code stringlengths 17 6.64M |
|---|
def cpu(partition):
partition.device = torch.device('cpu')
return nn.Module.cpu(partition)
|
def cuda(partition, device=None):
if (device is None):
device = torch.cuda.current_device()
partition.device = torch.device(device)
return nn.Module.cuda(partition, partition.device)
|
def to(partition, *args, **kwargs):
device = None
if ('device' in kwargs):
device = kwargs['device']
elif ('tensor' in kwargs):
device = kwargs['tensor'].device
if args:
if isinstance(args[0], (torch.device, int, str)):
device = args[0]
if torch.is_tensor(args[0]):
device = args[0].device
if (not (device is None)):
partition.device = torch.device(device)
return nn.Module.to(partition, *args, **kwargs)
|
def create_pipeline_configuration(DEBUG=False, batch_size=32):
config = {'batch_dim': 0, 'depth': 10000, 'basic_blocks': (T5LayerNorm, Linear, StatelessEmbedding, CrossEntropyLoss, T5Block, Dropout), 'model_inputs': {'attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'is_batched': True, 'used_by': [0, 1, 2, 3, 4, 5]}, 'decoder_attention_mask': {'shape': torch.Size([32, 1, 4, 4]), 'dtype': torch.float32, 'is_batched': True, 'used_by': [5, 6, 7]}, 'decoder_input_ids': {'shape': torch.Size([32, 4]), 'dtype': torch.int64, 'is_batched': True, 'used_by': [0, 5]}, 'input_ids': {'shape': torch.Size([32, 64]), 'dtype': torch.int64, 'is_batched': True, 'used_by': [0]}, 'inverted_encoder_attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'is_batched': True, 'used_by': [5, 6, 7]}, 'lm_labels': {'shape': torch.Size([32, 4]), 'dtype': torch.int64, 'is_batched': True, 'used_by': [7]}}, 'model_outputs': {'T5ForConditionalGeneration/CrossEntropyLoss[lm_loss]': {'shape': torch.Size([1]), 'dtype': torch.float32, 'is_batched': False, 'created_by': 7}}, 'stages': {0: {'stage_cls': Partition0, 'inputs': {'attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'decoder_input_ids': {'shape': torch.Size([32, 4]), 'dtype': torch.int64, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'input_ids': {'shape': torch.Size([32, 64]), 'dtype': torch.int64, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[decoder]/Tensor::size_117': {'shape': torch.Size([2]), 'dtype': torch.Size, 'req_grad': False, 'is_batched': False, 'used_by': [5]}, 'T5ForConditionalGeneration/Parameter[shared_embed_weight]': {'shape': torch.Size([32100, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': False, 'used_by': [5]}, 'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_1': {'shape': torch.Size([32, 32, 64, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [1]}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[3]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [1]}}, 'devices': [('cpu' if DEBUG else 'cuda:0')], 'stage_depth': 7}, 1: {'stage_cls': Partition1, 'inputs': {'attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_1': {'shape': torch.Size([32, 32, 64, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 0}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[3]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 0}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_2': {'shape': torch.Size([32, 32, 64, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [2]}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[7]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [2]}}, 'devices': [('cpu' if DEBUG else 'cuda:1')], 'stage_depth': 6}, 2: {'stage_cls': Partition2, 'inputs': {'attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_2': {'shape': torch.Size([32, 32, 64, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 1}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[7]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 1}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_3': {'shape': torch.Size([32, 32, 64, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [3]}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[11]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [3]}}, 'devices': [('cpu' if DEBUG else 'cuda:2')], 'stage_depth': 5}, 3: {'stage_cls': Partition3, 'inputs': {'attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_3': {'shape': torch.Size([32, 32, 64, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 2}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[11]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 2}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_4': {'shape': torch.Size([32, 32, 64, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [4]}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[15]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [4]}}, 'devices': [('cpu' if DEBUG else 'cuda:3')], 'stage_depth': 4}, 4: {'stage_cls': Partition4, 'inputs': {'attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_4': {'shape': torch.Size([32, 32, 64, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 3}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[15]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 3}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_5': {'shape': torch.Size([32, 32, 64, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [5]}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[19]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [5]}}, 'devices': [('cpu' if DEBUG else 'cuda:4')], 'stage_depth': 3}, 5: {'stage_cls': Partition5, 'inputs': {'attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'decoder_attention_mask': {'shape': torch.Size([32, 1, 4, 4]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'decoder_input_ids': {'shape': torch.Size([32, 4]), 'dtype': torch.int64, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'inverted_encoder_attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[decoder]/Tensor::size_117': {'shape': torch.Size([2]), 'dtype': torch.Size, 'req_grad': False, 'is_batched': False, 'created_by': 0}, 'T5ForConditionalGeneration/Parameter[shared_embed_weight]': {'shape': torch.Size([32100, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': False, 'created_by': 0}, 'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_5': {'shape': torch.Size([32, 32, 64, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 4}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[19]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 4}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]_6': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [6]}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___128': {'shape': torch.Size([32, 4, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [6]}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___130_6': {'shape': torch.Size([32, 32, 4, 4]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [6]}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___132_6': {'shape': torch.Size([32, 32, 4, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [6]}}, 'devices': [('cpu' if DEBUG else 'cuda:5')], 'stage_depth': 2}, 6: {'stage_cls': Partition6, 'inputs': {'decoder_attention_mask': {'shape': torch.Size([32, 1, 4, 4]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'inverted_encoder_attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]_6': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 5}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___128': {'shape': torch.Size([32, 4, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 5}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___130_6': {'shape': torch.Size([32, 32, 4, 4]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 5}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___132_6': {'shape': torch.Size([32, 32, 4, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 5}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]_7': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [7]}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___130_7': {'shape': torch.Size([32, 32, 4, 4]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [7]}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___132_7': {'shape': torch.Size([32, 32, 4, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [7]}, 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[12]': {'shape': torch.Size([32, 4, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [7]}}, 'devices': [('cpu' if DEBUG else 'cuda:6')], 'stage_depth': 1}, 7: {'stage_cls': Partition7, 'inputs': {'decoder_attention_mask': {'shape': torch.Size([32, 1, 4, 4]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'inverted_encoder_attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'lm_labels': {'shape': torch.Size([32, 4]), 'dtype': torch.int64, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]_7': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 6}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___130_7': {'shape': torch.Size([32, 32, 4, 4]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 6}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___132_7': {'shape': torch.Size([32, 32, 4, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 6}, 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[12]': {'shape': torch.Size([32, 4, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 6}}, 'outputs': {'T5ForConditionalGeneration/CrossEntropyLoss[lm_loss]': {'shape': torch.Size([1]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': False, 'used_by': [(- 1)]}}, 'devices': [('cpu' if DEBUG else 'cuda:7')], 'stage_depth': 0}}}
batch_dim = config['batch_dim']
for d in chain(config['model_inputs'].values(), config['model_outputs'].values()):
if d['is_batched']:
shape = d['shape']
d['shape'] = torch.Size(((shape[:batch_dim] + (batch_size,)) + shape[(batch_dim + 1):]))
for s in config['stages'].values():
for d in chain(s['inputs'].values(), s['outputs'].values()):
if d['is_batched']:
shape = d['shape']
d['shape'] = torch.Size(((shape[:batch_dim] + (batch_size,)) + shape[(batch_dim + 1):]))
return config
|
class Partition0(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[encoder]/StatelessEmbedding[embed_tokens]', 'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[0]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[1]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[2]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[3]']
TENSORS = ['T5ForConditionalGeneration/Parameter[shared_embed_weight]']
def __init__(self, layers, tensors, device='cuda:0'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1]
self.lookup = {'l_0': 'encoder.embed_tokens', 'l_1': 'encoder.dropout', 'l_2': 'encoder.0', 'l_3': 'encoder.1', 'l_4': 'encoder.2', 'l_5': 'encoder.3', 'p_0': 'shared_embed_weight'}
self.to(self.device)
def forward(self, *args):
(attention_mask, decoder_input_ids, input_ids) = unflatten(args, self.input_structure)
t_0 = decoder_input_ids.size()
t_1 = input_ids.size()
t_1 = t_1[(- 1)]
t_1 = input_ids.view((- 1), t_1)
t_1 = self.l_0(self.p_0, t_1)
t_1 = self.l_1(t_1)
t_1 = self.l_2(t_1, attention_mask=attention_mask, position_bias=None, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_2 = t_1[0]
t_1 = t_1[1]
t_2 = self.l_3(t_2, attention_mask=attention_mask, position_bias=t_1, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_2 = self.l_4(t_2, attention_mask=attention_mask, position_bias=t_1, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_2 = self.l_5(t_2, attention_mask=attention_mask, position_bias=t_1, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
return list(flatten((t_0, self.p_0, t_1, t_2)))
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, *args, **kwargs):
return load_state_dict(self, *args, **kwargs)
def named_parameters(self, *args, **kwargs):
return named_parameters(self, *args, **kwargs)
def named_buffers(self, *args, **kwargs):
return named_buffers(self, *args, **kwargs)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition1(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[encoder]/T5Block[4]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[5]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[6]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[7]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:1'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1]
self.lookup = {'l_0': 'encoder.4', 'l_1': 'encoder.5', 'l_2': 'encoder.6', 'l_3': 'encoder.7'}
self.to(self.device)
def forward(self, *args):
(attention_mask, x0, x1) = unflatten(args, self.input_structure)
t_0 = self.l_0(x1, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_1(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_2(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_3(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
return list(flatten((x0, t_0)))
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, *args, **kwargs):
return load_state_dict(self, *args, **kwargs)
def named_parameters(self, *args, **kwargs):
return named_parameters(self, *args, **kwargs)
def named_buffers(self, *args, **kwargs):
return named_buffers(self, *args, **kwargs)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition2(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[encoder]/T5Block[8]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[9]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[10]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[11]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:2'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1]
self.lookup = {'l_0': 'encoder.8', 'l_1': 'encoder.9', 'l_2': 'encoder.10', 'l_3': 'encoder.11'}
self.to(self.device)
def forward(self, *args):
(attention_mask, x0, x1) = unflatten(args, self.input_structure)
t_0 = self.l_0(x1, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_1(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_2(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_3(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
return list(flatten((x0, t_0)))
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, *args, **kwargs):
return load_state_dict(self, *args, **kwargs)
def named_parameters(self, *args, **kwargs):
return named_parameters(self, *args, **kwargs)
def named_buffers(self, *args, **kwargs):
return named_buffers(self, *args, **kwargs)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition3(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[encoder]/T5Block[12]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[13]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[14]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[15]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:3'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1]
self.lookup = {'l_0': 'encoder.12', 'l_1': 'encoder.13', 'l_2': 'encoder.14', 'l_3': 'encoder.15'}
self.to(self.device)
def forward(self, *args):
(attention_mask, x0, x1) = unflatten(args, self.input_structure)
t_0 = self.l_0(x1, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_1(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_2(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_3(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
return list(flatten((x0, t_0)))
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, *args, **kwargs):
return load_state_dict(self, *args, **kwargs)
def named_parameters(self, *args, **kwargs):
return named_parameters(self, *args, **kwargs)
def named_buffers(self, *args, **kwargs):
return named_buffers(self, *args, **kwargs)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition4(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[encoder]/T5Block[16]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[17]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[18]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[19]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:4'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1]
self.lookup = {'l_0': 'encoder.16', 'l_1': 'encoder.17', 'l_2': 'encoder.18', 'l_3': 'encoder.19'}
self.to(self.device)
def forward(self, *args):
(attention_mask, x0, x1) = unflatten(args, self.input_structure)
t_0 = self.l_0(x1, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_1(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_2(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_3(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
return list(flatten((x0, t_0)))
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, *args, **kwargs):
return load_state_dict(self, *args, **kwargs)
def named_parameters(self, *args, **kwargs):
return named_parameters(self, *args, **kwargs)
def named_buffers(self, *args, **kwargs):
return named_buffers(self, *args, **kwargs)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition5(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[encoder]/T5Block[20]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[21]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[22]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[23]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5LayerNorm[final_layer_norm]', 'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]', 'T5ForConditionalGeneration/T5Stack[decoder]/StatelessEmbedding[embed_tokens]', 'T5ForConditionalGeneration/T5Stack[decoder]/Dropout[dropout]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[0]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:5'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1, 1, 1, 1, 1, 1]
self.lookup = {'l_0': 'encoder.20', 'l_1': 'encoder.21', 'l_2': 'encoder.22', 'l_3': 'encoder.23', 'l_4': 'encoder.final_layer_norm', 'l_5': 'encoder.dropout', 'l_6': 'decoder.embed_tokens', 'l_7': 'decoder.dropout', 'l_8': 'decoder.0'}
self.to(self.device)
def forward(self, *args):
(attention_mask, decoder_attention_mask, decoder_input_ids, inverted_encoder_attention_mask, x0, x1, x2, x3) = unflatten(args, self.input_structure)
t_0 = self.l_0(x3, attention_mask=attention_mask, position_bias=x2, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_1(t_0, attention_mask=attention_mask, position_bias=x2, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_2(t_0, attention_mask=attention_mask, position_bias=x2, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_3(t_0, attention_mask=attention_mask, position_bias=x2, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_4(t_0)
t_0 = self.l_5(t_0)
t_1 = x0[(- 1)]
t_1 = decoder_input_ids.view((- 1), t_1)
t_1 = self.l_6(x1, t_1)
t_1 = self.l_7(t_1)
t_1 = self.l_8(t_1, attention_mask=decoder_attention_mask, position_bias=None, encoder_hidden_states=t_0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=None)
t_2 = t_1[0]
t_3 = t_1[1]
t_1 = t_1[2]
return list(flatten((t_0, t_2, t_3, t_1)))
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, *args, **kwargs):
return load_state_dict(self, *args, **kwargs)
def named_parameters(self, *args, **kwargs):
return named_parameters(self, *args, **kwargs)
def named_buffers(self, *args, **kwargs):
return named_buffers(self, *args, **kwargs)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition6(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[decoder]/T5Block[1]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[2]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[3]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[4]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[5]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[6]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[7]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[8]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[9]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[10]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[11]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[12]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:6'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1, 1, 1, 1]
self.lookup = {'l_0': 'decoder.1', 'l_1': 'decoder.2', 'l_2': 'decoder.3', 'l_3': 'decoder.4', 'l_4': 'decoder.5', 'l_5': 'decoder.6', 'l_6': 'decoder.7', 'l_7': 'decoder.8', 'l_8': 'decoder.9', 'l_9': 'decoder.10', 'l_10': 'decoder.11', 'l_11': 'decoder.12'}
self.to(self.device)
def forward(self, *args):
(decoder_attention_mask, inverted_encoder_attention_mask, x0, x1, x2, x3) = unflatten(args, self.input_structure)
t_0 = self.l_0(x1, attention_mask=decoder_attention_mask, position_bias=x2, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x3)
t_0 = self.l_1(t_0, attention_mask=decoder_attention_mask, position_bias=x2, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x3)
t_0 = self.l_2(t_0, attention_mask=decoder_attention_mask, position_bias=x2, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x3)
t_0 = self.l_3(t_0, attention_mask=decoder_attention_mask, position_bias=x2, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x3)
t_0 = self.l_4(t_0, attention_mask=decoder_attention_mask, position_bias=x2, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x3)
t_0 = self.l_5(t_0, attention_mask=decoder_attention_mask, position_bias=x2, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x3)
t_0 = self.l_6(t_0, attention_mask=decoder_attention_mask, position_bias=x2, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x3)
t_0 = self.l_7(t_0, attention_mask=decoder_attention_mask, position_bias=x2, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x3)
t_0 = self.l_8(t_0, attention_mask=decoder_attention_mask, position_bias=x2, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x3)
t_0 = self.l_9(t_0, attention_mask=decoder_attention_mask, position_bias=x2, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x3)
t_0 = self.l_10(t_0, attention_mask=decoder_attention_mask, position_bias=x2, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x3)
t_0 = self.l_11(t_0, attention_mask=decoder_attention_mask, position_bias=x2, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x3)
return list(flatten((x0, x2, x3, t_0)))
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, *args, **kwargs):
return load_state_dict(self, *args, **kwargs)
def named_parameters(self, *args, **kwargs):
return named_parameters(self, *args, **kwargs)
def named_buffers(self, *args, **kwargs):
return named_buffers(self, *args, **kwargs)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition7(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[decoder]/T5Block[13]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[14]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[15]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[16]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[17]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[18]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[19]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[20]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[21]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[22]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[23]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5LayerNorm[final_layer_norm]', 'T5ForConditionalGeneration/T5Stack[decoder]/Dropout[dropout]', 'T5ForConditionalGeneration/Linear[lm_head]', 'T5ForConditionalGeneration/CrossEntropyLoss[lm_loss]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:7'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1, 1, 1, 1, 1]
self.lookup = {'l_0': 'decoder.13', 'l_1': 'decoder.14', 'l_2': 'decoder.15', 'l_3': 'decoder.16', 'l_4': 'decoder.17', 'l_5': 'decoder.18', 'l_6': 'decoder.19', 'l_7': 'decoder.20', 'l_8': 'decoder.21', 'l_9': 'decoder.22', 'l_10': 'decoder.23', 'l_11': 'decoder.final_layer_norm', 'l_12': 'decoder.dropout', 'l_13': 'lm_head', 'l_14': 'lm_loss'}
self.to(self.device)
def forward(self, *args):
(decoder_attention_mask, inverted_encoder_attention_mask, lm_labels, x0, x1, x2, x3) = unflatten(args, self.input_structure)
t_0 = self.l_0(x3, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_1(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_2(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_3(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_4(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_5(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_6(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_7(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_8(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_9(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_10(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_11(t_0)
t_0 = self.l_12(t_0)
t_0 = (t_0 * 0.03125)
t_0 = self.l_13(t_0)
t_1 = t_0.size((- 1))
t_1 = t_0.view((- 1), t_1)
t_0 = lm_labels.view((- 1))
t_0 = self.l_14(t_1, t_0)
return (t_0,)
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, *args, **kwargs):
return load_state_dict(self, *args, **kwargs)
def named_parameters(self, *args, **kwargs):
return named_parameters(self, *args, **kwargs)
def named_buffers(self, *args, **kwargs):
return named_buffers(self, *args, **kwargs)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
def traverse_model(module: nn.Module, depth: int, prefix: Optional[str]=None, basic_blocks: Tuple[Type[nn.Module]]=(), full: bool=False) -> Iterator[Tuple[(nn.Module, str, nn.Module, Optional[bool])]]:
'\n iterate over model layers yielding the layer,layer_scope,encasing_module\n Parameters:\n -----------\n model:\n the model to iterate over\n depth:\n how far down in the model tree to go\n basic_blocks:\n a list of modules that if encountered will not be broken down\n full:\n whether to yield only layers specified by the depth and basic_block options or to yield all layers\n '
if (prefix is None):
prefix = type(module).__name__
for (name, sub_module) in module.named_children():
scope = (((prefix + '/') + type(sub_module).__name__) + f'[{name}]')
if ((len(list(sub_module.children())) == 0) or isinstance(sub_module, tuple(basic_blocks)) or (depth == 0)):
if full:
(yield (sub_module, scope, module, True))
else:
(yield (sub_module, scope, module))
else:
if full:
(yield (sub_module, scope, module, False))
(yield from traverse_model(sub_module, (depth - 1), scope, basic_blocks, full))
|
def layerDict(model: nn.Module, depth=1000, basic_blocks=()) -> Dict[(str, nn.Module)]:
return {s: l for (l, s, _) in traverse_model(model, depth, basic_blocks=basic_blocks)}
|
def traverse_params_buffs(module: nn.Module, prefix: Optional[str]=None) -> Iterator[Tuple[(torch.tensor, str)]]:
"\n iterate over model's buffers and parameters yielding obj,obj_scope\n\n Parameters:\n -----------\n model:\n the model to iterate over\n "
if (prefix is None):
prefix = type(module).__name__
for (param_name, param) in module.named_parameters(recurse=False):
param_scope = f'{prefix}/{type(param).__name__}[{param_name}]'
(yield (param, param_scope))
for (buffer_name, buffer) in module.named_buffers(recurse=False):
buffer_scope = f'{prefix}/{type(buffer).__name__}[{buffer_name}]'
(yield (buffer, buffer_scope))
for (name, sub_module) in module.named_children():
(yield from traverse_params_buffs(sub_module, (((prefix + '/') + type(sub_module).__name__) + f'[{name}]')))
|
def tensorDict(model: nn.Module) -> OrderedDict[(str, Tensor)]:
return collections.OrderedDict(((s, t) for (t, s) in traverse_params_buffs(model)))
|
def move_tensors(ts, device):
def move(t):
if isinstance(t, (nn.Module, Tensor)):
return t.to(device)
return t
return nested_map(move, ts)
|
def nested_map(func, ts, full=False):
if isinstance(ts, torch.Size):
return func(ts)
elif isinstance(ts, (list, tuple, set)):
return type(ts)((nested_map(func, t, full=full) for t in ts))
elif isinstance(ts, dict):
return {k: nested_map(func, v, full=full) for (k, v) in ts.items()}
elif (isinstance(ts, slice) and full):
start = nested_map(func, ts.start, full=full)
stop = nested_map(func, ts.stop, full=full)
step = nested_map(func, ts.step, full=full)
return slice(start, stop, step)
return func(ts)
|
def flatten(ts):
if isinstance(ts, torch.Size):
(yield ts)
elif isinstance(ts, (list, tuple, set)):
(yield from chain(*[flatten(t) for t in ts]))
elif isinstance(ts, dict):
(yield from chain(*[flatten(t) for (k, t) in sorted(ts.items(), key=(lambda t: t[0]))]))
else:
(yield ts)
|
def unflatten(xs, structure):
return _unflatten(xs, structure)[0]
|
def _unflatten(xs, structure):
if isinstance(structure, torch.Size):
return (xs[0], 1)
if (not isinstance(structure, (list, tuple, set, dict))):
return (xs[0], 1)
if isinstance(structure, (list, tuple, set)):
offset = 0
elements = []
for s in structure:
(e, n) = _unflatten(xs[offset:], s)
elements.append(e)
offset += n
return (type(structure)(elements), offset)
assert isinstance(structure, dict)
offset = 0
elements = dict()
for (k, v) in sorted(structure.items(), key=(lambda t: t[0])):
(e, n) = _unflatten(xs[offset:], v)
elements[k] = e
offset += n
return (elements, offset)
|
def state_dict(partition, *args, **kwargs):
state = nn.Module.state_dict(partition, *args, **kwargs)
lookup = partition.lookup
result = dict()
for (k, v) in state.items():
if (k in lookup):
result[lookup[k]] = v
else:
assert ('.' in k)
split_idx = k.find('.')
new_k = (lookup[k[:split_idx]] + k[split_idx:])
result[new_k] = v
return result
|
def load_state_dict(partition, state_dict, strict=True):
reverse_lookup = {v: k for (k, v) in partition.lookup.items()}
device = partition.device
keys = list(partition.state_dict(None).keys())
new_state = dict()
for k in keys:
if (k in reverse_lookup):
new_state[reverse_lookup[k]] = state_dict[k].to(device)
continue
idx = k.rfind('.')
to_replace = k[:idx]
if (to_replace in reverse_lookup):
key = (reverse_lookup[to_replace] + k[idx:])
new_state[key] = state_dict[k].to(device)
nn.Module.load_state_dict(partition, new_state, strict=strict)
|
def named_buffers(partition, prefix='', recurse=True):
params = nn.Module.named_buffers(partition, prefix=prefix, recurse=recurse)
lookup = partition.lookup
for (k, v) in params:
if (k in lookup):
(yield (lookup[k], v))
else:
assert ('.' in k)
split_idx = k.find('.')
new_k = (lookup[k[:split_idx]] + k[split_idx:])
(yield (new_k, v))
|
def named_parameters(partition, prefix='', recurse=True):
params = nn.Module.named_parameters(partition, prefix=prefix, recurse=recurse)
lookup = partition.lookup
for (k, v) in params:
if (k in lookup):
(yield (lookup[k], v))
else:
assert ('.' in k)
split_idx = k.find('.')
new_k = (lookup[k[:split_idx]] + k[split_idx:])
(yield (new_k, v))
|
def cpu(partition):
partition.device = torch.device('cpu')
return nn.Module.cpu(partition)
|
def cuda(partition, device=None):
if (device is None):
device = torch.cuda.current_device()
partition.device = torch.device(device)
return nn.Module.cuda(partition, partition.device)
|
def to(partition, *args, **kwargs):
device = None
if ('device' in kwargs):
device = kwargs['device']
elif ('tensor' in kwargs):
device = kwargs['tensor'].device
if args:
if isinstance(args[0], (torch.device, int, str)):
device = args[0]
if torch.is_tensor(args[0]):
device = args[0].device
if (not (device is None)):
partition.device = torch.device(device)
return nn.Module.to(partition, *args, **kwargs)
|
def create_pipeline_configuration(DEBUG=False, batch_size=32):
config = {'batch_dim': 0, 'depth': 10000, 'basic_blocks': (CrossEntropyLoss, Linear, Dropout, T5LayerNorm, T5Block, StatelessEmbedding), 'model_inputs': {'attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'is_batched': True, 'used_by': [1, 2, 3, 4, 5, 6, 7, 8]}, 'decoder_attention_mask': {'shape': torch.Size([32, 1, 4, 4]), 'dtype': torch.float32, 'is_batched': True, 'used_by': [9, 10, 11, 12, 13, 14, 15]}, 'decoder_input_ids': {'shape': torch.Size([32, 4]), 'dtype': torch.int64, 'is_batched': True, 'used_by': [9]}, 'input_ids': {'shape': torch.Size([32, 64]), 'dtype': torch.int64, 'is_batched': True, 'used_by': [0]}, 'inverted_encoder_attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'is_batched': True, 'used_by': [9, 10, 11, 12, 13, 14, 15]}, 'lm_labels': {'shape': torch.Size([32, 4]), 'dtype': torch.int64, 'is_batched': True, 'used_by': [15]}}, 'model_outputs': {'T5ForConditionalGeneration/CrossEntropyLoss[lm_loss]': {'shape': torch.Size([1]), 'dtype': torch.float32, 'is_batched': False, 'created_by': 15}}, 'stages': {0: {'stage_cls': Partition0, 'inputs': {'input_ids': {'shape': torch.Size([32, 64]), 'dtype': torch.int64, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}}, 'outputs': {'T5ForConditionalGeneration/Parameter[shared_embed_weight]': {'shape': torch.Size([32100, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': False, 'used_by': [9]}, 'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [1]}}, 'devices': [('cpu' if DEBUG else 'cuda:0')], 'stage_depth': 15}, 1: {'stage_cls': Partition1, 'inputs': {'attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 0}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_2': {'shape': torch.Size([32, 32, 64, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [2]}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[2]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [2]}}, 'devices': [('cpu' if DEBUG else 'cuda:1')], 'stage_depth': 14}, 2: {'stage_cls': Partition2, 'inputs': {'attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_2': {'shape': torch.Size([32, 32, 64, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 1}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[2]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 1}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_3': {'shape': torch.Size([32, 32, 64, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [3]}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[5]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [3]}}, 'devices': [('cpu' if DEBUG else 'cuda:2')], 'stage_depth': 13}, 3: {'stage_cls': Partition3, 'inputs': {'attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_3': {'shape': torch.Size([32, 32, 64, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 2}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[5]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 2}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_4': {'shape': torch.Size([32, 32, 64, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [4]}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[8]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [4]}}, 'devices': [('cpu' if DEBUG else 'cuda:3')], 'stage_depth': 12}, 4: {'stage_cls': Partition4, 'inputs': {'attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_4': {'shape': torch.Size([32, 32, 64, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 3}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[8]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 3}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_5': {'shape': torch.Size([32, 32, 64, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [5]}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[11]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [5]}}, 'devices': [('cpu' if DEBUG else 'cuda:4')], 'stage_depth': 11}, 5: {'stage_cls': Partition5, 'inputs': {'attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_5': {'shape': torch.Size([32, 32, 64, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 4}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[11]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 4}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_6': {'shape': torch.Size([32, 32, 64, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [6]}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[14]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [6]}}, 'devices': [('cpu' if DEBUG else 'cuda:5')], 'stage_depth': 10}, 6: {'stage_cls': Partition6, 'inputs': {'attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_6': {'shape': torch.Size([32, 32, 64, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 5}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[14]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 5}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_7': {'shape': torch.Size([32, 32, 64, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [7]}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[17]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [7]}}, 'devices': [('cpu' if DEBUG else 'cuda:6')], 'stage_depth': 9}, 7: {'stage_cls': Partition7, 'inputs': {'attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_7': {'shape': torch.Size([32, 32, 64, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 6}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[17]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 6}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_8': {'shape': torch.Size([32, 32, 64, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [8]}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[20]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [8]}}, 'devices': [('cpu' if DEBUG else 'cuda:7')], 'stage_depth': 8}, 8: {'stage_cls': Partition8, 'inputs': {'attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_8': {'shape': torch.Size([32, 32, 64, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 7}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[20]': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 7}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]_9': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [9]}}, 'devices': [('cpu' if DEBUG else 'cuda:8')], 'stage_depth': 7}, 9: {'stage_cls': Partition9, 'inputs': {'decoder_attention_mask': {'shape': torch.Size([32, 1, 4, 4]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'decoder_input_ids': {'shape': torch.Size([32, 4]), 'dtype': torch.int64, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'inverted_encoder_attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/Parameter[shared_embed_weight]': {'shape': torch.Size([32100, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': False, 'created_by': 0}, 'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]_9': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 8}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]_10': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [10]}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___130_10': {'shape': torch.Size([32, 32, 4, 4]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [10]}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___132_10': {'shape': torch.Size([32, 32, 4, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [10]}, 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[1]': {'shape': torch.Size([32, 4, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [10]}}, 'devices': [('cpu' if DEBUG else 'cuda:9')], 'stage_depth': 6}, 10: {'stage_cls': Partition10, 'inputs': {'decoder_attention_mask': {'shape': torch.Size([32, 1, 4, 4]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'inverted_encoder_attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]_10': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 9}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___130_10': {'shape': torch.Size([32, 32, 4, 4]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 9}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___132_10': {'shape': torch.Size([32, 32, 4, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 9}, 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[1]': {'shape': torch.Size([32, 4, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 9}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]_11': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [11]}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___130_11': {'shape': torch.Size([32, 32, 4, 4]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [11]}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___132_11': {'shape': torch.Size([32, 32, 4, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [11]}, 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[5]': {'shape': torch.Size([32, 4, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [11]}}, 'devices': [('cpu' if DEBUG else 'cuda:10')], 'stage_depth': 5}, 11: {'stage_cls': Partition11, 'inputs': {'decoder_attention_mask': {'shape': torch.Size([32, 1, 4, 4]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'inverted_encoder_attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]_11': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 10}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___130_11': {'shape': torch.Size([32, 32, 4, 4]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 10}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___132_11': {'shape': torch.Size([32, 32, 4, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 10}, 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[5]': {'shape': torch.Size([32, 4, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 10}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]_12': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [12]}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___130_12': {'shape': torch.Size([32, 32, 4, 4]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [12]}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___132_12': {'shape': torch.Size([32, 32, 4, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [12]}, 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[9]': {'shape': torch.Size([32, 4, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [12]}}, 'devices': [('cpu' if DEBUG else 'cuda:11')], 'stage_depth': 4}, 12: {'stage_cls': Partition12, 'inputs': {'decoder_attention_mask': {'shape': torch.Size([32, 1, 4, 4]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'inverted_encoder_attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]_12': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 11}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___130_12': {'shape': torch.Size([32, 32, 4, 4]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 11}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___132_12': {'shape': torch.Size([32, 32, 4, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 11}, 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[9]': {'shape': torch.Size([32, 4, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 11}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]_13': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [13]}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___130_13': {'shape': torch.Size([32, 32, 4, 4]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [13]}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___132_13': {'shape': torch.Size([32, 32, 4, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [13]}, 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[13]': {'shape': torch.Size([32, 4, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [13]}}, 'devices': [('cpu' if DEBUG else 'cuda:12')], 'stage_depth': 3}, 13: {'stage_cls': Partition13, 'inputs': {'decoder_attention_mask': {'shape': torch.Size([32, 1, 4, 4]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'inverted_encoder_attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]_13': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 12}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___130_13': {'shape': torch.Size([32, 32, 4, 4]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 12}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___132_13': {'shape': torch.Size([32, 32, 4, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 12}, 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[13]': {'shape': torch.Size([32, 4, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 12}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]_14': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [14]}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___130_14': {'shape': torch.Size([32, 32, 4, 4]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [14]}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___132_14': {'shape': torch.Size([32, 32, 4, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [14]}, 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[17]': {'shape': torch.Size([32, 4, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [14]}}, 'devices': [('cpu' if DEBUG else 'cuda:13')], 'stage_depth': 2}, 14: {'stage_cls': Partition14, 'inputs': {'decoder_attention_mask': {'shape': torch.Size([32, 1, 4, 4]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'inverted_encoder_attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]_14': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 13}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___130_14': {'shape': torch.Size([32, 32, 4, 4]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 13}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___132_14': {'shape': torch.Size([32, 32, 4, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 13}, 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[17]': {'shape': torch.Size([32, 4, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 13}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]_15': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [15]}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___130_15': {'shape': torch.Size([32, 32, 4, 4]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [15]}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___132_15': {'shape': torch.Size([32, 32, 4, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [15]}, 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[21]': {'shape': torch.Size([32, 4, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [15]}}, 'devices': [('cpu' if DEBUG else 'cuda:14')], 'stage_depth': 1}, 15: {'stage_cls': Partition15, 'inputs': {'decoder_attention_mask': {'shape': torch.Size([32, 1, 4, 4]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'inverted_encoder_attention_mask': {'shape': torch.Size([32, 1, 1, 64]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'lm_labels': {'shape': torch.Size([32, 4]), 'dtype': torch.int64, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]_15': {'shape': torch.Size([32, 64, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 14}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___130_15': {'shape': torch.Size([32, 32, 4, 4]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 14}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___132_15': {'shape': torch.Size([32, 32, 4, 64]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 14}, 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[21]': {'shape': torch.Size([32, 4, 1024]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 14}}, 'outputs': {'T5ForConditionalGeneration/CrossEntropyLoss[lm_loss]': {'shape': torch.Size([1]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': False, 'used_by': [(- 1)]}}, 'devices': [('cpu' if DEBUG else 'cuda:15')], 'stage_depth': 0}}}
batch_dim = config['batch_dim']
for d in chain(config['model_inputs'].values(), config['model_outputs'].values()):
if d['is_batched']:
shape = d['shape']
d['shape'] = torch.Size(((shape[:batch_dim] + (batch_size,)) + shape[(batch_dim + 1):]))
for s in config['stages'].values():
for d in chain(s['inputs'].values(), s['outputs'].values()):
if d['is_batched']:
shape = d['shape']
d['shape'] = torch.Size(((shape[:batch_dim] + (batch_size,)) + shape[(batch_dim + 1):]))
return config
|
class Partition0(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[encoder]/StatelessEmbedding[embed_tokens]', 'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]']
TENSORS = ['T5ForConditionalGeneration/Parameter[shared_embed_weight]']
def __init__(self, layers, tensors, device='cuda:0'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1]
self.lookup = {'l_0': 'encoder.embed_tokens', 'l_1': 'encoder.dropout', 'p_0': 'shared_embed_weight'}
self.to(self.device)
def forward(self, *args):
input_ids = unflatten(args, self.input_structure)[0]
t_0 = input_ids.size()
t_0 = t_0[(- 1)]
t_0 = input_ids.view((- 1), t_0)
t_0 = self.l_0(self.p_0, t_0)
t_0 = self.l_1(t_0)
return list(flatten((self.p_0, t_0)))
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition1(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[encoder]/T5Block[0]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[1]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[2]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:1'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1]
self.lookup = {'l_0': 'encoder.0', 'l_1': 'encoder.1', 'l_2': 'encoder.2'}
self.to(self.device)
def forward(self, *args):
(attention_mask, x0) = unflatten(args, self.input_structure)
t_0 = self.l_0(x0, attention_mask=attention_mask, position_bias=None, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_1 = t_0[0]
t_0 = t_0[1]
t_1 = self.l_1(t_1, attention_mask=attention_mask, position_bias=t_0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_1 = self.l_2(t_1, attention_mask=attention_mask, position_bias=t_0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
return list(flatten((t_0, t_1)))
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition2(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[encoder]/T5Block[3]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[4]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[5]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:2'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1]
self.lookup = {'l_0': 'encoder.3', 'l_1': 'encoder.4', 'l_2': 'encoder.5'}
self.to(self.device)
def forward(self, *args):
(attention_mask, x0, x1) = unflatten(args, self.input_structure)
t_0 = self.l_0(x1, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_1(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_2(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
return list(flatten((x0, t_0)))
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition3(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[encoder]/T5Block[6]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[7]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[8]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:3'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1]
self.lookup = {'l_0': 'encoder.6', 'l_1': 'encoder.7', 'l_2': 'encoder.8'}
self.to(self.device)
def forward(self, *args):
(attention_mask, x0, x1) = unflatten(args, self.input_structure)
t_0 = self.l_0(x1, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_1(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_2(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
return list(flatten((x0, t_0)))
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition4(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[encoder]/T5Block[9]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[10]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[11]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:4'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1]
self.lookup = {'l_0': 'encoder.9', 'l_1': 'encoder.10', 'l_2': 'encoder.11'}
self.to(self.device)
def forward(self, *args):
(attention_mask, x0, x1) = unflatten(args, self.input_structure)
t_0 = self.l_0(x1, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_1(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_2(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
return list(flatten((x0, t_0)))
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition5(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[encoder]/T5Block[12]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[13]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[14]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:5'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1]
self.lookup = {'l_0': 'encoder.12', 'l_1': 'encoder.13', 'l_2': 'encoder.14'}
self.to(self.device)
def forward(self, *args):
(attention_mask, x0, x1) = unflatten(args, self.input_structure)
t_0 = self.l_0(x1, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_1(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_2(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
return list(flatten((x0, t_0)))
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition6(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[encoder]/T5Block[15]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[16]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[17]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:6'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1]
self.lookup = {'l_0': 'encoder.15', 'l_1': 'encoder.16', 'l_2': 'encoder.17'}
self.to(self.device)
def forward(self, *args):
(attention_mask, x0, x1) = unflatten(args, self.input_structure)
t_0 = self.l_0(x1, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_1(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_2(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
return list(flatten((x0, t_0)))
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition7(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[encoder]/T5Block[18]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[19]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[20]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:7'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1]
self.lookup = {'l_0': 'encoder.18', 'l_1': 'encoder.19', 'l_2': 'encoder.20'}
self.to(self.device)
def forward(self, *args):
(attention_mask, x0, x1) = unflatten(args, self.input_structure)
t_0 = self.l_0(x1, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_1(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_2(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
return list(flatten((x0, t_0)))
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition8(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[encoder]/T5Block[21]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[22]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[23]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5LayerNorm[final_layer_norm]', 'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:8'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1]
self.lookup = {'l_0': 'encoder.21', 'l_1': 'encoder.22', 'l_2': 'encoder.23', 'l_3': 'encoder.final_layer_norm', 'l_4': 'encoder.dropout'}
self.to(self.device)
def forward(self, *args):
(attention_mask, x0, x1) = unflatten(args, self.input_structure)
t_0 = self.l_0(x1, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_1(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_2(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_3(t_0)
t_0 = self.l_4(t_0)
return (t_0,)
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition9(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[decoder]/StatelessEmbedding[embed_tokens]', 'T5ForConditionalGeneration/T5Stack[decoder]/Dropout[dropout]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[0]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[1]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:9'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1, 1, 1]
self.lookup = {'l_0': 'decoder.embed_tokens', 'l_1': 'decoder.dropout', 'l_2': 'decoder.0', 'l_3': 'decoder.1'}
self.to(self.device)
def forward(self, *args):
(decoder_attention_mask, decoder_input_ids, inverted_encoder_attention_mask, x0, x1) = unflatten(args, self.input_structure)
t_0 = decoder_input_ids.size()
t_0 = t_0[(- 1)]
t_0 = decoder_input_ids.view((- 1), t_0)
t_0 = self.l_0(x0, t_0)
t_0 = self.l_1(t_0)
t_0 = self.l_2(t_0, attention_mask=decoder_attention_mask, position_bias=None, encoder_hidden_states=x1, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=None)
t_1 = t_0[0]
t_2 = t_0[1]
t_0 = t_0[2]
t_1 = self.l_3(t_1, attention_mask=decoder_attention_mask, position_bias=t_2, encoder_hidden_states=x1, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=t_0)
return list(flatten((x1, t_2, t_0, t_1)))
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition10(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[decoder]/T5Block[2]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[3]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[4]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[5]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:10'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1, 1, 1, 1]
self.lookup = {'l_0': 'decoder.2', 'l_1': 'decoder.3', 'l_2': 'decoder.4', 'l_3': 'decoder.5'}
self.to(self.device)
def forward(self, *args):
(decoder_attention_mask, inverted_encoder_attention_mask, x0, x1, x2, x3) = unflatten(args, self.input_structure)
t_0 = self.l_0(x3, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_1(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_2(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_3(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
return list(flatten((x0, x1, x2, t_0)))
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition11(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[decoder]/T5Block[6]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[7]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[8]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[9]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:11'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1, 1, 1, 1]
self.lookup = {'l_0': 'decoder.6', 'l_1': 'decoder.7', 'l_2': 'decoder.8', 'l_3': 'decoder.9'}
self.to(self.device)
def forward(self, *args):
(decoder_attention_mask, inverted_encoder_attention_mask, x0, x1, x2, x3) = unflatten(args, self.input_structure)
t_0 = self.l_0(x3, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_1(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_2(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_3(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
return list(flatten((x0, x1, x2, t_0)))
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition12(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[decoder]/T5Block[10]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[11]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[12]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[13]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:12'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1, 1, 1, 1]
self.lookup = {'l_0': 'decoder.10', 'l_1': 'decoder.11', 'l_2': 'decoder.12', 'l_3': 'decoder.13'}
self.to(self.device)
def forward(self, *args):
(decoder_attention_mask, inverted_encoder_attention_mask, x0, x1, x2, x3) = unflatten(args, self.input_structure)
t_0 = self.l_0(x3, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_1(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_2(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_3(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
return list(flatten((x0, x1, x2, t_0)))
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition13(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[decoder]/T5Block[14]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[15]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[16]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[17]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:13'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1, 1, 1, 1]
self.lookup = {'l_0': 'decoder.14', 'l_1': 'decoder.15', 'l_2': 'decoder.16', 'l_3': 'decoder.17'}
self.to(self.device)
def forward(self, *args):
(decoder_attention_mask, inverted_encoder_attention_mask, x0, x1, x2, x3) = unflatten(args, self.input_structure)
t_0 = self.l_0(x3, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_1(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_2(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_3(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
return list(flatten((x0, x1, x2, t_0)))
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition14(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[decoder]/T5Block[18]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[19]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[20]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[21]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:14'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1, 1, 1, 1]
self.lookup = {'l_0': 'decoder.18', 'l_1': 'decoder.19', 'l_2': 'decoder.20', 'l_3': 'decoder.21'}
self.to(self.device)
def forward(self, *args):
(decoder_attention_mask, inverted_encoder_attention_mask, x0, x1, x2, x3) = unflatten(args, self.input_structure)
t_0 = self.l_0(x3, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_1(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_2(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_3(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
return list(flatten((x0, x1, x2, t_0)))
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition15(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[decoder]/T5Block[22]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[23]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5LayerNorm[final_layer_norm]', 'T5ForConditionalGeneration/T5Stack[decoder]/Dropout[dropout]', 'T5ForConditionalGeneration/Linear[lm_head]', 'T5ForConditionalGeneration/CrossEntropyLoss[lm_loss]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:15'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1, 1, 1, 1, 1]
self.lookup = {'l_0': 'decoder.22', 'l_1': 'decoder.23', 'l_2': 'decoder.final_layer_norm', 'l_3': 'decoder.dropout', 'l_4': 'lm_head', 'l_5': 'lm_loss'}
self.to(self.device)
def forward(self, *args):
(decoder_attention_mask, inverted_encoder_attention_mask, lm_labels, x0, x1, x2, x3) = unflatten(args, self.input_structure)
t_0 = self.l_0(x3, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_1(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_2(t_0)
t_0 = self.l_3(t_0)
t_0 = (t_0 * 0.03125)
t_0 = self.l_4(t_0)
t_1 = t_0.size((- 1))
t_1 = t_0.view((- 1), t_1)
t_0 = lm_labels.view((- 1))
t_0 = self.l_5(t_1, t_0)
return (t_0,)
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
def traverse_model(module: nn.Module, depth: int, prefix: Optional[str]=None, basic_blocks: Tuple[nn.Module]=(), full: bool=False) -> Iterator[Tuple[(nn.Module, str, nn.Module)]]:
'\n iterate over model layers yielding the layer,layer_scope,encasing_module\n Parameters:\n -----------\n model:\n the model to iterate over\n depth:\n how far down in the model tree to go\n basic_blocks:\n a list of modules that if encountered will not be broken down\n full:\n whether to yield only layers specified by the depth and basick_block options or to yield all layers\n '
if (prefix is None):
prefix = type(module).__name__
for (name, sub_module) in module.named_children():
scope = (((prefix + '/') + type(sub_module).__name__) + f'[{name}]')
if ((len(list(sub_module.children())) == 0) or isinstance(sub_module, tuple(basic_blocks)) or (depth == 0)):
if full:
(yield (sub_module, scope, module, True))
else:
(yield (sub_module, scope, module))
else:
if full:
(yield (sub_module, scope, module, False))
(yield from traverse_model(sub_module, (depth - 1), scope, basic_blocks, full))
|
def layerDict(model: nn.Module, depth=1000, basic_blocks=()) -> Dict[(str, nn.Module)]:
return {s: l for (l, s, _) in traverse_model(model, depth, basic_blocks=basic_blocks)}
|
def traverse_params_buffs(module: nn.Module, prefix: Optional[str]=None) -> Iterator[Tuple[(torch.tensor, str)]]:
"\n iterate over model's buffers and parameters yielding obj,obj_scope\n\n Parameters:\n -----------\n model:\n the model to iterate over\n "
if (prefix is None):
prefix = type(module).__name__
for (param_name, param) in module.named_parameters(recurse=False):
param_scope = f'{prefix}/{type(param).__name__}[{param_name}]'
(yield (param, param_scope))
for (buffer_name, buffer) in module.named_buffers(recurse=False):
buffer_scope = f'{prefix}/{type(buffer).__name__}[{buffer_name}]'
(yield (buffer, buffer_scope))
for (name, sub_module) in module.named_children():
(yield from traverse_params_buffs(sub_module, (((prefix + '/') + type(sub_module).__name__) + f'[{name}]')))
|
def tensorDict(model: nn.Module) -> OrderedDict[(str, Tensor)]:
return collections.OrderedDict(((s, t) for (t, s) in traverse_params_buffs(model)))
|
def move_tensors(ts, device):
def move(t):
if isinstance(t, (nn.Module, Tensor)):
return t.to(device)
return t
return nested_map(move, ts)
|
def nested_map(func, ts, full=False):
if isinstance(ts, torch.Size):
return func(ts)
elif isinstance(ts, (list, tuple, set)):
return type(ts)((nested_map(func, t, full=full) for t in ts))
elif isinstance(ts, dict):
return {k: nested_map(func, v, full=full) for (k, v) in ts.items()}
elif (isinstance(ts, slice) and full):
start = nested_map(func, ts.start, full=full)
stop = nested_map(func, ts.stop, full=full)
step = nested_map(func, ts.step, full=full)
return slice(start, stop, step)
return func(ts)
|
def flatten(ts):
if isinstance(ts, torch.Size):
(yield ts)
elif isinstance(ts, (list, tuple, set)):
(yield from chain(*[flatten(t) for t in ts]))
elif isinstance(ts, dict):
(yield from chain(*[flatten(t) for (k, t) in sorted(ts.items(), key=(lambda t: t[0]))]))
else:
(yield ts)
|
def unflatten(xs, structure):
return _unflatten(xs, structure)[0]
|
def _unflatten(xs, structure):
if isinstance(structure, torch.Size):
return (xs[0], 1)
if (not isinstance(structure, (list, tuple, set, dict))):
return (xs[0], 1)
if isinstance(structure, (list, tuple, set)):
offset = 0
elements = []
for s in structure:
(e, n) = _unflatten(xs[offset:], s)
elements.append(e)
offset += n
return (type(structure)(elements), offset)
assert isinstance(structure, dict)
offset = 0
elements = dict()
for (k, v) in sorted(structure.items(), key=(lambda t: t[0])):
(e, n) = _unflatten(xs[offset:], v)
elements[k] = e
offset += n
return (elements, offset)
|
def state_dict(partition, *args, **kwargs):
state = nn.Module.state_dict(partition, *args, **kwargs)
lookup = partition.lookup
result = dict()
for (k, v) in state.items():
if (k in lookup):
result[lookup[k]] = v
else:
assert ('.' in k)
split_idx = k.find('.')
new_k = (lookup[k[:split_idx]] + k[split_idx:])
result[new_k] = v
return result
|
def load_state_dict(partition, state):
reverse_lookup = {v: k for (k, v) in partition.lookup.items()}
device = partition.device
keys = list(partition.state_dict(None).keys())
new_state = dict()
for k in keys:
if (k in reverse_lookup):
new_state[reverse_lookup[k]] = state[k].to(device)
continue
idx = k.rfind('.')
to_replace = k[:idx]
if (to_replace in reverse_lookup):
key = (reverse_lookup[to_replace] + k[idx:])
new_state[key] = state[k].to(device)
nn.Module.load_state_dict(partition, new_state, strict=True)
|
def named_buffers(partition, recurse=True):
params = nn.Module.named_buffers(partition, recurse=recurse)
lookup = partition.lookup
for (k, v) in params:
if (k in lookup):
(yield (lookup[k], v))
else:
assert ('.' in k)
split_idx = k.find('.')
new_k = (lookup[k[:split_idx]] + k[split_idx:])
(yield (new_k, v))
|
def named_parameters(partition, recurse=True):
params = nn.Module.named_parameters(partition, recurse=recurse)
lookup = partition.lookup
for (k, v) in params:
if (k in lookup):
(yield (lookup[k], v))
else:
assert ('.' in k)
split_idx = k.find('.')
new_k = (lookup[k[:split_idx]] + k[split_idx:])
(yield (new_k, v))
|
def cpu(partition):
partition.device = torch.device('cpu')
return nn.Module.cpu(partition)
|
def cuda(partition, device=None):
if (device is None):
device = torch.cuda.current_device()
partition.device = torch.device(device)
return nn.Module.cuda(partition, partition.device)
|
def to(partition, *args, **kwargs):
device = None
if ('device' in kwargs):
device = kwargs['device']
elif ('tensor' in kwargs):
device = kwargs['tensor'].device
if args:
if isinstance(args[0], (torch.device, int, str)):
device = args[0]
if torch.is_tensor(args[0]):
device = args[0].device
if (not (device is None)):
partition.device = torch.device(device)
return nn.Module.to(partition, *args, **kwargs)
|
def create_pipeline_configuration(DEBUG=False, batch_size=32):
config = {'batch_dim': 0, 'depth': 10000, 'basic_blocks': (CrossEntropyLoss, T5Block, T5LayerNorm, StatelessEmbedding, Linear, Dropout), 'model_inputs': {'attention_mask': {'shape': torch.Size([32, 1, 1, 384]), 'dtype': torch.float32, 'is_batched': True, 'used_by': [0, 1, 2]}, 'decoder_attention_mask': {'shape': torch.Size([32, 1, 32, 32]), 'dtype': torch.float32, 'is_batched': True, 'used_by': [3]}, 'decoder_input_ids': {'shape': torch.Size([32, 32]), 'dtype': torch.int64, 'is_batched': True, 'used_by': [0]}, 'input_ids': {'shape': torch.Size([32, 384]), 'dtype': torch.int64, 'is_batched': True, 'used_by': [0]}, 'inverted_encoder_attention_mask': {'shape': torch.Size([32, 1, 1, 384]), 'dtype': torch.float32, 'is_batched': True, 'used_by': [3]}, 'lm_labels': {'shape': torch.Size([32, 32]), 'dtype': torch.int64, 'is_batched': True, 'used_by': [3]}}, 'model_outputs': {'T5ForConditionalGeneration/CrossEntropyLoss[lm_loss]': {'shape': torch.Size([1]), 'dtype': torch.float32, 'is_batched': False, 'created_by': 3}}, 'stages': {0: {'stage_cls': Partition0, 'inputs': {'attention_mask': {'shape': torch.Size([32, 1, 1, 384]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'decoder_input_ids': {'shape': torch.Size([32, 32]), 'dtype': torch.int64, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'input_ids': {'shape': torch.Size([32, 384]), 'dtype': torch.int64, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22': {'shape': torch.Size([32, 8, 384, 384]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [1, 2]}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[1]': {'shape': torch.Size([32, 384, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [1]}, 'T5ForConditionalGeneration/T5Stack[decoder]/StatelessEmbedding[embed_tokens]': {'shape': torch.Size([32, 32, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [3]}}, 'devices': [('cpu' if DEBUG else 'cuda:0')]}, 1: {'stage_cls': Partition1, 'inputs': {'attention_mask': {'shape': torch.Size([32, 1, 1, 384]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22': {'shape': torch.Size([32, 8, 384, 384]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 0}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[1]': {'shape': torch.Size([32, 384, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 0}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[3]': {'shape': torch.Size([32, 384, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [2]}}, 'devices': [('cpu' if DEBUG else 'cuda:1')]}, 2: {'stage_cls': Partition2, 'inputs': {'attention_mask': {'shape': torch.Size([32, 1, 1, 384]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22': {'shape': torch.Size([32, 8, 384, 384]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 0}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[3]': {'shape': torch.Size([32, 384, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 1}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]': {'shape': torch.Size([32, 384, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [3]}}, 'devices': [('cpu' if DEBUG else 'cuda:2')]}, 3: {'stage_cls': Partition3, 'inputs': {'decoder_attention_mask': {'shape': torch.Size([32, 1, 32, 32]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'inverted_encoder_attention_mask': {'shape': torch.Size([32, 1, 1, 384]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'lm_labels': {'shape': torch.Size([32, 32]), 'dtype': torch.int64, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]': {'shape': torch.Size([32, 384, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 2}, 'T5ForConditionalGeneration/T5Stack[decoder]/StatelessEmbedding[embed_tokens]': {'shape': torch.Size([32, 32, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 0}}, 'outputs': {'T5ForConditionalGeneration/CrossEntropyLoss[lm_loss]': {'shape': torch.Size([1]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': False, 'used_by': [(- 1)]}}, 'devices': [('cpu' if DEBUG else 'cuda:3')]}}}
batch_dim = config['batch_dim']
for d in chain(config['model_inputs'].values(), config['model_outputs'].values()):
if d['is_batched']:
shape = d['shape']
d['shape'] = torch.Size(((shape[:batch_dim] + (batch_size,)) + shape[(batch_dim + 1):]))
for s in config['stages'].values():
for d in chain(s['inputs'].values(), s['outputs'].values()):
if d['is_batched']:
shape = d['shape']
d['shape'] = torch.Size(((shape[:batch_dim] + (batch_size,)) + shape[(batch_dim + 1):]))
return config
|
class Partition0(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[encoder]/StatelessEmbedding[embed_tokens]', 'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[0]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[1]', 'T5ForConditionalGeneration/T5Stack[decoder]/StatelessEmbedding[embed_tokens]']
TENSORS = ['T5ForConditionalGeneration/Parameter[shared_embed_weight]']
def __init__(self, layers, tensors, device='cuda:0'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1]
self.lookup = {'l_0': 'encoder.embed_tokens', 'l_1': 'encoder.dropout', 'l_2': 'encoder.0', 'l_3': 'encoder.1', 'l_4': 'decoder.embed_tokens', 'p_0': 'shared_embed_weight'}
self.to(self.device)
def forward(self, *args):
(attention_mask, decoder_input_ids, input_ids) = move_tensors(unflatten(args, self.input_structure), self.device)
t_0 = input_ids.size()
t_0 = t_0[(- 1)]
t_0 = input_ids.view((- 1), t_0)
t_0 = self.l_0(self.p_0, t_0)
t_0 = self.l_1(t_0)
t_0 = self.l_2(t_0, attention_mask=attention_mask, position_bias=None, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_1 = t_0[0]
t_0 = t_0[1]
t_1 = self.l_3(t_1, attention_mask=attention_mask, position_bias=t_0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_2 = decoder_input_ids.size()
t_2 = t_2[(- 1)]
t_2 = decoder_input_ids.view((- 1), t_2)
t_2 = self.l_4(self.p_0, t_2)
return list(flatten((t_0, t_1, t_2)))
def state_dict(self, device=None):
return state_dict(self, device=device)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition1(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[encoder]/T5Block[2]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[3]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:1'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1]
self.lookup = {'l_0': 'encoder.2', 'l_1': 'encoder.3'}
self.to(self.device)
def forward(self, *args):
(attention_mask, x0, x1) = move_tensors(unflatten(args, self.input_structure), self.device)
t_0 = self.l_0(x1, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_1(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
return (t_0,)
def state_dict(self, device=None):
return state_dict(self, device=device)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition2(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[encoder]/T5Block[4]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[5]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5LayerNorm[final_layer_norm]', 'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:2'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1]
self.lookup = {'l_0': 'encoder.4', 'l_1': 'encoder.5', 'l_2': 'encoder.final_layer_norm', 'l_3': 'encoder.dropout'}
self.to(self.device)
def forward(self, *args):
(attention_mask, x0, x1) = move_tensors(unflatten(args, self.input_structure), self.device)
t_0 = self.l_0(x1, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_1(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_2(t_0)
t_0 = self.l_3(t_0)
return (t_0,)
def state_dict(self, device=None):
return state_dict(self, device=device)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition3(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[decoder]/Dropout[dropout]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[0]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[1]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[2]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[3]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[4]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[5]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5LayerNorm[final_layer_norm]', 'T5ForConditionalGeneration/T5Stack[decoder]/Dropout[dropout]', 'T5ForConditionalGeneration/Linear[lm_head]', 'T5ForConditionalGeneration/CrossEntropyLoss[lm_loss]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:3'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1, 1, 1]
self.lookup = {'l_0': 'decoder.dropout', 'l_1': 'decoder.0', 'l_2': 'decoder.1', 'l_3': 'decoder.2', 'l_4': 'decoder.3', 'l_5': 'decoder.4', 'l_6': 'decoder.5', 'l_7': 'decoder.final_layer_norm', 'l_8': 'decoder.dropout', 'l_9': 'lm_head', 'l_10': 'lm_loss'}
self.to(self.device)
def forward(self, *args):
(decoder_attention_mask, inverted_encoder_attention_mask, lm_labels, x0, x1) = move_tensors(unflatten(args, self.input_structure), self.device)
t_0 = self.l_0(x1)
t_0 = self.l_1(t_0, attention_mask=decoder_attention_mask, position_bias=None, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=None)
t_1 = t_0[0]
t_2 = t_0[1]
t_0 = t_0[2]
t_1 = self.l_2(t_1, attention_mask=decoder_attention_mask, position_bias=t_2, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=t_0)
t_1 = self.l_3(t_1, attention_mask=decoder_attention_mask, position_bias=t_2, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=t_0)
t_1 = self.l_4(t_1, attention_mask=decoder_attention_mask, position_bias=t_2, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=t_0)
t_1 = self.l_5(t_1, attention_mask=decoder_attention_mask, position_bias=t_2, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=t_0)
t_0 = self.l_6(t_1, attention_mask=decoder_attention_mask, position_bias=t_2, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=t_0)
t_0 = self.l_7(t_0)
t_0 = self.l_8(t_0)
t_0 = (t_0 * 0.04419417382415922)
t_0 = self.l_9(t_0)
t_2 = t_0.size((- 1))
t_2 = t_0.view((- 1), t_2)
t_0 = lm_labels.view((- 1))
t_0 = self.l_10(t_2, t_0)
return (t_0,)
def state_dict(self, device=None):
return state_dict(self, device=device)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
def traverse_model(module: nn.Module, depth: int, prefix: Optional[str]=None, basic_blocks: Tuple[nn.Module]=(), full: bool=False) -> Iterator[Tuple[(nn.Module, str, nn.Module)]]:
'\n iterate over model layers yielding the layer,layer_scope,encasing_module\n Parameters:\n -----------\n model:\n the model to iterate over\n depth:\n how far down in the model tree to go\n basic_blocks:\n a list of modules that if encountered will not be broken down\n full:\n whether to yield only layers specified by the depth and basick_block options or to yield all layers\n '
if (prefix is None):
prefix = type(module).__name__
for (name, sub_module) in module.named_children():
scope = (((prefix + '/') + type(sub_module).__name__) + f'[{name}]')
if ((len(list(sub_module.children())) == 0) or isinstance(sub_module, tuple(basic_blocks)) or (depth == 0)):
if full:
(yield (sub_module, scope, module, True))
else:
(yield (sub_module, scope, module))
else:
if full:
(yield (sub_module, scope, module, False))
(yield from traverse_model(sub_module, (depth - 1), scope, basic_blocks, full))
|
def layerDict(model: nn.Module, depth=1000, basic_blocks=()) -> Dict[(str, nn.Module)]:
return {s: l for (l, s, _) in traverse_model(model, depth, basic_blocks=basic_blocks)}
|
def traverse_params_buffs(module: nn.Module, prefix: Optional[str]=None) -> Iterator[Tuple[(torch.tensor, str)]]:
"\n iterate over model's buffers and parameters yielding obj,obj_scope\n\n Parameters:\n -----------\n model:\n the model to iterate over\n "
if (prefix is None):
prefix = type(module).__name__
for (param_name, param) in module.named_parameters(recurse=False):
param_scope = f'{prefix}/{type(param).__name__}[{param_name}]'
(yield (param, param_scope))
for (buffer_name, buffer) in module.named_buffers(recurse=False):
buffer_scope = f'{prefix}/{type(buffer).__name__}[{buffer_name}]'
(yield (buffer, buffer_scope))
for (name, sub_module) in module.named_children():
(yield from traverse_params_buffs(sub_module, (((prefix + '/') + type(sub_module).__name__) + f'[{name}]')))
|
def tensorDict(model: nn.Module) -> OrderedDict[(str, Tensor)]:
return collections.OrderedDict(((s, t) for (t, s) in traverse_params_buffs(model)))
|
def move_tensors(ts, device):
def move(t):
if isinstance(t, (nn.Module, Tensor)):
return t.to(device)
return t
return nested_map(move, ts)
|
def nested_map(func, ts, full=False):
if isinstance(ts, torch.Size):
return func(ts)
elif isinstance(ts, (list, tuple, set)):
return type(ts)((nested_map(func, t, full=full) for t in ts))
elif isinstance(ts, dict):
return {k: nested_map(func, v, full=full) for (k, v) in ts.items()}
elif (isinstance(ts, slice) and full):
start = nested_map(func, ts.start, full=full)
stop = nested_map(func, ts.stop, full=full)
step = nested_map(func, ts.step, full=full)
return slice(start, stop, step)
return func(ts)
|
def flatten(ts):
if isinstance(ts, torch.Size):
(yield ts)
elif isinstance(ts, (list, tuple, set)):
(yield from chain(*[flatten(t) for t in ts]))
elif isinstance(ts, dict):
(yield from chain(*[flatten(t) for (k, t) in sorted(ts.items(), key=(lambda t: t[0]))]))
else:
(yield ts)
|
def unflatten(xs, structure):
return _unflatten(xs, structure)[0]
|
def _unflatten(xs, structure):
if isinstance(structure, torch.Size):
return (xs[0], 1)
if (not isinstance(structure, (list, tuple, set, dict))):
return (xs[0], 1)
if isinstance(structure, (list, tuple, set)):
offset = 0
elements = []
for s in structure:
(e, n) = _unflatten(xs[offset:], s)
elements.append(e)
offset += n
return (type(structure)(elements), offset)
assert isinstance(structure, dict)
offset = 0
elements = dict()
for (k, v) in sorted(structure.items(), key=(lambda t: t[0])):
(e, n) = _unflatten(xs[offset:], v)
elements[k] = e
offset += n
return (elements, offset)
|
def state_dict(partition, device=None):
state = nn.Module.state_dict(partition)
lookup = partition.lookup
result = dict()
for (k, v) in state.items():
if (k in lookup):
result[lookup[k]] = (v if (device is None) else v.to(device))
else:
assert ('.' in k)
split_idx = k.find('.')
new_k = (lookup[k[:split_idx]] + k[split_idx:])
result[new_k] = (v if (device is None) else v.to(device))
return result
|
def load_state_dict(partition, state):
reverse_lookup = {v: k for (k, v) in partition.lookup.items()}
device = partition.device
keys = list(partition.state_dict(None).keys())
new_state = dict()
for k in keys:
if (k in reverse_lookup):
new_state[reverse_lookup[k]] = state[k].to(device)
continue
idx = k.rfind('.')
to_replace = k[:idx]
if (to_replace in reverse_lookup):
key = (reverse_lookup[to_replace] + k[idx:])
new_state[key] = state[k].to(device)
nn.Module.load_state_dict(partition, new_state, strict=True)
|
def named_buffers(partition, recurse=True):
params = nn.Module.named_buffers(partition, recurse=recurse)
lookup = partition.lookup
for (k, v) in params:
if (k in lookup):
(yield (lookup[k], v))
else:
assert ('.' in k)
split_idx = k.find('.')
new_k = (lookup[k[:split_idx]] + k[split_idx:])
(yield (new_k, v))
|
def named_parameters(partition, recurse=True):
params = nn.Module.named_parameters(partition, recurse=recurse)
lookup = partition.lookup
for (k, v) in params:
if (k in lookup):
(yield (lookup[k], v))
else:
assert ('.' in k)
split_idx = k.find('.')
new_k = (lookup[k[:split_idx]] + k[split_idx:])
(yield (new_k, v))
|
def cpu(partition):
partition.device = torch.device('cpu')
return nn.Module.cpu(partition)
|
def cuda(partition, device=None):
if (device is None):
device = torch.cuda.current_device()
partition.device = torch.device(device)
return nn.Module.cuda(partition, partition.device)
|
def to(partition, *args, **kwargs):
device = None
if ('device' in kwargs):
device = kwargs['device']
elif ('tensor' in kwargs):
device = kwargs['tensor'].device
if args:
if isinstance(args[0], (torch.device, int, str)):
device = args[0]
if torch.is_tensor(args[0]):
device = args[0].device
if (not (device is None)):
partition.device = torch.device(device)
return nn.Module.to(partition, *args, **kwargs)
|
def create_pipeline_configuration(DEBUG=False, batch_size=32):
config = {'batch_dim': 0, 'depth': 10000, 'basic_blocks': (CrossEntropyLoss, Linear, StatelessEmbedding, T5Block, Dropout, T5LayerNorm), 'model_inputs': {'attention_mask': {'shape': torch.Size([32, 1, 1, 512]), 'dtype': torch.float32, 'is_batched': True, 'used_by': [0, 1, 2]}, 'decoder_attention_mask': {'shape': torch.Size([32, 1, 4, 4]), 'dtype': torch.float32, 'is_batched': True, 'used_by': [4, 5]}, 'decoder_input_ids': {'shape': torch.Size([32, 4]), 'dtype': torch.int64, 'is_batched': True, 'used_by': [0]}, 'input_ids': {'shape': torch.Size([32, 512]), 'dtype': torch.int64, 'is_batched': True, 'used_by': [0]}, 'inverted_encoder_attention_mask': {'shape': torch.Size([32, 1, 1, 512]), 'dtype': torch.float32, 'is_batched': True, 'used_by': [4, 5]}, 'lm_labels': {'shape': torch.Size([32, 4]), 'dtype': torch.int64, 'is_batched': True, 'used_by': [5]}}, 'model_outputs': {'T5ForConditionalGeneration/CrossEntropyLoss[lm_loss]': {'shape': torch.Size([1]), 'dtype': torch.float32, 'is_batched': False, 'created_by': 5}}, 'stages': {0: {'stage_cls': Partition0, 'inputs': {'attention_mask': {'shape': torch.Size([32, 1, 1, 512]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'decoder_input_ids': {'shape': torch.Size([32, 4]), 'dtype': torch.int64, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'input_ids': {'shape': torch.Size([32, 512]), 'dtype': torch.int64, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_1': {'shape': torch.Size([32, 8, 512, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [1]}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[1]': {'shape': torch.Size([32, 512, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [1]}, 'T5ForConditionalGeneration/T5Stack[decoder]/StatelessEmbedding[embed_tokens]': {'shape': torch.Size([32, 4, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [4]}}, 'devices': [('cpu' if DEBUG else 'cuda:0')], 'stage_depth': 5}, 1: {'stage_cls': Partition1, 'inputs': {'attention_mask': {'shape': torch.Size([32, 1, 1, 512]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_1': {'shape': torch.Size([32, 8, 512, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 0}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[1]': {'shape': torch.Size([32, 512, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 0}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_2': {'shape': torch.Size([32, 8, 512, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [2]}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[3]': {'shape': torch.Size([32, 512, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [2]}}, 'devices': [('cpu' if DEBUG else 'cuda:1')], 'stage_depth': 4}, 2: {'stage_cls': Partition2, 'inputs': {'attention_mask': {'shape': torch.Size([32, 1, 1, 512]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/tuple::__getitem___22_2': {'shape': torch.Size([32, 8, 512, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 1}, 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[3]': {'shape': torch.Size([32, 512, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 1}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[5]': {'shape': torch.Size([32, 512, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [3]}}, 'devices': [('cpu' if DEBUG else 'cuda:2')], 'stage_depth': 3}, 3: {'stage_cls': Partition3, 'inputs': {'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[5]': {'shape': torch.Size([32, 512, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 2}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]_4': {'shape': torch.Size([32, 512, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [4]}}, 'devices': [('cpu' if DEBUG else 'cuda:3')], 'stage_depth': 2}, 4: {'stage_cls': Partition4, 'inputs': {'decoder_attention_mask': {'shape': torch.Size([32, 1, 4, 4]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'inverted_encoder_attention_mask': {'shape': torch.Size([32, 1, 1, 512]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]_4': {'shape': torch.Size([32, 512, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 3}, 'T5ForConditionalGeneration/T5Stack[decoder]/StatelessEmbedding[embed_tokens]': {'shape': torch.Size([32, 4, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 0}}, 'outputs': {'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]_5': {'shape': torch.Size([32, 512, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [5]}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___58': {'shape': torch.Size([32, 8, 4, 4]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [5]}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___60': {'shape': torch.Size([32, 8, 4, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [5]}, 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[2]': {'shape': torch.Size([32, 4, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'used_by': [5]}}, 'devices': [('cpu' if DEBUG else 'cuda:4')], 'stage_depth': 1}, 5: {'stage_cls': Partition5, 'inputs': {'decoder_attention_mask': {'shape': torch.Size([32, 1, 4, 4]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'inverted_encoder_attention_mask': {'shape': torch.Size([32, 1, 1, 512]), 'dtype': torch.float32, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'lm_labels': {'shape': torch.Size([32, 4]), 'dtype': torch.int64, 'req_grad': False, 'is_batched': True, 'created_by': (- 1)}, 'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]_5': {'shape': torch.Size([32, 512, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 4}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___58': {'shape': torch.Size([32, 8, 4, 4]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 4}, 'T5ForConditionalGeneration/T5Stack[decoder]/tuple::__getitem___60': {'shape': torch.Size([32, 8, 4, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 4}, 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[2]': {'shape': torch.Size([32, 4, 512]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': True, 'created_by': 4}}, 'outputs': {'T5ForConditionalGeneration/CrossEntropyLoss[lm_loss]': {'shape': torch.Size([1]), 'dtype': torch.float32, 'req_grad': True, 'is_batched': False, 'used_by': [(- 1)]}}, 'devices': [('cpu' if DEBUG else 'cuda:5')], 'stage_depth': 0}}}
batch_dim = config['batch_dim']
for d in chain(config['model_inputs'].values(), config['model_outputs'].values()):
if d['is_batched']:
shape = d['shape']
d['shape'] = torch.Size(((shape[:batch_dim] + (batch_size,)) + shape[(batch_dim + 1):]))
for s in config['stages'].values():
for d in chain(s['inputs'].values(), s['outputs'].values()):
if d['is_batched']:
shape = d['shape']
d['shape'] = torch.Size(((shape[:batch_dim] + (batch_size,)) + shape[(batch_dim + 1):]))
return config
|
class Partition0(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[encoder]/StatelessEmbedding[embed_tokens]', 'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[0]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[1]', 'T5ForConditionalGeneration/T5Stack[decoder]/StatelessEmbedding[embed_tokens]']
TENSORS = ['T5ForConditionalGeneration/Parameter[shared_embed_weight]']
def __init__(self, layers, tensors, device='cuda:0'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1]
self.lookup = {'l_0': 'encoder.embed_tokens', 'l_1': 'encoder.dropout', 'l_2': 'encoder.0', 'l_3': 'encoder.1', 'l_4': 'decoder.embed_tokens', 'p_0': 'shared_embed_weight'}
self.to(self.device)
def forward(self, *args):
(attention_mask, decoder_input_ids, input_ids) = unflatten(args, self.input_structure)
t_0 = decoder_input_ids.size()
t_1 = input_ids.size()
t_1 = t_1[(- 1)]
t_1 = input_ids.view((- 1), t_1)
t_1 = self.l_0(self.p_0, t_1)
t_1 = self.l_1(t_1)
t_1 = self.l_2(t_1, attention_mask=attention_mask, position_bias=None, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_2 = t_1[0]
t_1 = t_1[1]
t_2 = self.l_3(t_2, attention_mask=attention_mask, position_bias=t_1, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = t_0[(- 1)]
t_0 = decoder_input_ids.view((- 1), t_0)
t_0 = self.l_4(self.p_0, t_0)
return list(flatten((t_1, t_2, t_0)))
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition1(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[encoder]/T5Block[2]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[3]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:1'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1]
self.lookup = {'l_0': 'encoder.2', 'l_1': 'encoder.3'}
self.to(self.device)
def forward(self, *args):
(attention_mask, x0, x1) = unflatten(args, self.input_structure)
t_0 = self.l_0(x1, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_1(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
return list(flatten((x0, t_0)))
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition2(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[encoder]/T5Block[4]', 'T5ForConditionalGeneration/T5Stack[encoder]/T5Block[5]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:2'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1]
self.lookup = {'l_0': 'encoder.4', 'l_1': 'encoder.5'}
self.to(self.device)
def forward(self, *args):
(attention_mask, x0, x1) = unflatten(args, self.input_structure)
t_0 = self.l_0(x1, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
t_0 = self.l_1(t_0, attention_mask=attention_mask, position_bias=x0, encoder_hidden_states=None, encoder_attention_mask=None, encoder_decoder_position_bias=None)
return (t_0,)
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition3(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[encoder]/T5LayerNorm[final_layer_norm]', 'T5ForConditionalGeneration/T5Stack[encoder]/Dropout[dropout]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:3'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1]
self.lookup = {'l_0': 'encoder.final_layer_norm', 'l_1': 'encoder.dropout'}
self.to(self.device)
def forward(self, *args):
x0 = unflatten(args, self.input_structure)[0]
t_0 = self.l_0(x0)
t_0 = self.l_1(t_0)
return (t_0,)
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition4(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[decoder]/Dropout[dropout]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[0]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[1]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[2]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:4'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1, 1]
self.lookup = {'l_0': 'decoder.dropout', 'l_1': 'decoder.0', 'l_2': 'decoder.1', 'l_3': 'decoder.2'}
self.to(self.device)
def forward(self, *args):
(decoder_attention_mask, inverted_encoder_attention_mask, x0, x1) = unflatten(args, self.input_structure)
t_0 = self.l_0(x1)
t_0 = self.l_1(t_0, attention_mask=decoder_attention_mask, position_bias=None, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=None)
t_1 = t_0[0]
t_2 = t_0[1]
t_0 = t_0[2]
t_1 = self.l_2(t_1, attention_mask=decoder_attention_mask, position_bias=t_2, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=t_0)
t_1 = self.l_3(t_1, attention_mask=decoder_attention_mask, position_bias=t_2, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=t_0)
return list(flatten((x0, t_2, t_0, t_1)))
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
class Partition5(nn.Module):
LAYER_SCOPES = ['T5ForConditionalGeneration/T5Stack[decoder]/T5Block[3]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[4]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5Block[5]', 'T5ForConditionalGeneration/T5Stack[decoder]/T5LayerNorm[final_layer_norm]', 'T5ForConditionalGeneration/T5Stack[decoder]/Dropout[dropout]', 'T5ForConditionalGeneration/Linear[lm_head]', 'T5ForConditionalGeneration/CrossEntropyLoss[lm_loss]']
TENSORS = []
def __init__(self, layers, tensors, device='cuda:5'):
super().__init__()
for (idx, layer_scope) in enumerate(self.LAYER_SCOPES):
self.add_module(f'l_{idx}', layers[layer_scope])
b = p = 0
for tensor_scope in self.TENSORS:
tensor = tensors[tensor_scope]
if isinstance(tensor, nn.Parameter):
self.register_parameter(f'p_{p}', tensor)
p += 1
else:
self.register_buffer(f'b_{b}', tensor)
b += 1
self.device = torch.device(device)
self.input_structure = [1, 1, 1, 1, 1, 1, 1]
self.lookup = {'l_0': 'decoder.3', 'l_1': 'decoder.4', 'l_2': 'decoder.5', 'l_3': 'decoder.final_layer_norm', 'l_4': 'decoder.dropout', 'l_5': 'lm_head', 'l_6': 'lm_loss'}
self.to(self.device)
def forward(self, *args):
(decoder_attention_mask, inverted_encoder_attention_mask, lm_labels, x0, x1, x2, x3) = unflatten(args, self.input_structure)
t_0 = self.l_0(x3, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_1(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_2(t_0, attention_mask=decoder_attention_mask, position_bias=x1, encoder_hidden_states=x0, encoder_attention_mask=inverted_encoder_attention_mask, encoder_decoder_position_bias=x2)
t_0 = self.l_3(t_0)
t_0 = self.l_4(t_0)
t_0 = (t_0 * 0.04419417382415922)
t_0 = self.l_5(t_0)
t_1 = t_0.size((- 1))
t_1 = t_0.view((- 1), t_1)
t_0 = lm_labels.view((- 1))
t_0 = self.l_6(t_1, t_0)
return (t_0,)
def state_dict(self, *args, **kwargs):
return state_dict(self, *args, **kwargs)
def load_state_dict(self, state):
return load_state_dict(self, state)
def named_parameters(self, recurse=True):
return named_parameters(self, recurse=recurse)
def named_buffers(self, recurse=True):
return named_buffers(self, recurse=recurse)
def cpu(self):
return cpu(self)
def cuda(self, device=None):
return cuda(self, device=device)
def to(self, *args, **kwargs):
return to(self, *args, **kwargs)
|
def traverse_model(module: nn.Module, depth: int, prefix: Optional[str]=None, basic_blocks: Tuple[nn.Module]=(), full: bool=False) -> Iterator[Tuple[(nn.Module, str, nn.Module)]]:
'\n iterate over model layers yielding the layer,layer_scope,encasing_module\n Parameters:\n -----------\n model:\n the model to iterate over\n depth:\n how far down in the model tree to go\n basic_blocks:\n a list of modules that if encountered will not be broken down\n full:\n whether to yield only layers specified by the depth and basick_block options or to yield all layers\n '
if (prefix is None):
prefix = type(module).__name__
for (name, sub_module) in module.named_children():
scope = (((prefix + '/') + type(sub_module).__name__) + f'[{name}]')
if ((len(list(sub_module.children())) == 0) or isinstance(sub_module, tuple(basic_blocks)) or (depth == 0)):
if full:
(yield (sub_module, scope, module, True))
else:
(yield (sub_module, scope, module))
else:
if full:
(yield (sub_module, scope, module, False))
(yield from traverse_model(sub_module, (depth - 1), scope, basic_blocks, full))
|
def layerDict(model: nn.Module, depth=1000, basic_blocks=()) -> Dict[(str, nn.Module)]:
return {s: l for (l, s, _) in traverse_model(model, depth, basic_blocks=basic_blocks)}
|
def traverse_params_buffs(module: nn.Module, prefix: Optional[str]=None) -> Iterator[Tuple[(torch.tensor, str)]]:
"\n iterate over model's buffers and parameters yielding obj,obj_scope\n\n Parameters:\n -----------\n model:\n the model to iterate over\n "
if (prefix is None):
prefix = type(module).__name__
for (param_name, param) in module.named_parameters(recurse=False):
param_scope = f'{prefix}/{type(param).__name__}[{param_name}]'
(yield (param, param_scope))
for (buffer_name, buffer) in module.named_buffers(recurse=False):
buffer_scope = f'{prefix}/{type(buffer).__name__}[{buffer_name}]'
(yield (buffer, buffer_scope))
for (name, sub_module) in module.named_children():
(yield from traverse_params_buffs(sub_module, (((prefix + '/') + type(sub_module).__name__) + f'[{name}]')))
|
def tensorDict(model: nn.Module) -> OrderedDict[(str, Tensor)]:
return collections.OrderedDict(((s, t) for (t, s) in traverse_params_buffs(model)))
|
def move_tensors(ts, device):
def move(t):
if isinstance(t, (nn.Module, Tensor)):
return t.to(device)
return t
return nested_map(move, ts)
|
def nested_map(func, ts, full=False):
if isinstance(ts, torch.Size):
return func(ts)
elif isinstance(ts, (list, tuple, set)):
return type(ts)((nested_map(func, t, full=full) for t in ts))
elif isinstance(ts, dict):
return {k: nested_map(func, v, full=full) for (k, v) in ts.items()}
elif (isinstance(ts, slice) and full):
start = nested_map(func, ts.start, full=full)
stop = nested_map(func, ts.stop, full=full)
step = nested_map(func, ts.step, full=full)
return slice(start, stop, step)
return func(ts)
|
def flatten(ts):
if isinstance(ts, torch.Size):
(yield ts)
elif isinstance(ts, (list, tuple, set)):
(yield from chain(*[flatten(t) for t in ts]))
elif isinstance(ts, dict):
(yield from chain(*[flatten(t) for (k, t) in sorted(ts.items(), key=(lambda t: t[0]))]))
else:
(yield ts)
|
def unflatten(xs, structure):
return _unflatten(xs, structure)[0]
|
def _unflatten(xs, structure):
if isinstance(structure, torch.Size):
return (xs[0], 1)
if (not isinstance(structure, (list, tuple, set, dict))):
return (xs[0], 1)
if isinstance(structure, (list, tuple, set)):
offset = 0
elements = []
for s in structure:
(e, n) = _unflatten(xs[offset:], s)
elements.append(e)
offset += n
return (type(structure)(elements), offset)
assert isinstance(structure, dict)
offset = 0
elements = dict()
for (k, v) in sorted(structure.items(), key=(lambda t: t[0])):
(e, n) = _unflatten(xs[offset:], v)
elements[k] = e
offset += n
return (elements, offset)
|
def state_dict(partition, *args, **kwargs):
state = nn.Module.state_dict(partition, *args, **kwargs)
lookup = partition.lookup
result = dict()
for (k, v) in state.items():
if (k in lookup):
result[lookup[k]] = v
else:
assert ('.' in k)
split_idx = k.find('.')
new_k = (lookup[k[:split_idx]] + k[split_idx:])
result[new_k] = v
return result
|
def load_state_dict(partition, state):
reverse_lookup = {v: k for (k, v) in partition.lookup.items()}
device = partition.device
keys = list(partition.state_dict(None).keys())
new_state = dict()
for k in keys:
if (k in reverse_lookup):
new_state[reverse_lookup[k]] = state[k].to(device)
continue
idx = k.rfind('.')
to_replace = k[:idx]
if (to_replace in reverse_lookup):
key = (reverse_lookup[to_replace] + k[idx:])
new_state[key] = state[k].to(device)
nn.Module.load_state_dict(partition, new_state, strict=True)
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.