[Bugfix][Model] Reject encoder-backbone jina-embeddings-v5 checkpoints with a clear error (fixes #50337) (#50352)
Signed-off-by: seewoo <seewoo@ucsc.edu> Co-authored-by: Claude <noreply@anthropic.com>
This commit is contained in:
@@ -0,0 +1,64 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""Backbone validation for Jina Embeddings V5 (issue #50337).
|
||||
|
||||
The V5 family ships two backbones under one `architectures` entry: `-small` is
|
||||
a Qwen3 decoder (which vLLM implements), while `-nano` is a bidirectional
|
||||
EuroBERT encoder (which it does not). Upstream ships a separate
|
||||
`configuration_*.py` per repository, so the only signal distinguishing them is
|
||||
`is_decoder`, which the encoder variant sets to False.
|
||||
"""
|
||||
|
||||
from types import SimpleNamespace
|
||||
from typing import cast
|
||||
|
||||
import pytest
|
||||
from transformers import PretrainedConfig
|
||||
|
||||
from vllm.config import ModelConfig
|
||||
from vllm.model_executor.models.config import (
|
||||
MODELS_CONFIG_MAP,
|
||||
JinaEmbeddingsV5ModelConfig,
|
||||
)
|
||||
|
||||
|
||||
def _model_config(hf_config: PretrainedConfig) -> ModelConfig:
|
||||
"""Minimal stand-in for ModelConfig; only hf_config is read."""
|
||||
return cast(ModelConfig, SimpleNamespace(hf_config=hf_config))
|
||||
|
||||
|
||||
@pytest.mark.cpu_test
|
||||
def test_registered_for_the_architecture():
|
||||
"""The handler only runs if it is wired to the architecture name."""
|
||||
assert MODELS_CONFIG_MAP["JinaEmbeddingsV5Model"] is JinaEmbeddingsV5ModelConfig
|
||||
|
||||
|
||||
@pytest.mark.cpu_test
|
||||
def test_encoder_backbone_is_rejected():
|
||||
"""An encoder checkpoint must fail fast, naming the reason.
|
||||
|
||||
Without this, loading proceeds until every `q_norm`/`k_norm` weight is
|
||||
reported as uninitialized, which does not point at the actual problem.
|
||||
"""
|
||||
model_config = _model_config(PretrainedConfig(is_decoder=False))
|
||||
|
||||
with pytest.raises(NotImplementedError, match="is_decoder=False"):
|
||||
JinaEmbeddingsV5ModelConfig.verify_and_update_model_config(model_config)
|
||||
|
||||
|
||||
@pytest.mark.cpu_test
|
||||
def test_supported_decoder_backbone_is_accepted():
|
||||
"""The Qwen3-based variants must keep loading.
|
||||
|
||||
`-small` omits `is_decoder` entirely, so an absent attribute has to be
|
||||
treated as a decoder. The first assertion pins that assumption: if
|
||||
PretrainedConfig ever gains an `is_decoder=False` default, this fails here
|
||||
rather than silently rejecting a supported checkpoint.
|
||||
"""
|
||||
absent = PretrainedConfig()
|
||||
assert not hasattr(absent, "is_decoder")
|
||||
|
||||
JinaEmbeddingsV5ModelConfig.verify_and_update_model_config(_model_config(absent))
|
||||
JinaEmbeddingsV5ModelConfig.verify_and_update_model_config(
|
||||
_model_config(PretrainedConfig(is_decoder=True))
|
||||
)
|
||||
@@ -474,6 +474,37 @@ class JambaForSequenceClassificationConfig(VerifyAndUpdateConfig):
|
||||
pooler_config.use_activation = False
|
||||
|
||||
|
||||
class JinaEmbeddingsV5ModelConfig(VerifyAndUpdateConfig):
|
||||
"""Config handler for Jina Embeddings V5 embedding models."""
|
||||
|
||||
@staticmethod
|
||||
def verify_and_update_model_config(model_config: "ModelConfig") -> None:
|
||||
"""Reject checkpoints whose backbone is not the Qwen3 decoder.
|
||||
|
||||
The V5 family ships more than one backbone under a single
|
||||
`architectures` entry: the `-small` variants are Qwen3 decoders, while
|
||||
`-nano` is a bidirectional EuroBERT encoder. Upstream ships a separate
|
||||
`configuration_*.py` per repository, so the config carries no backbone
|
||||
field and the encoder variants are only identifiable by
|
||||
`is_decoder=False`. `JinaEmbeddingsV5Model` extends `Qwen3ForCausalLM`
|
||||
and has no path to an encoder backbone, so without this check such a
|
||||
checkpoint fails later with every `q_norm` and `k_norm` weight reported
|
||||
as uninitialized, which does not point at the actual problem.
|
||||
|
||||
Raises:
|
||||
NotImplementedError: If the checkpoint uses an encoder backbone.
|
||||
"""
|
||||
if getattr(model_config.hf_config, "is_decoder", True):
|
||||
return
|
||||
|
||||
raise NotImplementedError(
|
||||
"This jina-embeddings-v5 checkpoint uses a bidirectional encoder "
|
||||
"backbone (is_decoder=False). JinaEmbeddingsV5Model is built on "
|
||||
"Qwen3 and cannot serve it, so only the Qwen3-based decoder "
|
||||
"variants are supported, such as jina-embeddings-v5-text-small."
|
||||
)
|
||||
|
||||
|
||||
class JinaForRankingConfig(VerifyAndUpdateConfig):
|
||||
@staticmethod
|
||||
def verify_and_update_model_config(model_config: "ModelConfig") -> None:
|
||||
@@ -915,6 +946,7 @@ MODELS_CONFIG_MAP: dict[str, type[VerifyAndUpdateConfig]] = {
|
||||
"GteNewForSequenceClassification": GteNewModelConfig,
|
||||
"GteNewModel": GteNewModelConfig,
|
||||
"JambaForSequenceClassification": JambaForSequenceClassificationConfig,
|
||||
"JinaEmbeddingsV5Model": JinaEmbeddingsV5ModelConfig,
|
||||
"JinaForRanking": JinaForRankingConfig,
|
||||
"JinaVLForRanking": JinaVLForSequenceClassificationConfig,
|
||||
"KimiK3ForConditionalGeneration": KimiK3ForConditionalGenerationConfig,
|
||||
|
||||
Reference in New Issue
Block a user