biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Model] Classification models support logit_bias / sigmoid_normalize (#24031)

Browse Source 
Signed-off-by: wang.yuqi <noooop@126.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
This commit is contained in:
wang.yuqi
2025-09-03 00:48:57 +08:00
committed by GitHub
parent 38ba061f6f
commit e0653f6c0b
4 changed files with 38 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

45
vllm/config/__init__.py
View File

@@ -2651,24 +2651,46 @@ class PoolerConfig:
## for embeddings models
normalize: Optional[bool] = None
"""
Whether to normalize the embeddings outputs.
Whether to normalize the embeddings outputs. Defaults to True.
"""
dimensions: Optional[int] = None
"""
Reduce the dimensions of embeddings if model
support matryoshka representation.
support matryoshka representation. Defaults to None.
"""
enable_chunked_processing: Optional[bool] = None
"""
Whether to enable chunked processing for long inputs that exceed the model's
maximum position embeddings. When enabled, long inputs will be split into
chunks, processed separately, and then aggregated using weighted averaging.
This allows embedding models to handle arbitrarily long text without CUDA
errors. Defaults to False.
"""
max_embed_len: Optional[int] = None
"""
Maximum input length allowed for embedding generation. When set, allows
inputs longer than max_embed_len to be accepted for embedding models.
When an input exceeds max_embed_len, it will be handled according to
the original max_model_len validation logic.
Defaults to None (i.e. set to max_model_len).
"""
## for classification models
activation: Optional[bool] = None
"""
Whether to apply activation function to the classification outputs.
Defaults to True.
"""
logit_bias: Optional[float] = None
"""
If provided, apply classification logit biases. Defaults to None.
"""
## for reward models
softmax: Optional[bool] = None
"""
Whether to apply softmax to the reward outputs.
Defaults to True.
"""
step_tag_id: Optional[int] = None
"""
@@ -2683,25 +2705,6 @@ class PoolerConfig:
``math-shepherd-mistral-7b-prm`` model.
"""
enable_chunked_processing: Optional[bool] = None
"""
Whether to enable chunked processing for long inputs that exceed the model's
maximum position embeddings. When enabled, long inputs will be split into
chunks, processed separately, and then aggregated using weighted averaging.
This allows embedding models to handle arbitrarily long text without CUDA
errors. Defaults to False.
"""
max_embed_len: Optional[int] = None
"""
Maximum input length allowed for embedding generation. When set, allows
inputs longer than max_embed_len to be accepted for embedding models.
This parameter enables accepting long inputs without requiring
VLLM_ALLOW_LONG_MAX_MODEL_LEN environment variable. When an input exceeds
max_embed_len, it will be handled according to the original max_model_len
validation logic. Defaults to None (i.e. set to max_model_len).
"""
def compute_hash(self) -> str:
"""
WARNING: Whenever a new field is added to this config,