[new model] add COLQwen3 code & Inference (#34398)
Signed-off-by: craftsangjae <craftsangjae@gmail.com> Signed-off-by: katacoder <craftsangjae@gmail.com>
This commit is contained in:
@@ -74,6 +74,8 @@ _CONFIG_REGISTRY: dict[str, type[PretrainedConfig]] = LazyConfigDict(
|
||||
afmoe="AfmoeConfig",
|
||||
bagel="BagelConfig",
|
||||
chatglm="ChatGLMConfig",
|
||||
colqwen3="ColQwen3Config",
|
||||
ops_colqwen3="OpsColQwen3Config",
|
||||
deepseek_vl_v2="DeepseekVLV2Config",
|
||||
deepseek_v32="DeepseekV3Config",
|
||||
flex_olmo="FlexOlmoConfig",
|
||||
|
||||
@@ -18,6 +18,9 @@ _CLASS_TO_MODULE: dict[str, str] = {
|
||||
"AfmoeConfig": "vllm.transformers_utils.configs.afmoe",
|
||||
"BagelConfig": "vllm.transformers_utils.configs.bagel",
|
||||
"ChatGLMConfig": "vllm.transformers_utils.configs.chatglm",
|
||||
"ColQwen3Config": "vllm.transformers_utils.configs.colqwen3",
|
||||
"OpsColQwen3Config": "vllm.transformers_utils.configs.colqwen3",
|
||||
"Qwen3VLNemotronEmbedConfig": "vllm.transformers_utils.configs.colqwen3",
|
||||
"DeepseekVLV2Config": "vllm.transformers_utils.configs.deepseek_vl2",
|
||||
"DotsOCRConfig": "vllm.transformers_utils.configs.dotsocr",
|
||||
"EAGLEConfig": "vllm.transformers_utils.configs.eagle",
|
||||
@@ -68,6 +71,9 @@ __all__ = [
|
||||
"AfmoeConfig",
|
||||
"BagelConfig",
|
||||
"ChatGLMConfig",
|
||||
"ColQwen3Config",
|
||||
"OpsColQwen3Config",
|
||||
"Qwen3VLNemotronEmbedConfig",
|
||||
"DeepseekVLV2Config",
|
||||
"DeepseekV3Config",
|
||||
"DotsOCRConfig",
|
||||
|
||||
@@ -0,0 +1,58 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
ColQwen3 configuration that extends Qwen3VLConfig with embedding projection
|
||||
fields. This allows ColQwen3 models to be loaded without trust_remote_code
|
||||
by mapping their custom model_type (colqwen3, ops_colqwen3, etc.) to a
|
||||
standard config class that vLLM understands.
|
||||
|
||||
Supported model_types:
|
||||
- colqwen3 (TomoroAI/tomoro-colqwen3-embed-8b)
|
||||
- ops_colqwen3 (OpenSearch-AI/Ops-Colqwen3-4B)
|
||||
- qwen3_vl_nemotron_embed (nvidia/nemotron-colembed-vl-8b-v2)
|
||||
"""
|
||||
|
||||
from transformers.models.qwen3_vl.configuration_qwen3_vl import Qwen3VLConfig
|
||||
|
||||
|
||||
class ColQwen3Config(Qwen3VLConfig):
|
||||
"""Configuration class for ColQwen3 models.
|
||||
|
||||
Extends Qwen3VLConfig with additional fields used by ColQwen3 variants
|
||||
for the embedding projection layer.
|
||||
"""
|
||||
|
||||
# Accept any ColQwen3 variant model_type
|
||||
model_type = "colqwen3"
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
embed_dim: int | None = None,
|
||||
dims: int | None = None,
|
||||
dim: int | None = None,
|
||||
projection_dim: int | None = None,
|
||||
colbert_dim: int | None = None,
|
||||
pooling: str | None = None,
|
||||
**kwargs,
|
||||
):
|
||||
# Store embedding projection config fields
|
||||
self.embed_dim = embed_dim
|
||||
self.dims = dims
|
||||
self.dim = dim
|
||||
self.projection_dim = projection_dim
|
||||
self.colbert_dim = colbert_dim
|
||||
self.pooling = pooling
|
||||
|
||||
super().__init__(**kwargs)
|
||||
|
||||
|
||||
class OpsColQwen3Config(ColQwen3Config):
|
||||
"""Configuration for OpenSearch-AI ColQwen3 variants."""
|
||||
|
||||
model_type = "ops_colqwen3"
|
||||
|
||||
|
||||
class Qwen3VLNemotronEmbedConfig(ColQwen3Config):
|
||||
"""Configuration for NVIDIA Nemotron ColEmbed variants."""
|
||||
|
||||
model_type = "qwen3_vl_nemotron_embed"
|
||||
Reference in New Issue
Block a user