put checkpoints on every encoder layer
h = recomputation_checkpoint(layer)
self._hooks.append(h)
ipu = layer_ipu[index]
logger.info(f"Encoder {index:<2} --> IPU {ipu}")
self.vit.encoder.layer[index] = poptorch.BeginBlock(layer, f"Encoder{index}", ipu_id=ipu)
last_ipu = self.ipu_config.ipus_per_replica - 1
logger.info(f"Head --> IPU {last_ipu}")
logger.info("---------------------------------------")
self.vit.layernorm = poptorch.BeginBlock(self.vit.layernorm, "LayerNorm", ipu_id=last_ipu)
self.classifier = poptorch.BeginBlock(self.classifier, "Classifier", ipu_id=last_ipu)
return selfAs you can see, you can specify where each part of the model should be put by wrapping them around poptorch.BeginBlock, which takes a layer, a block name, and an IPU ID as inputs.
To know which IPU ID to use, you can use the ipu_config.layers_per_ipu attribute, for more information check here
PipelineMixin
Transforms the model to run in an IPU pipeline.
Undoes the changes to the model done by parallelize.
You should call this function before calling save_pretrained so that the model.state_dict dictionary is fully compatible with the
original model.
from_transformers
< source >( model: PreTrainedModel ipu_config: IPUConfig )
Parameters
- model (PreTrainedModel) — The model to convert to a pipelined model.
- ipu_config (IPUConfig) —
The
IPUConfiginstance of the pipelined model.
Creates a pipelined version of model from a PreTrainedModel instance.
from_pretrained_transformers
< source >( model_name_or_path: str ipu_config: IPUConfig *model_args **kwargs )
Parameters
- model_name_or_path (
str) — The model name or path. - ipu_config (IPUConfig) —
The
IPUConfigof the pipelined model. - model_args (
Tuple[Any]) — The positional arguments to use when instantiating the model. - kwargs (
Dict[str, Any]) — The keyword arguments to use when instantiating the model.
Creates a pipelined version of a model by using the from_pretrained function.
Checks that the model has an IPUConfig attached, and returns it.