[Doc] Update V1 status of various pooling models (#23189)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-08-20 10:33:41 +08:00
committed by GitHub
parent e58c5a9768
commit 64ab3c7253
4 changed files with 28 additions and 24 deletions
+3 -3
View File
@@ -20,7 +20,7 @@ from vllm.sequence import PoolerOutput
from vllm.tasks import PoolingTask
from vllm.transformers_utils.tokenizer import cached_tokenizer_from_config
from .interfaces import SupportsV0Only
from .interfaces import default_pooling_type
logger = init_logger(__name__)
@@ -215,7 +215,8 @@ class GritLMPooler(Pooler):
return build_output(pooled_data)
class GritLM(LlamaForCausalLM, SupportsV0Only):
@default_pooling_type("MEAN")
class GritLM(LlamaForCausalLM):
"""This class implements the embedding model for parasail-ai/GritLM-7B-vllm.
The class inherits from LlamaForCausalLM and provides a custom pooling
@@ -241,7 +242,6 @@ class GritLM(LlamaForCausalLM, SupportsV0Only):
prefix: str = "",
**kwargs,
) -> None:
# Use full attention for pooling (this is why V1 is not supported yet)
if vllm_config.model_config.runner_type == "pooling":
hf_config = vllm_config.model_config.hf_config
hf_config.is_causal = False
+7 -4
View File
@@ -3,7 +3,7 @@
from collections.abc import Iterable, Mapping, MutableSequence
from typing import (TYPE_CHECKING, ClassVar, Literal, Optional, Protocol,
Union, overload, runtime_checkable)
TypeVar, Union, overload, runtime_checkable)
import numpy as np
import torch
@@ -641,11 +641,14 @@ def supports_cross_encoding(
return is_pooling_model(model) and _supports_cross_encoding(model)
def default_pooling_type(pooling_type: str) -> object:
_T = TypeVar("_T", bound=type[torch.nn.Module])
def default_pooling_type(pooling_type: str):
"""Set default_pooling_type decorator. """
def func(model: object):
model.default_pooling_type = pooling_type
def func(model: _T) -> _T:
model.default_pooling_type = pooling_type # type: ignore
return model
return func