51 lines
1.6 KiB
Python
51 lines
1.6 KiB
Python
# adapted from
|
|
# https://github.com/vllm-project/vllm/blob/82a1b1a82b1fbb454c82a9ef95730b929c9b270c/vllm/model_executor/layers/pooler.py
|
|
|
|
from dataclasses import dataclass
|
|
from enum import IntEnum
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
|
|
from sglang.srt.model_executor.model_runner import ForwardBatch
|
|
|
|
|
|
class PoolingType(IntEnum):
|
|
LAST = 0
|
|
|
|
|
|
@dataclass
|
|
class EmbeddingPoolerOutput:
|
|
embeddings: torch.Tensor
|
|
|
|
|
|
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, AVERAGE, MAX).
|
|
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, forward_batch: ForwardBatch
|
|
) -> EmbeddingPoolerOutput:
|
|
if self.pooling_type == PoolingType.LAST:
|
|
last_token_indices = torch.cumsum(forward_batch.extend_seq_lens, dim=0) - 1
|
|
pooled_data = hidden_states[last_token_indices]
|
|
else:
|
|
raise ValueError(f"Invalid pooling type: {self.pooling_type}")
|
|
|
|
if self.normalize:
|
|
pooled_data = nn.functional.normalize(pooled_data, p=2, dim=1)
|
|
|
|
return EmbeddingPoolerOutput(embeddings=pooled_data)
|