def lower_modules_to_accelerator(model: nn.Module, trace,
export_options: ExportConfig):
import torch_glow
if hasattr(model, "encoder") and isinstance(model.encoder, RoBERTaEncoder):
backend = "NNPI"
(
submod_modelpath,
compilation_spec_dict,
inputs_function,
) = accelerator.get_modules(model, backend)[0]
submod_tracepath = accelerator.model2trace_path(submod_modelpath)
spec = torch_glow.CompilationSpec()
spec.get_settings().set_glow_backend(backend)
compilation_group = torch_glow.CompilationGroup()
spec.compilation_groups_append(compilation_group)
compilation_group_settings = compilation_group.get_settings()
compilation_group_settings.set_convert_to_fp16(True)
for k, v in compilation_spec_dict.items():
compilation_group.get_settings().backend_specific_opts_insert(k, v)
if inputs_function is not None:
input_sets = inputs_function(model, trace, export_options, None,
submod_modelpath)
compilation_group.set_input_sets(input_sets)
trace = torch_glow.to_glow_selective(
trace,
{submod_tracepath: spec},
inplace=False,
)
return trace
else:
return trace
def lower_modules_to_accelerator(model: nn.Module, trace,
export_options: ExportConfig):
import torch_glow
if hasattr(model, "encoder") and isinstance(model.encoder, RoBERTaEncoder):
backend = "NNPI"
submod_modelpath, compilation_spec_dict = accelerator.get_modules(
model, backend)[0]
submod_tracepath = accelerator.model2trace_path(submod_modelpath)
spec = torch_glow.CompilationSpec()
spec.get_settings().set_glow_backend(backend)
compilation_group = torch_glow.CompilationGroup()
spec.compilation_groups_append(compilation_group)
compilation_group_settings = compilation_group.get_settings()
compilation_group_settings.set_convert_to_fp16(True)
for k, v in compilation_spec_dict.items():
compilation_group.get_settings().backend_specific_opts_insert(k, v)
# Todod: @input decorator dose not work properly, fixing it later
# input_sets = inputs.input_process(model, export_options, None, submod_tracepath)
input_sets = accelerator_transformerLayers_inputs(
model, trace, export_options, None, submod_tracepath)
compilation_group.set_input_sets(input_sets)
trace = torch_glow.to_glow_selective(
trace,
{submod_tracepath: spec},
inplace=False,
)
return trace
else:
return trace
def lower_modules_to_accelerator(model: nn.Module,
trace,
export_options: ExportConfig,
throughput_optimize=False):
# Raise error if accelerator could not be imported
if not accelerator_lowering_supported:
raise RuntimeError("Accelerator Lowering not supported!")
import torch_glow
log_accelerator_feature_usage("build.NNPI")
if ((hasattr(model, "encoder")
and isinstance(model.encoder, RoBERTaEncoder))
or (hasattr(model, "representation")
and isinstance(model.representation, AcceleratorBiLSTM)) or
(hasattr(model, "lower_module")
# Internal CNN LM module to add accelerator support.
and type(model.lower_module).__qualname__ == "CNNLowerModule")):
backend = "NNPI"
(
submod_modelpath,
compilation_spec_dict,
inputs_function,
) = accelerator.get_modules(model, backend)[0]
submod_tracepath = accelerator.model2trace_path(submod_modelpath)
spec = torch_glow.CompilationSpec()
spec.get_settings().set_glow_backend(backend)
compilation_group = torch_glow.CompilationGroup()
spec.compilation_groups_append(compilation_group)
compilation_group_settings = compilation_group.get_settings()
compilation_group_settings.set_convert_to_fp16(True)
# Override the options for throughput-optimized case
if throughput_optimize:
compilation_spec_dict["NNPI_IceCores"] = "4"
compilation_spec_dict["NNPINumParallelChunks"] = "4"
compilation_group_settings.set_replication_count(3)
for k, v in compilation_spec_dict.items():
compilation_group.get_settings().backend_specific_opts_insert(k, v)
if inputs_function is not None:
input_sets = inputs_function(model, trace, export_options, None,
submod_modelpath)
else:
raise RuntimeError(
"inputs_function needs to be specified in accelerator decorator"
)
compilation_group.set_input_sets(input_sets)
trace = torch_glow.to_glow_selective(
trace,
{submod_tracepath: spec},
inplace=False,
)
return trace
else:
return trace
def lower_modules_to_accelerator(model, trace, seq_padding_control,
batch_padding_control):
import torch_glow
if hasattr(model, "encoder") and isinstance(model.encoder, RoBERTaEncoder):
backend = "NNPI"
submod_modelpath, compilation_spec_dict = accelerator.get_modules(
model, backend)[0]
submod_tracepath = accelerator.model2trace_path(submod_modelpath)
embedding_dim = model.encoder.encoder.transformer.token_embedding.embedding_dim
spec = torch_glow.CompilationSpec()
spec.get_settings().set_glow_backend(backend)
compilation_group = torch_glow.CompilationGroup()
spec.compilation_groups_append(compilation_group)
compilation_group_settings = compilation_group.get_settings()
compilation_group_settings.set_convert_to_fp16(True)
for k, v in compilation_spec_dict.items():
compilation_group.get_settings().backend_specific_opts_insert(k, v)
for seq_len in seq_padding_control:
if seq_len <= 0:
continue
for batch_size in batch_padding_control:
if batch_size <= 0:
continue
input1 = torch.randn([seq_len, batch_size, embedding_dim],
dtype=torch.float32)
input2 = torch.randn([batch_size, seq_len]).bool()
input_specs = torch_glow.input_specs_from_tensors(
[input1, input2])
compilation_group.input_sets_append(input_specs)
trace = torch_glow.to_glow_selective(
trace,
{submod_tracepath: spec},
inplace=False,
)
return trace
else:
return trace
def lower_modules_to_accelerator(
model: nn.Module, trace, export_options: ExportConfig, throughput_optimize=False
):
# Raise error if accelerator could not be imported
if not accelerator_lowering_supported:
raise RuntimeError("Accelerator Lowering not supported!")
import torch_glow
log_accelerator_feature_usage("build.NNPI")
if (
(hasattr(model, "encoder") and isinstance(model.encoder, RoBERTaEncoder))
or (
hasattr(model, "representation")
and isinstance(model.representation, AcceleratorBiLSTM)
)
or (
hasattr(model, "lower_module")
# Internal CNN LM module to add accelerator support.
and type(model.lower_module).__qualname__ == "CNNLowerModule"
)
):
backend = "NNPI"
backend_qualifier = ""
if throughput_optimize:
backend_qualifier = ":throughput_optimized"
modules_to_lower = accelerator.get_modules(model, backend + backend_qualifier)
if len(modules_to_lower) < 1:
raise RuntimeError("Need at least one module to lower to accelerator")
elif len(modules_to_lower) > 1:
print(f"Warning. Received {len(modules_to_lower)} modules to lower.")
print("Warning. Only lowering first module.")
(
submod_modelpath,
compilation_spec_dict,
inputs_function,
) = modules_to_lower[0]
submod_tracepath = accelerator.model2trace_path(submod_modelpath)
spec = torch_glow.CompilationSpec()
spec.get_settings().set_glow_backend(backend)
compilation_group = torch_glow.CompilationGroup()
spec.compilation_groups_append(compilation_group)
compilation_group_settings = compilation_group.get_settings()
# Set values from dict that are not set via backend-specific opts
compilation_group_settings.set_convert_to_fp16(
compilation_spec_dict.pop("glow:ConvertToFP16", "true") in ["true", "True"]
)
compilation_group_settings.set_replication_count(
int(compilation_spec_dict.pop("glow:ReplicationCount", "1"))
)
for k, v in compilation_spec_dict.items():
compilation_group.get_settings().backend_specific_opts_insert(k, v)
if inputs_function is not None:
input_sets = inputs_function(
model, trace, export_options, None, submod_modelpath
)
else:
raise RuntimeError(
"inputs_function needs to be specified in accelerator decorator"
)
compilation_group.set_input_sets(input_sets)
trace = torch_glow.to_glow_selective(
trace,
{submod_tracepath: spec},
inplace=False,
)
return trace
else:
return trace
