IpuStrategy¶

class paddle.static. IpuStrategy [source]

Help users precisely control the graph building in paddle.static.IpuCompiledProgram .

Returns: The IpuStrategy instance.

Examples

           >>> 

>>> import paddle
>>> import paddle.static as static

>>> paddle.enable_static()

>>> ipu_strategy = static.IpuStrategy()

register_patch ( ) register_patch¶

Register patch function to support dynamic to static on IPU. This operation would break the dy2static functionality on CPU. Use release_patch to release the patch.

Examples

             >>> 

>>> import paddle
>>> import paddle.static as static

>>> ipu_strategy = static.IpuStrategy()

>>> ipu_strategy.register_patch()

release_patch ( ) release_patch¶

Release the registered IPU functions.

Examples

             >>> 

>>> import paddle
>>> import paddle.static as static

>>> ipu_strategy = static.IpuStrategy()

>>> ipu_strategy.release_patch()

set_optimizer ( optimizer ) set_optimizer¶

Set optimizer to ipu_strategy in dynamic mode.

Args:

optimizer (Optimizer): Optimizer to be used in training.

Returns:

None.

Examples:

                 >>> 
>>> import paddle
>>> import paddle.static as static

>>> linear = paddle.nn.Linear(10, 10)
>>> optimizer = paddle.optimizer.SGD(learning_rate=0.01,
...                                 parameters=linear.parameters())
>>> ipu_strategy = static.IpuStrategy()
>>> ipu_strategy.set_optimizer(optimizer)

                

parse_optimizer ( optimizer ) parse_optimizer¶

Parse optimizer attributes for IPU dynamic to static support. Currently only support parse lr.

Args:

optimizer (Optimizer): Optimizer to be parsed.

Returns:

Dict.

Examples:

                 >>> 

>>> import paddle
>>> import paddle.static as static

>>> linear = paddle.nn.Linear(10, 10)
>>> optimizer = paddle.optimizer.SGD(learning_rate=0.01,
...                                 parameters=linear.parameters())
>>> ipu_strategy = static.IpuStrategy()
>>> attrs = ipu_strategy.parse_optimizer(optimizer)

                

set_graph_config ( num_ipus=1, is_training=True, micro_batch_size=1, enable_manual_shard=False ) set_graph_config¶

Set graph configuration to the IpuStrategy instance.

Parameters

num_ipus (int, optional) – Number of IPU devices. Default 1, which means only use 1 IPU.
is_training (bool, optional) – True is training graph, False is inference graph. Default True, which means is training mode.
batch_size (int, optional) – The batch-size in the graph. Used to make the graph batch-size fixed, if the batch-size in the graph is dynamic. Default 1, which means the batch-size would be set 1, if the batch-size is dynamic.
enable_manual_shard (bool, optional) – Enable graph sharding or not. Only if num_ipus > 1, enable_manual_shard is able to be set True. Default False, which means disabled.

Returns

None.

Examples

             >>> 

>>> import paddle
>>> import paddle.static as static

>>> paddle.enable_static()

>>> ipu_strategy = static.IpuStrategy()
>>> ipu_strategy.set_graph_config(num_ipus=1,
...                             is_training=True,
...                             micro_batch_size=1,
...                             enable_manual_shard=False)

            

set_pipelining_config ( enable_pipelining=False, batches_per_step=1, enable_gradient_accumulation=False, accumulation_factor=1 ) set_pipelining_config¶

Set pipelining configuration to the IpuStrategy instance. Used to optimize the throughput performance.

Parameters

enable_pipelining (bool, optional) – Enable data pipelining between subgraphs. Only if enable_manual_shard=True, enable_pipelining is able to be set True. Default False, which means disabled.
batches_per_step (int, optional) – Set the batches per run in data pipelining mode. Only if enable_pipelining=True, batches_per_step is able to be set > 1. Default 1, which means no data pipelining.
enable_gradient_accumulation (bool, optional) – Enable to accumulate gradients before updating the weights in training mode. Only if enable_pipelining=True, enable_gradient_accumulation is able to be set True. Default False, which means no gradient accumulation.
accumulation_factor (int, optional) – Specify the number of micro-batches to accumulate before applying the varUpdate. Default 1, which means disable the accumulation.

Returns

None.

Examples

             >>> 

>>> import paddle
>>> import paddle.static as static

>>> paddle.enable_static()

>>> ipu_strategy = static.IpuStrategy()
>>> ipu_strategy.set_pipelining_config(enable_pipelining=False,
...                                     batches_per_step=1,
...                                     enable_gradient_accumulation=False,
...                                     accumulation_factor=1)

            

set_precision_config ( enable_fp16=False ) set_precision_config¶

Set half computation configuration to the IpuStrategy instance. Used to optimize the performance.

Parameters: enable_fp16 (bool, optional) – Enable FLOAT16 mode and transform FLOAT32 to FLOAT16. Default False, which means disable FLOAT16 mode.
Returns: None.

Examples

             >>> 

>>> import paddle
>>> import paddle.static as static

>>> paddle.enable_static()

>>> ipu_strategy = static.IpuStrategy()
>>> ipu_strategy.set_precision_config(enable_fp16=False)

add_custom_op ( paddle_op, popart_op=None, domain='custom.ops', version=1 ) add_custom_op¶

Add a mapping to use popart custom ops running on the IPU.

Parameters

paddle_op (str) – the name of custom op in paddle.
popart_op (str) – the name of custom op in popart.
domain (str) – domain name of custom op in popart.
version (int) – version of custom op in popart.

Returns

None.

Examples

             >>> 

>>> import paddle
>>> import paddle.static as static

>>> paddle.enable_static()

>>> ipu_strategy = static.IpuStrategy()
>>> ipu_strategy.add_custom_op('paddle_relu', 'popart_relu')

set_options ( options ) set_options¶

Set options from dict.

Parameters: options (dict) – dict of options.
Returns: None.

Examples

             >>> 

>>> import paddle
>>> import paddle.static as static

>>> paddle.enable_static()

>>> ipu_strategy = static.IpuStrategy()
>>> options = {'num_ipus':1, 'enable_fp16': True}
>>> ipu_strategy.set_options(options)

get_option ( option ) get_option¶

Get option.

Parameters: option (str) – name of option.
Returns: option value.

Examples

             >>> 

>>> import paddle
>>> import paddle.static as static

>>> paddle.enable_static()

>>> ipu_strategy = static.IpuStrategy()
>>> num_ipus = ipu_strategy.get_option('num_ipus')

enable_pattern ( pattern ) enable_pattern¶

Enable PopART pattern to optimize the graph.

Parameters: pattern (string) – the name of the pattern.
Returns: None.

Examples

             >>> 

>>> import paddle
>>> import paddle.static as static

>>> paddle.enable_static()

>>> ipu_strategy = static.IpuStrategy()
>>> ipu_strategy.enable_pattern("ViewSimplifyPattern")

disable_pattern ( pattern ) disable_pattern¶

Disable PopART pattern.

Parameters: pattern (string) – the name of the pattern.
Returns: None.

Examples

             >>> 

>>> import paddle
>>> import paddle.static as static

>>> paddle.enable_static()

>>> ipu_strategy = static.IpuStrategy()
>>> ipu_strategy.disable_pattern("ViewSimplifyPattern")

property num_ipus: Get the number of IPU devices from IpuStrategy instance.

property is_training: Get the boolean of training or inference from IpuStrategy instance.

property enable_pipelining: Get the boolean of enable pipelining or not from IpuStrategy instance.

property enable_fp16: Get the boolean of float16 mode or not from IpuStrategy instance.