chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Add Mistral Small 4 (Pixtral) support (#20708)

Browse Source 
Signed-off-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
Co-authored-by: Alex Nails <alexnails@radixark.ai>
Co-authored-by: Dimitrios Bariamis <12195802+dbari@users.noreply.github.com>
Co-authored-by: dbari <dbari@users.noreply.github.com>
This commit is contained in:
Xinyuan Tong
2026-03-18 21:15:32 +00:00
committed by GitHub
parent df1d046de2
commit 6b8a6545b2
18 changed files with 360 additions and 124 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/srt/configs/deepseek_ocr.py
View File

@@ -781,8 +781,8 @@ class DeepseekV2Config(PretrainedConfig):
class DeepseekVLV2Config(PretrainedConfig):
# model_type = "deepseek_vl_v2"
model_type = "deepseek-ocr"
vision_config: VisionEncoderConfig
projector_config: MlpProjectorConfig
vision_config: VisionEncoderConfig = None
projector_config: MlpProjectorConfig = None
tile_tag: str = "2D"
global_view_pos: str = "head"

6
python/sglang/srt/configs/deepseekvl2.py
View File

@@ -649,9 +649,9 @@ class DeepseekV2Config(PretrainedConfig):
class DeepseekVL2Config(PretrainedConfig):
model_type = "deepseek_vl_v2"
vision_config: DeepseekVL2VisionEncoderConfig
projector_config: DeepseekVL2MlpProjectorConfig
language_config: DeepseekV2Config
vision_config: DeepseekVL2VisionEncoderConfig = None
projector_config: DeepseekVL2MlpProjectorConfig = None
language_config: DeepseekV2Config = None
tile_tag: str = "2D"
global_view_pos: str = "head"

24
python/sglang/srt/configs/janus_pro.py
View File

@@ -123,14 +123,14 @@ class SigLIPVisionCfg:
class MultiModalityConfig(PretrainedConfig):
model_type = "multi_modality"
vision_config: VisionConfig
aligner_config: AlignerConfig
vision_config: VisionConfig = None
aligner_config: AlignerConfig = None
gen_vision_config: GenVisionConfig
gen_aligner_config: GenAlignerConfig
gen_head_config: GenHeadConfig
gen_vision_config: GenVisionConfig = None
gen_aligner_config: GenAlignerConfig = None
gen_head_config: GenHeadConfig = None
language_config: LlamaConfig
language_config: LlamaConfig = None
def __init__(self, **kwargs):
super().__init__(**kwargs)
@@ -595,12 +595,12 @@ class VLChatProcessor(ProcessorMixin):
class VLMImageProcessorConfig(PretrainedConfig):
model_type = "deepseek_vlm"
image_size: int
min_size: int
image_mean: Union[Tuple[float, float, float], List[float]]
image_std: Union[Tuple[float, float, float], List[float]]
rescale_factor: float
do_normalize: bool
image_size: int = None
min_size: int = None
image_mean: Union[Tuple[float, float, float], List[float]] = None
image_std: Union[Tuple[float, float, float], List[float]] = None
rescale_factor: float = None
do_normalize: bool = None
def __init__(
self,

24
python/sglang/srt/configs/jet_nemotron.py
View File

@@ -25,18 +25,18 @@ class JetBlockConfig:
class JetNemotronConfig(PretrainedConfig):
model_type: str = "jet_nemotron"
efficient_attention_config: dict[str, dict[str, Any]]
hidden_act: str
hidden_size: int
initializer_range: float
intermediate_size: int
layer_types: list[str]
max_position_embeddings: int
num_attention_heads: int
num_key_value_heads: int
rms_norm_eps: float
rope_scaling: None
rope_theta: float
efficient_attention_config: dict[str, dict[str, Any]] = None
hidden_act: str = None
hidden_size: int = None
initializer_range: float = None
intermediate_size: int = None
layer_types: list[str] = None
max_position_embeddings: int = None
num_attention_heads: int = None
num_key_value_heads: int = None
rms_norm_eps: float = None
rope_scaling: None = None
rope_theta: float = None
@property
def full_attention_layer_ids(self) -> list[int]: