Support BERTModel (first encoder-only embedding model) (#9056)

Signed-off-by: Max de Bayser <maxdebayser@gmail.com>
Signed-off-by: Max de Bayser <mbayser@br.ibm.com>
Co-authored-by: Andrew Feldman <afeldman@neuralmagic.com>
Co-authored-by: afeldman-nm <156691304+afeldman-nm@users.noreply.github.com>
Co-authored-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
Co-authored-by: laishzh <laishengzhang@gmail.com>
Co-authored-by: Max de Bayser <maxdebayser@gmail.com>
Co-authored-by: Max de Bayser <mbayser@br.ibm.com>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
This commit is contained in:
Robert Shaw
2024-10-17 23:21:01 +00:00
committed by GitHub
co-authored by Andrew Feldman afeldman-nm Woosuk Kwon laishzh Max de Bayser Max de Bayser Cyrus Leung
parent bb76538bbd
commit 343f8e0905
6 changed files with 497 additions and 15 deletions
+10 -2
View File
@@ -12,6 +12,7 @@ class PoolingType(IntEnum):
"""Enumeration for different types of pooling methods."""
LAST = 0
ALL = 1
CLS = 2
class Pooler(nn.Module):
@@ -23,12 +24,13 @@ class Pooler(nn.Module):
3. Returns structured results as `PoolerOutput`.
Attributes:
pooling_type: The type of pooling to use (LAST, AVERAGE, MAX).
pooling_type: The type of pooling to use (LAST, ALL, CLS).
normalize: Whether to normalize the pooled data.
"""
def __init__(self, pooling_type: PoolingType, normalize: bool):
super().__init__()
self.pooling_type = pooling_type
self.normalize = normalize
@@ -38,10 +40,16 @@ class Pooler(nn.Module):
pooling_metadata: PoolingMetadata,
) -> PoolerOutput:
"""Pools specific information from hidden states based on metadata."""
prompt_lens = PoolingTensors.from_pooling_metadata(
pooling_metadata, hidden_states.device).prompt_lens
if self.pooling_type == PoolingType.LAST:
if self.pooling_type is PoolingType.CLS:
first_token_flat_indices = torch.zeros_like(prompt_lens)
first_token_flat_indices[1:] += torch.cumsum(prompt_lens,
dim=0)[:-1]
pooled_data = hidden_states[first_token_flat_indices]
elif self.pooling_type == PoolingType.LAST:
last_token_flat_indices = torch.cumsum(prompt_lens, dim=0) - 1
pooled_data = hidden_states[last_token_flat_indices]
elif self.pooling_type == PoolingType.ALL: