SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

FX tracing improvement (#14321)

Browse Source 
* Change the way tracing happens, enabling dynamic axes out of the box

* Update the tests and modeling xlnet

* Add the non recoding of leaf modules to avoid recording more values for the methods to record than what will be seen at tracing time (which would otherwise desynchronize the recorded values and the values that need to be given to the proxies during tracing, causing errors).

* Comments and making tracing work for gpt-j and xlnet

* Refactore things related to num_choices (and batch_size, sequence_length)

* Update fx to work on PyTorch 1.10

* Postpone autowrap_function feature usage for later

* Add copyrights

* Remove unnecessary file

* Fix issue with add_new_model_like

* Apply suggestions
This commit is contained in:
Michael Benayoun
2022-02-07 22:25:33 +01:00
committed by GitHub
parent 552f8d3091
commit 0fe17f375a
31 changed files with 349 additions and 689 deletions
Download Patch File Download Diff File Expand all files Collapse all files

43
tests/test_modeling_common.py
View File

@@ -116,8 +116,7 @@ class ModelTesterMixin:
model_tester = None
all_model_classes = ()
all_generative_model_classes = ()
fx_ready_model_classes = ()
fx_dynamic_ready_model_classes = ()
fx_compatible = False
test_torchscript = True
test_pruning = True
test_resize_embeddings = True
@@ -666,19 +665,14 @@ class ModelTesterMixin:
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
self._create_and_check_torch_fx_tracing(config, inputs_dict, output_loss=True)
def test_torch_fx_dynamic_axes(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
self._create_and_check_torch_fx_tracing(config, inputs_dict, dynamic_axes=True)
def _create_and_check_torch_fx_tracing(self, config, inputs_dict, output_loss=False, dynamic_axes=False):
if not is_torch_fx_available():
def _create_and_check_torch_fx_tracing(self, config, inputs_dict, output_loss=False):
if not is_torch_fx_available() or not self.fx_compatible:
return
configs_no_init = _config_zero_init(config) # To be sure we have no Nan
configs_no_init.return_dict = False
model_classes = self.fx_ready_model_classes if not dynamic_axes else self.fx_dynamic_ready_model_classes
for model_class in model_classes:
for model_class in self.all_model_classes:
model = model_class(config=configs_no_init)
model.to(torch_device)
model.eval()
@@ -687,8 +681,6 @@ class ModelTesterMixin:
try:
if model.config.is_encoder_decoder:
model.config.use_cache = False # FSTM still requires this hack -> FSTM should probably be refactored similar to BART afterward
input_ids = inputs["input_ids"]
decoder_attention_mask = inputs["decoder_attention_mask"]
labels = inputs.get("labels", None)
input_names = ["input_ids", "attention_mask", "decoder_input_ids", "decoder_attention_mask"]
if labels is not None:
@@ -697,17 +689,7 @@ class ModelTesterMixin:
model_output = model(**filtered_inputs)
batch_size = input_ids.shape[0]
encoder_sequence_length = input_ids.shape[1]
decoder_sequence_length = decoder_attention_mask.shape[1]
traced_model = symbolic_trace(
model,
input_names,
batch_size=batch_size if not dynamic_axes else -1,
sequence_length=[encoder_sequence_length, decoder_sequence_length] if not dynamic_axes else -1,
)
traced_model = symbolic_trace(model, input_names)
traced_output = traced_model(**filtered_inputs)
else:
input_names = ["input_ids", "attention_mask", "token_type_ids"]
@@ -729,23 +711,12 @@ class ModelTesterMixin:
model_output = model(**filtered_inputs)
rank = len(input_ids.shape)
if rank == 2:
batch_size, sequence_length = input_ids.shape
num_choices = -1
elif rank == 3:
batch_size, num_choices, sequence_length = input_ids.shape
else:
if rank not in [2, 3]:
raise NotImplementedError(
f"symbolic_trace automatic parameters inference not implemented for input of rank {rank}."
)
traced_model = symbolic_trace(
model,
input_names,
batch_size=batch_size if not dynamic_axes else -1,
sequence_length=sequence_length if not dynamic_axes else -1,
num_choices=num_choices,
)
traced_model = symbolic_trace(model, input_names)
traced_output = traced_model(**filtered_inputs)
except RuntimeError: