343f8e0905
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>
72 lines
2.4 KiB
Python
72 lines
2.4 KiB
Python
from enum import IntEnum
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
|
|
from vllm.model_executor.pooling_metadata import (PoolingMetadata,
|
|
PoolingTensors)
|
|
from vllm.sequence import EmbeddingSequenceGroupOutput, PoolerOutput
|
|
|
|
|
|
class PoolingType(IntEnum):
|
|
"""Enumeration for different types of pooling methods."""
|
|
LAST = 0
|
|
ALL = 1
|
|
CLS = 2
|
|
|
|
|
|
class Pooler(nn.Module):
|
|
"""A layer that pools specific information from hidden states.
|
|
|
|
This layer does the following:
|
|
1. Extracts specific tokens or aggregates data based on pooling method.
|
|
2. Normalizes output if specified.
|
|
3. Returns structured results as `PoolerOutput`.
|
|
|
|
Attributes:
|
|
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
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states: torch.Tensor,
|
|
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 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:
|
|
offset = 0
|
|
pooled_data = []
|
|
for prompt_len in prompt_lens:
|
|
pooled_data.append(hidden_states[offset:offset + prompt_len])
|
|
offset += prompt_len
|
|
else:
|
|
raise ValueError(f"Invalid pooling type: {self.pooling_type}")
|
|
|
|
if self.normalize:
|
|
pooled_data = nn.functional.normalize(pooled_data, p=2, dim=1)
|
|
|
|
pooled_outputs = [
|
|
EmbeddingSequenceGroupOutput(data.tolist()) for data in pooled_data
|
|
]
|
|
|
|
return PoolerOutput(outputs=pooled_outputs)
|