in server/text_generation_server/models/seq2seq_lm.py [0:0]
def concatenate(cls, batches: List["Seq2SeqLMBatch"]) -> "Seq2SeqLMBatch":
"""Concatenate multiple batches together by padding internal torch tensors"""
# Used for padding
total_batch_size = 0
max_input_length = 0
max_decoder_input_length = 0
padding_right_offset = 0
for batch in batches:
total_batch_size += len(batch)
max_input_length = max(max_input_length, batch.max_input_length)
max_decoder_input_length = max(
max_decoder_input_length, batch.max_decoder_input_length
)
padding_right_offset = max(padding_right_offset, batch.padding_right_offset)
# Batch attributes
requests = []
requests_idx_mapping = {}
all_decoder_input_ids = []
input_lengths = []
decoder_input_lengths = []
prefix_offsets = []
read_offsets = []
next_token_choosers = []
stopping_criterias = []
top_n_tokens = []
max_tokens = 0
# Batch tensors
attention_mask = None
decoder_input_ids = None
decoder_attention_mask = None
encoder_last_hidden_state = None
top_n_tokens_tensor = None
past_key_values = []
# Used for slicing correctly inside the tensors
# Equivalent to a cumsum on batch sizes
start_index = 0
for i, batch in enumerate(batches):
# Extend all list attributes
requests.extend(batch.requests)
all_decoder_input_ids.extend(batch.all_decoder_input_ids)
input_lengths.extend(batch.input_lengths)
decoder_input_lengths.extend(batch.decoder_input_lengths)
prefix_offsets.extend(batch.prefix_offsets)
read_offsets.extend(batch.read_offsets)
next_token_choosers.extend(batch.next_token_choosers)
stopping_criterias.extend(batch.stopping_criterias)
top_n_tokens.extend(batch.top_n_tokens)
if i == 0:
requests_idx_mapping = batch.requests_idx_mapping
else:
# We need to offset the mapping for each batch by the cumulative batch size
for k, v in batch.requests_idx_mapping.items():
requests_idx_mapping[k] = v + start_index
# Slicing end index for this batch
end_index = start_index + len(batch)
# We only concatenate batches that did at least one step
if batch.encoder_last_hidden_state is None:
raise ValueError("Batch encoder_last_hidden_state cannot be None")
# Create padded tensor
if attention_mask is None:
attention_mask = batch.attention_mask.new_zeros(
(total_batch_size, max_input_length),
)
# Copy to correct indices
attention_mask[start_index:end_index, -batch.max_input_length :] = (
batch.attention_mask[:, -batch.max_input_length :]
)
# Create padded tensor
if decoder_input_ids is None:
decoder_input_ids = batch.decoder_input_ids.new_zeros(
(total_batch_size, 1),
)
# Copy to correct indices
decoder_input_ids[start_index:end_index] = batch.decoder_input_ids
# Create padded tensor
if decoder_attention_mask is None:
# As decoder_attention_mask might not exist, we use `batch.attention_mask` for device here
decoder_attention_mask = batch.attention_mask.new_zeros(
(total_batch_size, max_decoder_input_length + padding_right_offset),
)
# If the decoder mask does not exist yet, all generations started at the same time and we never concatenated
# this batch. All generations are of length `batch.max_decoder_input_length`.
left_offset = max_decoder_input_length - batch.max_decoder_input_length
if batch.decoder_attention_mask is None:
decoder_attention_mask[
start_index:end_index,
left_offset:-padding_right_offset,
] = 1
# If it exists, we need to index
else:
batch_left_offset = (
batch.decoder_attention_mask.shape[1]
- batch.max_decoder_input_length
- batch.padding_right_offset
)
decoder_attention_mask[
start_index:end_index,
left_offset:-padding_right_offset,
] = batch.decoder_attention_mask[
:,
batch_left_offset : -batch.padding_right_offset,
]
# Create padded tensor
if encoder_last_hidden_state is None:
encoder_last_hidden_state = batch.encoder_last_hidden_state.new_zeros(
(
total_batch_size,
max_input_length,
batch.encoder_last_hidden_state.shape[-1],
),
)
if top_n_tokens_tensor is None:
top_n_tokens_tensor = batches[0].top_n_tokens_tensor.new_zeros(
total_batch_size,
)
top_n_tokens_tensor[start_index:end_index] = batch.top_n_tokens_tensor
# Copy to correct indices
encoder_last_hidden_state[
start_index:end_index, -batch.max_input_length :, :
] = batch.encoder_last_hidden_state[:, -batch.max_input_length :, :]
batch.encoder_last_hidden_state = None
# Ensure that we can update tensors in-place
if isinstance(batch.past_key_values[0], tuple):
batch.past_key_values = [
[t for t in layer] for layer in batch.past_key_values
]
# Add eventual padding tokens that were added while concatenating
max_tokens += batch.max_tokens + (
max_input_length
- batch.max_input_length
+ max_decoder_input_length
- batch.max_decoder_input_length
) * len(batch)
start_index = end_index
# Determine shapes for new past kv tensors
first_past_kvs = batches[0].past_key_values
_, num_heads, _, head_dim = first_past_kvs[0][0].shape
padded_dec_t_shape = (
total_batch_size,
num_heads,
(max_decoder_input_length - 1),
head_dim,
)
padded_enc_t_shape = (
total_batch_size,
num_heads,
max_input_length,
head_dim,
)
# Iterate over attention layers
for j in range(len(first_past_kvs)):
past_key_values.append([])
# Decoder past
for k in range(0, 2):
# Initialize tensors
padded_past_values = first_past_kvs[j][k].new_zeros(padded_dec_t_shape)
past_key_values[j].append(padded_past_values)
start_index = 0
for batch in batches:
t = batch.past_key_values[j][k]
# Clear reference to the original tensor
batch.past_key_values[j][k] = None
# Slicing end index for this batch
end_index = start_index + len(batch)
# We slice the past keys and values to remove the padding from previous batches
past_seq_len = batch.max_decoder_input_length - 1
padded_past_values[start_index:end_index, :, -past_seq_len:, :] = t[
:, :, -past_seq_len:, :
]
del t
start_index = end_index
# Encoder past
for k in range(2, 4):
# Initialize tensors
padded_past_values = first_past_kvs[j][k].new_zeros(padded_enc_t_shape)
past_key_values[j].append(padded_past_values)
start_index = 0
for batch in batches:
t = batch.past_key_values[j][k]
# Clear reference to the original tensor
batch.past_key_values[j][k] = None
# Slicing end index for this batch
end_index = start_index + len(batch)
# We slice the past keys and values to remove the padding from previous batches
padded_past_values[
start_index:end_index, :, -batch.max_input_length :, :
] = t[:, :, -batch.max_input_length :, :]
del t
start_index = end_index
return cls(
batch_id=batches[0].batch_id,
requests=requests,
requests_idx_mapping=requests_idx_mapping,
input_ids=None,
attention_mask=attention_mask,
decoder_input_ids=decoder_input_ids,
all_decoder_input_ids=all_decoder_input_ids,
decoder_attention_mask=decoder_attention_mask,
encoder_last_hidden_state=encoder_last_hidden_state,
past_key_values=past_key_values,
input_lengths=input_lengths,
decoder_input_lengths=decoder_input_lengths,
prefix_offsets=prefix_offsets,
read_offsets=read_offsets,
next_token_choosers=next_token_choosers,
stopping_criterias=stopping_criterias,
top_n_tokens=top_n_tokens,
top_n_tokens_tensor=top_n_tokens_tensor,
max_input_length=max_input_length,
max_decoder_input_length=max_decoder_input_length,
padding_right_offset=padding_right_offset,
max_tokens=max_tokens,
)