def optimize_fp16_onnx_with_cast(input_onnx_path, optimized_onnx_path,
epsilon):
m = onnx.load(input_onnx_path)
onnx_model = OnnxModel(m)
nodes_to_remove = onnx_model.nodes()
nodes_to_add = [
onnx.helper.make_node("Cast", ["input"], ["fp32_input"],
"cast_input",
to=1),
onnx.helper.make_node("Cast", ["layer_norm.weight"],
["fp32_layer_norm.weight"],
"cast_weight",
to=1),
onnx.helper.make_node("Cast", ["layer_norm.bias"],
["fp32_layer_norm.bias"],
"cast_bias",
to=1),
onnx.helper.make_node(
"LayerNormalization",
["fp32_input", "fp32_layer_norm.weight", "fp32_layer_norm.bias"],
["fp32_output"],
"layer_norm",
epsilon=epsilon), # use fp32 epsilon
onnx.helper.make_node("Cast", ["fp32_output"], ["output"],
"cast_output",
to=10)
]
onnx_model.remove_nodes(nodes_to_remove)
onnx_model.add_nodes(nodes_to_add)
onnx_model.prune_graph()
onnx_model.save_model_to_file(optimized_onnx_path)
def optimize_fp16_onnx_no_cast(input_onnx_path, optimized_onnx_path, epsilon):
m = onnx.load(input_onnx_path)
onnx_model = OnnxModel(m)
weight_name = get_weight(onnx_model)
bias_name = get_bias(onnx_model)
nodes_to_remove = [n for n in onnx_model.nodes() if n.output[0] != weight_name and n.output[0] != bias_name]
nodes_to_remove = onnx_model.nodes()
node_to_add = onnx.helper.make_node("LayerNormalization", ["input", weight_name, bias_name], ["output"],
"layer_norm",
epsilon=epsilon)
onnx_model.remove_nodes(nodes_to_remove)
onnx_model.add_node(node_to_add)
onnx_model.prune_graph()
onnx_model.save_model_to_file(optimized_onnx_path)
def optimize_fp16_onnx_no_cast(input_onnx_path, optimized_onnx_path, epsilon):
m = onnx.load(input_onnx_path)
onnx_model = OnnxModel(m)
nodes_to_remove = onnx_model.nodes()
node_to_add = onnx.helper.make_node(
"LayerNormalization",
["input", "layer_norm.weight", "layer_norm.bias"], ["output"],
"layer_norm",
epsilon=epsilon)
onnx_model.remove_nodes(nodes_to_remove)
onnx_model.add_node(node_to_add)
onnx_model.prune_graph()
onnx_model.save_model_to_file(optimized_onnx_path)
def remove_useless_reshape_nodes(model: OnnxModel):
"""Remove reshape node that is not needed based on symbolic shape inference: input and output has same shape
"""
shape_infer = model.infer_runtime_shape(update=True)
if shape_infer is None:
return
nodes_to_remove = []
for node in model.nodes():
if node.op_type == 'Reshape':
input_shape = shape_infer.get_edge_shape(node.input[0])
output_shape = shape_infer.get_edge_shape(node.output[0])
if input_shape and output_shape and input_shape == output_shape:
logger.info(
f"Remove reshape node {node.name} since its input shape is same as output: {input_shape}"
)
nodes_to_remove.append(node)
if nodes_to_remove:
for node in nodes_to_remove:
model.replace_input_of_all_nodes(node.output[0], node.input[0])
model.remove_node(node)
model.prune_graph()
def optimize_fp16_onnx_with_cast(input_onnx_path, optimized_onnx_path,
epsilon):
m = onnx.load(input_onnx_path)
onnx_model = OnnxModel(m)
weight_name = get_weight(onnx_model)
bias_name = get_bias(onnx_model)
nodes_to_remove = [
n for n in onnx_model.nodes()
if n.output[0] != weight_name and n.output[0] != bias_name
]
nodes_to_add = [
onnx.helper.make_node("Cast", ["input"], ["fp32_input"],
"cast_input",
to=1),
onnx.helper.make_node("Cast", [weight_name],
["fp32_layer_norm.weight"],
"cast_weight",
to=1),
onnx.helper.make_node("Cast", [bias_name], ["fp32_layer_norm.bias"],
"cast_bias",
to=1),
onnx.helper.make_node(
"LayerNormalization",
["fp32_input", "fp32_layer_norm.weight", "fp32_layer_norm.bias"],
["fp32_output"],
"layer_norm",
epsilon=epsilon,
), # use fp32 epsilon
onnx.helper.make_node("Cast", ["fp32_output"], ["output"],
"cast_output",
to=10),
]
onnx_model.remove_nodes(nodes_to_remove)
onnx_model.add_nodes(nodes_to_add)
onnx_model.prune_graph()
onnx_model.save_model_to_file(optimized_onnx_path)
def auto_mixed_precision(onnx_model: OnnxModel,
op_block_list: List[str] = [
'Add', 'LayerNormalization', 'FastGelu'
]):
"""Convert GPT-2 model to mixed precision.
It detects whether original model has fp16 precision weights, and set parameters for float16 conversion automatically.
Args:
onnx_model (OnnxModel): optimized ONNX model
op_block_list (List[str], optional): . Defaults to ['Add', 'LayerNormalization', 'FastGelu']
Returns:
parameters(dict): a dictionary of parameters used in float16 conversion
"""
op_full_set = set([node.op_type for node in onnx_model.nodes()])
fp32_op_set = set(op_block_list)
fp16_op_set = op_full_set.difference(fp32_op_set)
logger.info(f"fp32 op: {fp32_op_set} fp16 op: {fp16_op_set}")
# logits is the first output
logits_output_name = onnx_model.graph().output[0].name
# We use the weight in last MatMul node to detect whether the model is stored with float16 weights from training.
is_weight_fp16_precision = False
output_name_to_node = onnx_model.output_name_to_node()
assert logits_output_name in output_name_to_node
node = output_name_to_node[logits_output_name]
last_matmul_node = None
if node.op_type == "MatMul":
last_matmul_node = node
logger.info(f"Found last MatMul node for logits: {node.name}")
initializer = None
for input in node.input:
initializer = onnx_model.get_initializer(input)
if initializer is not None:
break
# when the max difference of value after converting float to float16 is lower than a threshold (1e-6),
# we can deduce that the weights are stored in float16 precision.
max_diff = float_to_float16_max_diff(initializer)
logger.debug(
f"max diff of converting weights in last MatMul node {node.name}: {max_diff}"
)
is_weight_fp16_precision = (max_diff < 1E-6)
else:
logger.warning(
f"Failed to find MatMul node for logits. Found {node.op_type} of node {node.name}"
)
if is_weight_fp16_precision:
keep_io_types = []
node_block_list = []
else:
# When original weight is float32 precision, keep logits and last MatMul in float32 could get better precision.
keep_io_types = [logits_output_name]
node_block_list = [last_matmul_node.name]
parameters = {
"keep_io_types": keep_io_types,
"op_block_list": op_block_list,
"node_block_list": node_block_list,
"force_fp16_initializers": is_weight_fp16_precision
}
logger.info(f"auto_mixed_precision parameters: {parameters}")
onnx_model.convert_float_to_float16(use_symbolic_shape_infer=True,
**parameters)
fusion_utils = FusionUtils(onnx_model)
fusion_utils.remove_cascaded_cast_nodes()
fusion_utils.remove_useless_cast_nodes()
return parameters