Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

feat(server): reduce memory requirement #214

Merged
merged 4 commits into from
Apr 24, 2023
Merged

feat(server): reduce memory requirement #214

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
12326ef
feat(server): reduce memory requirement
njhill Apr 20, 2023
0b1d001
update unit tests
njhill Apr 24, 2023
ab20142
trim to new max input length in filter()
njhill Apr 24, 2023
8e34bee
equivalent changes for seq2seq_lm
njhill Apr 24, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
23 changes: 17 additions & 6 deletions server/tests/models/test_bloom.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -175,12 +175,14 @@ def test_causal_lm_generate_token_completion_multi(
generations[1].generated_text.generated_tokens
== default_multi_requests_bloom_batch.stopping_criterias[1].max_new_tokens
)
# Copy stopping_criterias before filtering
stopping_criterias = default_multi_requests_bloom_batch.stopping_criterias.copy()

next_batch = next_batch.filter([next_batch.requests[0]])

for _ in range(
default_multi_requests_bloom_batch.stopping_criterias[0].max_new_tokens
- default_multi_requests_bloom_batch.stopping_criterias[1].max_new_tokens
stopping_criterias[0].max_new_tokens
- stopping_criterias[1].max_new_tokens
- 1
):
generations, next_batch = default_bloom.generate_token(next_batch)
Expand Down Expand Up @@ -212,6 +214,15 @@ def test_batch_concatenate(
next_batch_1 = default_multi_requests_bloom_batch
_, next_batch_1 = default_bloom.generate_token(next_batch_1)

# Clone past_key_values before concatenating to compare after,
# because they are removed from the concatenated batches
next_batch_0_past_key_values = [
(k.clone(), v.clone()) for (k, v) in next_batch_0.past_key_values
]
next_batch_1_past_key_values = [
(k.clone(), v.clone()) for (k, v) in next_batch_1.past_key_values
]

next_batch = BloomCausalLMBatch.concatenate([next_batch_0, next_batch_1])

assert torch.equal(next_batch.all_input_ids[0], next_batch_0.all_input_ids[0])
Expand Down Expand Up @@ -246,15 +257,15 @@ def test_batch_concatenate(
assert all([p[1].shape == (3, 16, 2, 64) for p in 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][:, :, -2:], past[0][0])
assert torch.equal(next_batch_0_past_key_values[i][0][:, :, -2:], past[0][0])
assert torch.equal(
next_batch_1.past_key_values[i][0][:, :, -1:],
next_batch_1_past_key_values[i][0][:, :, -1:],
past[0][1:, :, :, -1].reshape(-1, 64, 1),
)

assert torch.equal(next_batch_0.past_key_values[i][1][:, -2:, :], past[1][0])
assert torch.equal(next_batch_0_past_key_values[i][1][:, -2:, :], past[1][0])
assert torch.equal(
next_batch_1.past_key_values[i][1][:, -1:, :],
next_batch_1_past_key_values[i][1][:, -1:, :],
past[1][1:, :, -1, :].reshape(-1, 1, 64),
)

Expand Down
23 changes: 17 additions & 6 deletions server/tests/models/test_causal_lm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -173,12 +173,14 @@ def test_causal_lm_generate_token_completion_multi(
generations[1].generated_text.generated_tokens
== default_multi_requests_causal_lm_batch.stopping_criterias[1].max_new_tokens
)
# Copy stopping_criterias before filtering
stopping_criterias = default_multi_requests_causal_lm_batch.stopping_criterias.copy()

next_batch = next_batch.filter([next_batch.requests[0]])

for _ in range(
default_multi_requests_causal_lm_batch.stopping_criterias[0].max_new_tokens
- default_multi_requests_causal_lm_batch.stopping_criterias[1].max_new_tokens
stopping_criterias[0].max_new_tokens
- stopping_criterias[1].max_new_tokens
- 1
):
generations, next_batch = default_causal_lm.generate_token(next_batch)
Expand Down Expand Up @@ -209,6 +211,15 @@ def test_batch_concatenate(
next_batch_1 = default_multi_requests_causal_lm_batch
_, next_batch_1 = default_causal_lm.generate_token(next_batch_1)

# Clone past_key_values before concatenating to compare after,
# because they are removed from the concatenated batches
next_batch_0_past_key_values = [
(k.clone(), v.clone()) for (k, v) in next_batch_0.past_key_values
]
next_batch_1_past_key_values = [
(k.clone(), v.clone()) for (k, v) in next_batch_1.past_key_values
]

next_batch = CausalLMBatch.concatenate([next_batch_0, next_batch_1])

assert torch.equal(next_batch.all_input_ids[0], next_batch_0.all_input_ids[0])
Expand Down Expand Up @@ -244,14 +255,14 @@ def test_batch_concatenate(
assert all([p[1].shape == (3, 12, 2, 64) for p in 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(
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(
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:, :]
)

for _ in range(
Expand Down
33 changes: 23 additions & 10 deletions server/tests/models/test_seq2seq_lm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -219,6 +219,19 @@ def test_batch_concatenate(
next_batch_1 = default_multi_requests_seq2seq_lm_batch
_, 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])

assert next_batch.batch_id == 0
Expand All @@ -239,11 +252,11 @@ def test_batch_concatenate(

assert torch.equal(
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(
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]
Expand Down Expand Up @@ -275,24 +288,24 @@ def test_batch_concatenate(
)

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(
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(
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(
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(
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):
Expand Down
Loading