[Model] Replace embedding models with pooling adapter (#10769)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2024-12-01 08:02:54 +08:00
committed by GitHub
parent 7e4bbda573
commit 133707123e
32 changed files with 383 additions and 319 deletions
@@ -4,6 +4,8 @@ Run `pytest tests/models/embedding/language/test_embedding.py`.
"""
import pytest
from vllm.config import PoolerConfig
from ..utils import check_embeddings_close
@@ -33,6 +35,9 @@ def test_models(
dtype: str,
) -> None:
vllm_extra_kwargs = {}
if model == "ssmits/Qwen2-7B-Instruct-embed-base":
vllm_extra_kwargs["override_pooler_config"] = \
PoolerConfig(pooling_type="MEAN")
if model == "Alibaba-NLP/gte-Qwen2-7B-instruct":
vllm_extra_kwargs["hf_overrides"] = {"is_causal": False}
+12 -15
View File
@@ -6,11 +6,8 @@ import torch.cuda
from vllm.model_executor.models import (is_embedding_model,
is_text_generation_model,
supports_multimodal)
# yapf conflicts with isort for this block
# yapf: disable
from vllm.model_executor.models.registry import (_CROSS_ENCODER_MODELS,
_EMBEDDING_MODELS,
_MULTIMODAL_MODELS,
from vllm.model_executor.models.adapters import as_embedding_model
from vllm.model_executor.models.registry import (_MULTIMODAL_MODELS,
_SPECULATIVE_DECODING_MODELS,
_TEXT_GENERATION_MODELS,
ModelRegistry)
@@ -26,18 +23,18 @@ def test_registry_imports(model_arch):
model_cls, _ = ModelRegistry.resolve_model_cls(model_arch)
if model_arch in _SPECULATIVE_DECODING_MODELS:
pass # Ignore these models which do not have a unified format
else:
assert is_text_generation_model(model_cls) is (
model_arch in _TEXT_GENERATION_MODELS
or model_arch in _MULTIMODAL_MODELS)
return # Ignore these models which do not have a unified format
embedding_models = {**_EMBEDDING_MODELS, **_CROSS_ENCODER_MODELS}
assert is_embedding_model(model_cls) is (model_arch
in embedding_models)
if (model_arch in _TEXT_GENERATION_MODELS
or model_arch in _MULTIMODAL_MODELS):
assert is_text_generation_model(model_cls)
assert supports_multimodal(model_cls) is (model_arch
in _MULTIMODAL_MODELS)
# All vLLM models should be convertible to an embedding model
embed_model = as_embedding_model(model_cls)
assert is_embedding_model(embed_model)
if model_arch in _MULTIMODAL_MODELS:
assert supports_multimodal(model_cls)
@fork_new_process_for_each_test