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

Fix alibi #222

Merged
merged 1 commit into from
Jan 6, 2022
Merged

Fix alibi #222

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
104 changes: 61 additions & 43 deletions megatron/model/transformer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -303,11 +303,27 @@ def forward(self, hidden_states, attention_mask, layer_past=None,
query_layer, key_layer = apply_rotary_fn(query_layer, key_layer, cos, sin, offset=offset)

# Raw attention scores. [b * np, sq, sk]
matmul_result = torch.baddbmm(
matmul_result,
query_layer.transpose(0, 1), # [b * np, sq, hn]
key_layer.transpose(0, 1).transpose(1, 2), # [b * np, hn, sk]
beta=0.0 if alibi is None else 1.0, alpha=(1.0/self.norm_factor))
if alibi is None:
matmul_result = torch.baddbmm(
matmul_result,
query_layer.transpose(0, 1), # [b * np, sq, hn]
key_layer.transpose(0, 1).transpose(1, 2), # [b * np, hn, sk]
beta=0.0, alpha=(1.0/self.norm_factor))
else:
if not hasattr(self, "logged_alibi"):
logger.debug("Using Alibi.")
self.logged_alibi = True
Comment on lines +313 to +315
Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Testing purposes.

Copy link
Collaborator

@SaulLu SaulLu Jan 5, 2022 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Just for information, on my side I was going to propose to isolate the 3 ways to calculate the score according to the positional embedding by creating a method for each positional embedding method and then wrapping these methods in log_debug_usage to have a log (as for the activation functions) to detect in the tests.

The micro advantage is that it also allows to test rotary and absolute (in all cases) but I don"t mind if you think it's easier to keep it like you did.

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah that's true that your version looks nice. I just think it's an incorrect abstraction. There's no reason to group all the positional embeddings together (they get applied in different places, they do different things, they have different constraints despite having the common purpose on given sequential information). One could argue that using a pure causal mask is a position embedding mechanism.

What I was thinking of is abstracting only the alibi function in a seperate function to use the pretty decorator, but I was lazy ^^' @log_debug_usage(logger, msg)

Copy link
Contributor

@ofirpress ofirpress Jan 7, 2022 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

using a pure causal mask is a position embedding mechanism

I agree with this statement :) Models that don't have any position embeddings (like sinusoidal or learned or alibi) are actually able to achieve good (but not great) PPL because the causal mask encodes some kind of order.


if self.apply_query_key_layer_scaling:
beta = 1.0 / self.layer_number
else:
beta = 1.0
thomasw21 marked this conversation as resolved.
Show resolved Hide resolved

matmul_result = torch.baddbmm(
matmul_result,
query_layer.transpose(0, 1), # [b * np, sq, hn]
key_layer.transpose(0, 1).transpose(1, 2), # [b * np, hn, sk]
beta=beta, alpha=(1.0 / self.norm_factor))

# change view to [b, np, sq, sk]
attention_scores = matmul_result.view(*output_size)
Expand Down Expand Up @@ -470,9 +486,19 @@ def __init__(self, init_method, output_layer_init_method,
self.mlp = ParallelMLP(init_method,
output_layer_init_method)

# Alibi
if args.position_embedding_type == PositionEmbeddingType.alibi:
self.alibi = self._build_alibi_tensor(args.seq_length, args.num_attention_heads, args.micro_batch_size).to(torch.cuda.current_device())
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Does args.micro_batch_size work with batch size rampup? And if not, do we care?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

micro batch size doesn't increase during batch size rampup. It's constant.

We should care since all our experiment run with batch size rampup, but I would expect it to crash badly if it doesn't match.

if args.params_dtype == torch.float16:
self.alibi = self.alibi.to(torch.float16)
elif args.params_dtype == torch.bfloat16:
self.alibi = self.alibi.to(torch.bfloat16)
else:
self.alibi = None

def forward(self, hidden_states, attention_mask,
encoder_output=None, enc_dec_attn_mask=None,
layer_past=None, get_key_value=False, alibi=None):
layer_past=None, get_key_value=False):
# hidden_states: [b, s, h]

# Layer norm at the beginning of the transformer layer.
Expand All @@ -483,7 +509,7 @@ def forward(self, hidden_states, attention_mask,
attention_mask,
layer_past=layer_past,
get_key_value=get_key_value,
alibi=alibi)
alibi=self.alibi)

if get_key_value:
attention_output, presents = attention_output
Expand Down Expand Up @@ -561,6 +587,30 @@ def forward(self, hidden_states, attention_mask,

return output

@staticmethod
def _build_alibi_tensor(max_seq_len, num_attention_heads, batch_size):
# Based on https://github.com/ofirpress/attention_with_linear_biases/blob/a35aaca144e0eb6b789dfcb46784c4b8e31b7983/fairseq/models/transformer.py#L742
"""Returns tensor shaped (batch_size * num_attention_heads, 1, max_seq_len)"""

def get_slopes(n):
def get_slopes_power_of_2(n):
start = (2 ** (-2 ** -(math.log2(n) - 3)))
ratio = start
return [start * ratio ** i for i in range(n)]

if math.log2(n).is_integer():
return get_slopes_power_of_2(n)
else:
closest_power_of_2 = 2 ** math.floor(math.log2(n))
return get_slopes_power_of_2(closest_power_of_2) + get_slopes(2 * closest_power_of_2)[0::2][
:n - closest_power_of_2]

slopes = torch.Tensor(get_slopes(num_attention_heads))
alibi = slopes.unsqueeze(1).unsqueeze(1) * torch.arange(max_seq_len).unsqueeze(0).unsqueeze(0).expand(
num_attention_heads, -1, -1)
alibi = alibi.repeat(batch_size, 1, 1)
return alibi

class ParallelTransformerLayerPipe(ParallelTransformerLayer):
"""Extends ParallelTransformerLayer to forward attention_mask through the pipeline.

Expand Down Expand Up @@ -600,27 +650,6 @@ def forward(self, inputs, **kwargs):
class ParallelTransformer(MegatronModule):
"""Transformer class."""

@staticmethod
def _build_alibi_tensor(max_seq_len, num_attention_heads, batch_size):
# Based on https://github.com/ofirpress/attention_with_linear_biases/blob/a35aaca144e0eb6b789dfcb46784c4b8e31b7983/fairseq/models/transformer.py#L742
"""Returns tensor shaped (batch_size * num_attention_heads, 1, max_seq_len)"""
def get_slopes(n):
def get_slopes_power_of_2(n):
start = (2 ** (-2 ** -(math.log2(n) - 3)))
ratio = start
return [start * ratio ** i for i in range(n)]

if math.log2(n).is_integer():
return get_slopes_power_of_2(n)
else:
closest_power_of_2 = 2 ** math.floor(math.log2(n))
return get_slopes_power_of_2(closest_power_of_2) + get_slopes(2 * closest_power_of_2)[0::2][
:n - closest_power_of_2]
slopes = torch.Tensor(get_slopes(num_attention_heads))
alibi = slopes.unsqueeze(1).unsqueeze(1) * torch.arange(max_seq_len).unsqueeze(0).unsqueeze(0).expand(num_attention_heads, -1, -1)
alibi = alibi.repeat(batch_size, 1, 1)
return alibi

def __init__(self, init_method, output_layer_init_method,
layer_type=LayerType.encoder,
self_attn_mask_type=AttnMaskType.padding,
Expand Down Expand Up @@ -687,20 +716,11 @@ def build_layer(layer_number):
get_cuda_rng_tracker = deepspeed.checkpointing.get_cuda_rng_tracker
checkpoint = deepspeed.checkpointing.checkpoint

if args.position_embedding_type == PositionEmbeddingType.alibi:
self.alibi = self._build_alibi_tensor(args.seq_length, args.num_attention_heads, args.micro_batch_size).to(torch.cuda.current_device())
if args.params_dtype == torch.float16:
self.alibi = self.alibi.to(torch.float16)
elif args.params_dtype == torch.bfloat16:
self.alibi = self.alibi.to(torch.bfloat16)
else:
self.alibi = None

def _get_layer(self, layer_number):
return self.layers[layer_number]

def _checkpointed_forward(self, hidden_states, attention_mask,
encoder_output, enc_dec_attn_mask, alibi=None):
encoder_output, enc_dec_attn_mask):
"""Forward method with activation checkpointing."""
def custom(start, end):
def custom_forward(*inputs):
Expand All @@ -710,7 +730,7 @@ def custom_forward(*inputs):
enc_dec_attn_mask = inputs[3]
for index in range(start, end):
layer = self._get_layer(index)
x_ = layer(x_, attention_mask, encoder_output, enc_dec_attn_mask, alibi=alibi)
x_ = layer(x_, attention_mask, encoder_output, enc_dec_attn_mask)
return x_
return custom_forward

Expand Down Expand Up @@ -767,8 +787,7 @@ def forward(self, hidden_states, attention_mask, layer_past=None,
hidden_states = self._checkpointed_forward(hidden_states,
attention_mask,
encoder_output,
enc_dec_attn_mask,
alibi=self.alibi)
enc_dec_attn_mask)
else:
if get_key_value:
presents = []
Expand All @@ -782,8 +801,7 @@ def forward(self, hidden_states, attention_mask, layer_past=None,
encoder_output=encoder_output,
enc_dec_attn_mask=enc_dec_attn_mask,
layer_past=past,
get_key_value=get_key_value,
alibi=self.alibi)
get_key_value=get_key_value)
if get_key_value:
hidden_states, present = hidden_states
presents.append(present)
Expand Down
19 changes: 18 additions & 1 deletion tests/test_training.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -225,6 +225,20 @@ def get_variation_config(self, variation, output_dir, n_samples=None):
--deepspeed_config {self.test_file_dir_str}/ds_config.json
""".split()

elif variation == "alibi":
new_args = f"""
--rampup-batch-size 2 2 {n_samples}
--train-samples {n_samples}

--lr-decay-samples 6

--position-embedding-type alibi
""".split()

new_ds_args = f"""
--deepspeed_config {self.test_file_dir_str}/ds_config.json
""".split()

else:
raise ValueError(f"Don't know of variation {variation}")

Expand All @@ -234,7 +248,7 @@ def get_variation_config(self, variation, output_dir, n_samples=None):
return args, ds_args, num_gpus


@parameterized.expand(["base", "cl", "bnb", "glu"])
@parameterized.expand(["base", "cl", "bnb", "glu", "alibi"])
def test_training_all(self, variation):

# optional runs
Expand Down Expand Up @@ -277,6 +291,9 @@ def test_training_all(self, variation):
if variation == "glu":
self.assertIn("Using GLU activation: GELU", cs.out)

if variation == "alibi":
self.assertIn("Using Alibi", cs.out)

# 2. test training from checkpoint: resume
# now do it again, this time resuming from the checkpoint
with CaptureStdout() as cs:
Expand Down