[Model] Add native Dots3 NOTE multimodal support (#51255)

Signed-off-by: KurodaKanbei <mistergalahad@gmail.com>
Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>
Co-authored-by: OpenAI Codex <codex@openai.com>
Co-authored-by: Jee Jee Li <jeejeelee@inferact.ai>
Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>
This commit is contained in:
范裕达
2026-08-12 23:13:20 +08:00
committed by GitHub
co-authored by OpenAI Codex Jee Jee Li Jee Jee Li
parent b1b752042f
commit 9035151d6c
28 changed files with 6468 additions and 7 deletions
+1
View File
@@ -87,6 +87,7 @@ _CONFIG_REGISTRY: dict[str, type[PretrainedConfig]] = LazyConfigDict(
deepseek_vl_v2="DeepseekVLV2Config",
deepseek_v32="DeepseekV3Config",
deepseek_v4="DeepseekV4Config",
dots3_note="Dots3NoteConfig",
k3_dspark="K3DSparkConfig",
flex_olmo="FlexOlmoConfig",
fireredlid="FireRedLIDConfig",
@@ -35,6 +35,7 @@ _CLASS_TO_MODULE: dict[str, str] = {
"DiffusionGemmaTextConfig": "vllm.transformers_utils.configs.diffusion_gemma",
"DeepseekVLV2Config": "vllm.transformers_utils.configs.deepseek_vl2",
"DeepseekV4Config": "vllm.transformers_utils.configs.deepseek_v4",
"Dots3NoteConfig": "vllm.transformers_utils.configs.dots3_note",
"K3DSparkConfig": "vllm.transformers_utils.configs.k3_dspark",
"DotsOCRConfig": "vllm.transformers_utils.configs.dotsocr",
"EAGLEConfig": "vllm.transformers_utils.configs.eagle",
@@ -127,6 +128,7 @@ __all__ = [
"DeepseekVLV2Config",
"DeepseekV3Config",
"DeepseekV4Config",
"Dots3NoteConfig",
"K3DSparkConfig",
"DotsOCRConfig",
"EAGLEConfig",
@@ -0,0 +1,25 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from transformers import DeepseekV3Config
class Dots3NoteConfig(DeepseekV3Config):
"""Configuration for the Dots3Note hybrid DSA/SWA model."""
model_type = "dots3_note"
def __init__(self, **kwargs):
# These fields are absent from existing Note config.json files. Do
# not inherit DeepSeek-V3's 8-group/4-group router defaults: Note was
# trained with an ungrouped (1/1) noaux_tc router, as defined by the
# reference Dots3Config in SGLang haimain. A different grouping
# changes the selected experts at every MoE layer.
kwargs.setdefault("n_group", 1)
kwargs.setdefault("topk_group", 1)
# Dots3 projects the indexer RoPE coordinates in adjacent (GPT-J)
# pairs. DeepSeek-V3.2 defaults to the split-half NeoX layout when
# this flag is absent, which rotates different learned coordinates.
kwargs.setdefault("indexer_rope_interleave", True)
kwargs.setdefault("num_nextn_predict_layers", 1)
super().__init__(**kwargs)
@@ -317,6 +317,7 @@ class ModelArchConfigConvertorBase:
"deepseek_v3",
"deepseek_v32",
"deepseek_v4",
"dots3_note",
"deepseek_mtp",
"k3_dspark",
"glm_moe_dsa",