huggingface/text-generation-inference
5
0
Fork 0
mirror of https://github.com/huggingface/text-generation-inference.git synced 2025-09-10 03:44:54 +00:00
Code Issues Packages Projects Releases Wiki Activity

equivalent changes for seq2seq_lm

Browse Source
This commit is contained in:
Nick Hill 2023-04-24 07:25:25 +01:00
parent ab20142c14
commit 8e34beed32
2 changed files with 125 additions and 83 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

33
server/tests/models/test_seq2seq_lm.py
View File

@ -219,6 +219,19 @@ def test_batch_concatenate(
next_batch_1 = default_multi_requests_seq2seq_lm_batch next_batch_1 = default_multi_requests_seq2seq_lm_batch
_, next_batch_1 = default_seq2seq_lm.generate_token(next_batch_1) _, next_batch_1 = default_seq2seq_lm.generate_token(next_batch_1)
# Copy hidden state because it is removed from the concatenated branches
next_batch_0_encoder_last_hidden_state = next_batch_0.encoder_last_hidden_state
next_batch_1_encoder_last_hidden_state = next_batch_1.encoder_last_hidden_state
# Clone past_key_values before concatenating to compare after,
# because they are removed from the concatenated batches
next_batch_0_past_key_values = [
[t.clone() for t in layer] for layer in next_batch_0.past_key_values
]
next_batch_1_past_key_values = [
[t.clone() for t in layer] for layer in next_batch_1.past_key_values
]
next_batch = Seq2SeqLMBatch.concatenate([next_batch_0, next_batch_1]) next_batch = Seq2SeqLMBatch.concatenate([next_batch_0, next_batch_1])
assert next_batch.batch_id == 0 assert next_batch.batch_id == 0
@ -239,11 +252,11 @@ def test_batch_concatenate(
assert torch.equal( assert torch.equal(
next_batch.encoder_last_hidden_state[0], next_batch.encoder_last_hidden_state[0],
next_batch_0.encoder_last_hidden_state[0, -2:], next_batch_0_encoder_last_hidden_state[0, -2:],
) )
assert torch.equal( assert torch.equal(
next_batch.encoder_last_hidden_state[1:], next_batch.encoder_last_hidden_state[1:],
next_batch_1.encoder_last_hidden_state[:, -2:], next_batch_1_encoder_last_hidden_state[:, -2:],
) )
assert next_batch.input_lengths == [2, 2, 2] assert next_batch.input_lengths == [2, 2, 2]
@ -275,24 +288,24 @@ def test_batch_concatenate(
) )
for i, past in enumerate(next_batch.past_key_values): for i, past in enumerate(next_batch.past_key_values):
assert torch.equal(next_batch_0.past_key_values[i][0][0, :, -2:, :], past[0][0]) assert torch.equal(next_batch_0_past_key_values[i][0][0, :, -2:, :], past[0][0])
assert torch.equal( assert torch.equal(
next_batch_1.past_key_values[i][0][:, :, -1:, :], past[0][1:, :, -1:, :] next_batch_1_past_key_values[i][0][:, :, -1:, :], past[0][1:, :, -1:, :]
) )
assert torch.equal(next_batch_0.past_key_values[i][1][0, :, -2:, :], past[1][0]) assert torch.equal(next_batch_0_past_key_values[i][1][0, :, -2:, :], past[1][0])
assert torch.equal( assert torch.equal(
next_batch_1.past_key_values[i][1][:, :, -1:, :], past[1][1:, :, -1:, :] next_batch_1_past_key_values[i][1][:, :, -1:, :], past[1][1:, :, -1:, :]
) )
assert torch.equal(next_batch_0.past_key_values[i][2][0, :, -2:, :], past[2][0]) assert torch.equal(next_batch_0_past_key_values[i][2][0, :, -2:, :], past[2][0])
assert torch.equal( assert torch.equal(
next_batch_1.past_key_values[i][2][:, :, -2:, :], past[2][1:] next_batch_1_past_key_values[i][2][:, :, -2:, :], past[2][1:]
) )
assert torch.equal(next_batch_0.past_key_values[i][3][0, :, -2:, :], past[3][0]) assert torch.equal(next_batch_0_past_key_values[i][3][0, :, -2:, :], past[3][0])
assert torch.equal( assert torch.equal(
next_batch_1.past_key_values[i][3][:, :, -2:, :], past[3][1:] next_batch_1_past_key_values[i][3][:, :, -2:, :], past[3][1:]
) )
for _ in range(3): for _ in range(3):

159
server/text_generation_server/models/seq2seq_lm.py
View File

@ -25,7 +25,7 @@ class Seq2SeqLMBatch(Batch):
requests_idx_mapping: Dict[int, int] requests_idx_mapping: Dict[int, int]
# Encoder values # Encoder values
input_ids: torch.Tensor input_ids: Optional[torch.Tensor]
attention_mask: torch.Tensor attention_mask: torch.Tensor
# Decoder values # Decoder values
@ -164,6 +164,7 @@ class Seq2SeqLMBatch(Batch):
max_input_length = 0 max_input_length = 0
max_decoder_input_length = 0 max_decoder_input_length = 0
padding_right_offset = 0
for i, r in enumerate(requests): for i, r in enumerate(requests):
idx = self.requests_idx_mapping[r.id] idx = self.requests_idx_mapping[r.id]
@ -184,45 +185,53 @@ class Seq2SeqLMBatch(Batch):
max_decoder_input_length = max( max_decoder_input_length = max(
max_decoder_input_length, request_decoder_input_length max_decoder_input_length, request_decoder_input_length
) )
padding_right_offset = max(
padding_right_offset,
self.stopping_criterias[idx].max_new_tokens - self.stopping_criterias[idx].current_tokens
)
next_token_choosers.append(self.next_token_choosers[idx]) next_token_choosers.append(self.next_token_choosers[idx])
stopping_criterias.append(self.stopping_criterias[idx]) stopping_criterias.append(self.stopping_criterias[idx])
# Apply indices to input_ids, attention mask, past key values and other items that need to be cached # Apply indices to input_ids, attention mask, past key values and other items that need to be cached
decoder_input_ids = self.decoder_input_ids[keep_indices] self.decoder_input_ids = self.decoder_input_ids[keep_indices]
attention_mask = self.attention_mask[keep_indices] self.attention_mask = self.attention_mask[keep_indices, -max_input_length:]
if self.decoder_attention_mask is not None: if self.decoder_attention_mask is not None:
decoder_attention_mask = self.decoder_attention_mask[keep_indices] self.decoder_attention_mask = self.decoder_attention_mask[
else: keep_indices,
decoder_attention_mask = None -(self.padding_right_offset + max_decoder_input_length):
(self.decoder_attention_mask.shape[1] - self.padding_right_offset) + padding_right_offset,
encoder_last_hidden_state = self.encoder_last_hidden_state[keep_indices]
past_key_values = [
[t[keep_indices] for t in layer] for layer in self.past_key_values
] ]
return Seq2SeqLMBatch( self.encoder_last_hidden_state = self.encoder_last_hidden_state[keep_indices, -max_input_length:]
batch_id=self.batch_id,
requests=requests, # Ensure that past_key_values tensors can be updated in-place
requests_idx_mapping=requests_idx_mapping, if type(self.past_key_values[0]) == tuple:
input_ids=None, self.past_key_values = [[t for t in layer] for layer in self.past_key_values]
attention_mask=attention_mask,
decoder_input_ids=decoder_input_ids, decoder_past_seq_len = max_decoder_input_length - 1
all_decoder_input_ids=all_decoder_input_ids, for layer in self.past_key_values:
decoder_attention_mask=decoder_attention_mask, layer[0] = layer[0][keep_indices, :, -decoder_past_seq_len:]
encoder_last_hidden_state=encoder_last_hidden_state, layer[1] = layer[1][keep_indices, :, -decoder_past_seq_len:]
past_key_values=past_key_values, layer[2] = layer[2][keep_indices, :, -max_input_length:]
input_lengths=input_lengths, layer[3] = layer[3][keep_indices, :, -max_input_length:]
decoder_input_lengths=decoder_input_lengths,
offsets=offsets, self.requests = requests
token_offsets=token_offsets, self.requests_idx_mapping = requests_idx_mapping
next_token_choosers=next_token_choosers, self.input_ids = None
stopping_criterias=stopping_criterias, self.all_decoder_input_ids = all_decoder_input_ids
max_input_length=max_input_length, self.input_lengths = input_lengths
max_decoder_input_length=max_decoder_input_length, self.decoder_input_lengths = decoder_input_lengths
padding_right_offset=self.padding_right_offset, self.offsets = offsets
) self.token_offsets = token_offsets
self.next_token_choosers = next_token_choosers
self.stopping_criterias = stopping_criterias
self.max_input_length = max_input_length
self.max_decoder_input_length = max_decoder_input_length
self.padding_right_offset = padding_right_offset
return self
@classmethod @classmethod
@tracer.start_as_current_span("concatenate") @tracer.start_as_current_span("concatenate")
@ -350,58 +359,78 @@ class Seq2SeqLMBatch(Batch):
encoder_last_hidden_state[ encoder_last_hidden_state[
start_index:end_index, -batch.max_input_length :, : start_index:end_index, -batch.max_input_length :, :
] = batch.encoder_last_hidden_state[:, -batch.max_input_length :, :] ] = batch.encoder_last_hidden_state[:, -batch.max_input_length :, :]
batch.encoder_last_hidden_state = None
# Iterate over attention layers # Ensure that we can update tensors in-place
for j, past in enumerate(batch.past_key_values): if type(batch.past_key_values[0]) == tuple:
_, num_heads, _, head_dim = past[0].shape batch.past_key_values = [[t for t in layer] for layer in batch.past_key_values]
# This will run only once per layer start_index = end_index
if j == len(past_key_values):
past_key_values.append([])
# Decoder past # Determine shapes for new past kv tensors
for k, t in enumerate(past[:2]): first_past_kvs = batches[0].past_key_values
padded_t_shape = ( _, num_heads, _, head_dim = first_past_kvs[0][0].shape
padded_dec_t_shape = (
total_batch_size, total_batch_size,
num_heads, num_heads,
(max_decoder_input_length - 1), (max_decoder_input_length - 1),
head_dim, head_dim,
) )
# Initialize tensors padded_enc_t_shape = (
# This will run only once per layer and per past tensor
if k == len(past_key_values[j]):
past_key_values[j].append(t.new_zeros(padded_t_shape))
# We slice the past keys and values to remove the padding from previous batches
past_key_values[j][k][
start_index:end_index,
:,
-(batch.max_decoder_input_length - 1) :,
:,
] = t[:, :, -(batch.max_decoder_input_length - 1) :, :]
# encoder past
for k, t in enumerate(past[2:]):
padded_t_shape = (
total_batch_size, total_batch_size,
num_heads, num_heads,
max_input_length, max_input_length,
head_dim, head_dim,
) )
idx = k + 2 # 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 # Initialize tensors
# This will run only once per layer and per past tensor padded_past_values = first_past_kvs[j][k].new_zeros(padded_dec_t_shape)
if idx == len(past_key_values[j]): past_key_values[j].append(padded_past_values)
past_key_values[j].append(t.new_zeros(padded_t_shape))
past_key_values[j][idx][ start_index = 0
start_index:end_index, :, -batch.max_input_length :, : for batch in batches:
] = t[:, :, -batch.max_input_length :, :] 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 += len(batch) 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( return cls(
batch_id=batches[0].batch_id, batch_id=batches[0].batch_id,