[Bugfix] Fix Dense module loading for sentence-transformers embedding models (simplified V2) (#23408)

Signed-off-by: FFFfff1FFFfff <yifanli0919@gmail.com>
This commit is contained in:
LIYIFAN_liyifan
2025-08-25 05:39:24 +00:00
committed by GitHub
parent 787cdb3829
commit c9abb10489
5 changed files with 175 additions and 2 deletions
+11
View File
@@ -422,12 +422,23 @@ _ACTIVATION_REGISTRY = LazyDict({
lambda: nn.SiLU(),
"quick_gelu":
lambda: QuickGELU(),
"tanh":
lambda: nn.Tanh(),
"sigmoid":
lambda: nn.Sigmoid(),
})
def get_act_fn(act_fn_name: str) -> nn.Module:
"""Get an activation function by name."""
act_fn_name = act_fn_name.lower()
if act_fn_name.startswith("torch.nn.modules."):
activation_name = act_fn_name.split(".")[-1]
if activation_name == "identity":
return nn.Identity()
act_fn_name = activation_name
if act_fn_name not in _ACTIVATION_REGISTRY:
raise ValueError(
f"Activation function {act_fn_name!r} is not supported.")
+23 -1
View File
@@ -5,7 +5,7 @@ from collections.abc import Mapping, Set
from dataclasses import dataclass
from enum import IntEnum
from itertools import groupby
from typing import Callable, Optional, TypeVar, Union
from typing import Callable, Optional, TypeVar, Union, cast
import torch
import torch.nn as nn
@@ -435,9 +435,31 @@ class EmbeddingPoolerHead(PoolerHead):
def __init__(self) -> None:
super().__init__(activation=PoolerNormalize())
# Load ST projector if available
from vllm.config import get_current_vllm_config
from vllm.model_executor.models.adapters import _load_st_projector
vllm_config = get_current_vllm_config()
self.projector = _load_st_projector(
vllm_config.model_config) if vllm_config else None
def forward(self, pooled_data: Union[list[torch.Tensor], torch.Tensor],
pooling_metadata: PoolingMetadata):
# Apply ST projector
if self.projector is not None:
projector = cast(nn.Module, self.projector)
def _proj(x: torch.Tensor) -> torch.Tensor:
orig_dtype = x.dtype
y = projector(x.to(torch.float32))
return y.to(orig_dtype)
if isinstance(pooled_data, torch.Tensor):
pooled_data = _proj(pooled_data)
else:
pooled_data = [_proj(t) for t in pooled_data]
pooling_params = get_pooling_params(pooling_metadata)
if isinstance(pooled_data, list):