Skip to content

Commit

Permalink
Add engine options (PaddlePaddle#266)
Browse files Browse the repository at this point in the history 
* add engine_options for ipu_strategy

* update 01
  • Loading branch information
@gglin001
gglin001 authored Nov 5, 2021
1 parent d90295e commit 65a734d
Show file tree
Hide file tree
Showing 2 changed files with 64 additions and 97 deletions.
146 changes: 49 additions & 97 deletions paddle/fluid/pybind/pybind.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -3211,14 +3211,55 @@ All parameter, weight, gradient are variables in Paddle.
.def("enable_pattern", &ipu::IpuStrategy::enablePattern)
.def("disable_pattern", &ipu::IpuStrategy::disablePattern)
.def("is_pattern_enabled", &ipu::IpuStrategy::isPatternEnabled)
.def_property("num_ipus",
[](const ipu::IpuStrategy &self) { return self.num_ipus; },
[](ipu::IpuStrategy &self, int num_ipus) {
self.num_ipus = num_ipus;
.def_readwrite(
"num_ipus", &ipu::IpuStrategy::num_ipus,
R"DOC(Int type, set the number ipu we need. Default 1.)DOC")
.def_readwrite("batches_per_step", &ipu::IpuStrategy::batches_per_step,
R"DOC(Int type, set batches_per_step. Default 1.)DOC")
.def_readwrite(
"is_training", &ipu::IpuStrategy::is_training,
R"DOC(Bool type, True for training, False inference. Default True.)DOC")
.def_readwrite(
"need_avg_shard", &ipu::IpuStrategy::need_avg_shard,
R"DOC( Bool type, True enable avg shard, otherwise disable. Default False.)DOC")
.def_readwrite(
"batch_size", &ipu::IpuStrategy::batch_size,
R"DOC(Int type, used to make batch size fixed. Default 1.)DOC")
.def_readwrite(
"enable_fp16", &ipu::IpuStrategy::enable_fp16,
R"DOC(Bool type, True enable float16 mode, otherwise disable. Default False.)DOC")
.def_readwrite(
"save_init_onnx", &ipu::IpuStrategy::save_init_onnx,
R"DOC(Bool type, True enable save init onnx. Default False.)DOC")
.def_readwrite(
"save_last_onnx", &ipu::IpuStrategy::save_last_onnx,
R"DOC(Bool type, True enable save last onnx. Default False.)DOC")
.def_readwrite(
"save_per_n_step", &ipu::IpuStrategy::save_per_n_step,
R"DOC(Int type, Copy weights D2H per n steps. Default 1.)DOC")
.def_readwrite(
"available_mem_proportion",
&ipu::IpuStrategy::available_memory_proportion,
R"DOC(Float type. Set the available memory proportion for matmul/conv,
bigger value means more memory occupy, range [0.0f, 1.0f], 0.0 no effect,
default 0.0f.)DOC")
.def_readwrite(
"loss_scaling", &ipu::IpuStrategy::loss_scaling,
R"DOC(Float type. Set the loss scaling for mixed-precision training.
Default 1.0f.)DOC")
.def_property("engine_options",
[](const ipu::IpuStrategy &self) {
return self.popart_options.engineOptions;
},
R"DOC(
Int type, set the number ipu we need. Default 1.
)DOC")
[](ipu::IpuStrategy &self, py::dict dict) {
{
for (auto item : dict) {
auto k = item.first.cast<std::string>();
auto v = item.second.cast<std::string>();
self.popart_options.engineOptions[k] = v;
}
}
})
.def_property(
"enableGradientAccumulation",
[](const ipu::IpuStrategy &self) {
Expand Down Expand Up @@ -3256,24 +3297,6 @@ All parameter, weight, gradient are variables in Paddle.
self.popart_options.replicatedGraphCount =
replicatedGraphCount;
})
.def_property(
"batches_per_step",
[](const ipu::IpuStrategy &self) { return self.batches_per_step; },
[](ipu::IpuStrategy &self, int batches_per_step) {
self.batches_per_step = batches_per_step;
},
R"DOC(
Int type, set batches_per_step. Default 1.
)DOC")
.def_property(
"is_training",
[](const ipu::IpuStrategy &self) { return self.is_training; },
[](ipu::IpuStrategy &self, bool is_training) {
self.is_training = is_training;
},
R"DOC(
Bool type, True for training, False inference. Default True.
)DOC")
.def_property("enable_pipelining",
[](const ipu::IpuStrategy &self) {
return self.popart_options.enablePipelining;
Expand Down Expand Up @@ -3302,56 +3325,6 @@ All parameter, weight, gradient are variables in Paddle.
Bool type, True enable model sharding, otherwise disable. Default "
"False.
)DOC")
.def_property(
"need_avg_shard",
[](const ipu::IpuStrategy &self) { return self.need_avg_shard; },
[](ipu::IpuStrategy &self, bool need_avg_shard) {
self.need_avg_shard = need_avg_shard;
},
R"DOC(
Bool type, True enable avg shard, otherwise disable. Default False.
)DOC")
.def_property(
"batch_size",
[](const ipu::IpuStrategy &self) { return self.batch_size; },
[](ipu::IpuStrategy &self, int batch_size) {
self.batch_size = batch_size;
},
R"DOC(
Int type, used to make batch size fixed. Default 1.
)DOC")
.def_property(
"enable_fp16",
[](const ipu::IpuStrategy &self) { return self.enable_fp16; },
[](ipu::IpuStrategy &self, bool enable_fp16) {
self.enable_fp16 = enable_fp16;
},
R"DOC(
Bool type, True enable float16 mode, otherwise disable. Default False.)DOC")
.def_property(
"save_init_onnx",
[](const ipu::IpuStrategy &self) { return self.save_init_onnx; },
[](ipu::IpuStrategy &self, bool save_init_onnx) {
self.save_init_onnx = save_init_onnx;
},
R"DOC(
Bool type, True enable save init onnx. Default False.)DOC")
.def_property(
"save_last_onnx",
[](const ipu::IpuStrategy &self) { return self.save_last_onnx; },
[](ipu::IpuStrategy &self, bool save_last_onnx) {
self.save_last_onnx = save_last_onnx;
},
R"DOC(
Bool type, True enable save last onnx. Default False.)DOC")
.def_property(
"save_per_n_step",
[](const ipu::IpuStrategy &self) { return self.save_per_n_step; },
[](ipu::IpuStrategy &self, int save_per_n_step) {
self.save_per_n_step = save_per_n_step;
},
R"DOC(
Int type, Copy weights D2H per n steps. Default 1.)DOC")
.def_property(
"auto_recomputation",
[](const ipu::IpuStrategy &self) {
Expand All @@ -3362,7 +3335,7 @@ All parameter, weight, gradient are variables in Paddle.
static_cast<ipu::RecomputationType>(auto_recomputation);
},
R"DOC(
Int type:"
Int type:"
"0: None"
"1: Standard (Algorithm to pick checkpoints to try and minimise max liveness)"
"2: NormOnly (Only Norm Ops)"
Expand All @@ -3378,27 +3351,6 @@ All parameter, weight, gradient are variables in Paddle.
},
R"DOC(
Str type. half for fp16 partial, only work with fp16. Default float.
)DOC")
.def_property(
"available_mem_proportion",
[](const ipu::IpuStrategy &self) {
return self.available_memory_proportion;
},
[](ipu::IpuStrategy &self, float available_memory_proportion) {
self.available_memory_proportion = available_memory_proportion;
},
R"DOC(
Float type. Set the available memory proportion for matmul/conv, bigger value
means more memory occupy, range [0.0f, 1.0f], 0.0 no effect, default 0.0f.
)DOC")
.def_property(
"loss_scaling",
[](const ipu::IpuStrategy &self) { return self.loss_scaling; },
[](ipu::IpuStrategy &self, float loss_scaling) {
self.loss_scaling = loss_scaling;
},
R"DOC(
Float type. Set the loss scaling for mixed-precision training. Default 1.0f.
)DOC");

py::class_<framework::ipu::IpuCustomOpIdentifier>(m, "IpuCustomOpIdentifier")
Expand Down
15 changes: 15 additions & 0 deletions python/paddle/fluid/tests/unittests/ipu/test_ipu_strategy_ipu.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -120,5 +120,20 @@ def test_enable_patern(self):
assert v == getattr(ipu_strategy, k)


@unittest.skipIf(not paddle.is_compiled_with_ipu(),
"core is not compiled with IPU")
class TestIpuStrategyEngineOptions(unittest.TestCase):
def test_enable_patern(self):
ipu_strategy = compiler.get_ipu_strategy()
engine_conf = {
'debug.allowOutOfMemory': 'true',
'autoReport.directory': 'path',
'autoReport.all': 'true'
}
ipu_strategy.engine_options = engine_conf
for k, v in engine_conf.items():
assert v == ipu_strategy.engine_options[k]


if __name__ == "__main__":
unittest.main()

0 comments on commit 65a734d

Please sign in to comment.