chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

model: support Kimi-K2.5 (#17789)

Browse Source 
Co-authored-by: Mick <mickjagger19@icloud.com>
This commit is contained in:
Yuhao Yang
2026-01-27 10:57:00 +08:00
committed by GitHub
parent 0519b0935f
commit 479ab7a4e7
11 changed files with 1053 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
python/sglang/srt/configs/__init__.py
View File

@@ -9,6 +9,7 @@ from sglang.srt.configs.falcon_h1 import FalconH1Config
from sglang.srt.configs.janus_pro import MultiModalityConfig
from sglang.srt.configs.jet_nemotron import JetNemotronConfig
from sglang.srt.configs.jet_vlm import JetVLMConfig
from sglang.srt.configs.kimi_k25 import KimiK25Config
from sglang.srt.configs.kimi_linear import KimiLinearConfig
from sglang.srt.configs.kimi_vl import KimiVLConfig
from sglang.srt.configs.kimi_vl_moonvit import MoonViTConfig
@@ -39,6 +40,7 @@ __all__ = [
"Step3VisionEncoderConfig",
"Olmo3Config",
"KimiLinearConfig",
"KimiK25Config",
"Qwen3NextConfig",
"DotsVLMConfig",
"DotsOCRConfig",

171
python/sglang/srt/configs/kimi_k25.py Normal file
View File

@@ -0,0 +1,171 @@
"""
Kimi K25 Model Configuration.
"""
from transformers import DeepseekV3Config
from transformers.configuration_utils import PretrainedConfig
class KimiK25VisionConfig(PretrainedConfig):
"""Vision configuration for K2-VL (vision tower + mm projector).
Args:
Vision Tower Parameters:
patch_size: Patch size for vision tower.
init_pos_emb_height: Initial position embedding height.
init_pos_emb_width: Initial position embedding width.
init_pos_emb_time: Initial position embedding time dimension.
pos_emb_type: Type of position embedding.
num_attention_heads: Number of attention heads in vision tower.
num_hidden_layers: Number of hidden layers in vision tower.
hidden_size: Hidden size of vision tower.
intermediate_size: Intermediate size in vision tower FFN.
merge_kernel_size: Kernel size for spatial patch merging.
video_attn_type: Type of video attention.
merge_type: Type of merge operation.
MM Projector Parameters:
mm_projector_type: Type of multimodal projector.
mm_hidden_size: Hidden size for projector (defaults to hidden_size).
projector_hidden_act: Activation function for projector.
projector_ln_eps: Layer norm epsilon for projector.
"""
model_type = "kimi_k25"
def __init__(
self,
# Vision Tower
patch_size: int = 14,
init_pos_emb_height: int = 64,
init_pos_emb_width: int = 64,
init_pos_emb_time: int = 4,
pos_emb_type: str = "divided_fixed",
num_attention_heads: int = 16,
num_hidden_layers: int = 27,
hidden_size: int = 1152,
intermediate_size: int = 4304,
merge_kernel_size: tuple[int, int] = (2, 2),
video_attn_type: str = "spatial_temporal",
merge_type: str = "sd2_tpool",
# MM Projector
mm_projector_type: str = "patchmerger",
mm_hidden_size: int | None = None,
projector_hidden_act: str = "gelu",
projector_ln_eps: float = 1e-5,
text_hidden_size: int = 7168,
**kwargs,
):
super().__init__(**kwargs)
# Vision Tower
self.patch_size = patch_size
self.init_pos_emb_height = init_pos_emb_height
self.init_pos_emb_width = init_pos_emb_width
self.init_pos_emb_time = init_pos_emb_time
self.pos_emb_type = pos_emb_type
self.num_attention_heads = num_attention_heads
self.num_hidden_layers = num_hidden_layers
self.hidden_size = hidden_size
self.intermediate_size = intermediate_size
self.merge_kernel_size = merge_kernel_size
self.video_attn_type = video_attn_type
self.merge_type = merge_type
# MM Projector
self.mm_projector_type = mm_projector_type
if mm_hidden_size is not None:
self.mm_hidden_size = mm_hidden_size
else:
self.mm_hidden_size = hidden_size
self.projector_hidden_act = projector_hidden_act
self.projector_ln_eps = projector_ln_eps
self.text_hidden_size = text_hidden_size
class KimiK25Config(PretrainedConfig):
"""K2-VL model configuration.
K2-VL extends Kimi-VL with video support using video-chunks.
A video-chunk consists of multiple consecutive frames (default: 4)
that are processed together with temporal pooling.
Args:
text_config: Configuration for the text model (DeepseekV3).
Vision Tower Parameters:
patch_size: Patch size for vision tower.
init_pos_emb_height: Initial position embedding height.
init_pos_emb_width: Initial position embedding width.
init_pos_emb_time: Initial position embedding time dimension.
pos_emb_type: Type of position embedding.
vt_num_attention_heads: Number of attention heads in vision tower.
vt_num_hidden_layers: Number of hidden layers in vision tower.
vt_hidden_size: Hidden size of vision tower.
vt_intermediate_size: Intermediate size in vision tower FFN.
merge_kernel_size: Kernel size for spatial patch merging.
video_attn_type: Type of video attention.
merge_type: Type of merge operation.
Video-Chunk Parameters:
temporal_merge_kernel_size: Number of frames per video chunk.
Default is 4, meaning 4 frames are merged into 1 chunk.
sample_fps: Video sampling frame rate.
timestamp_mode: Format for chunk timestamps.
MM Projector Parameters:
mm_projector_type: Type of multimodal projector.
mm_hidden_size: Hidden size from vision tower.
projector_hidden_act: Activation function for projector.
projector_ln_eps: Layer norm epsilon for projector.
Other Parameters:
ignore_index: The ignore index for the loss function.
media_placeholder_token_id: The token ID for media placeholders.
pad_token_id: The token ID for padding.
"""
model_type = "kimi_k25"
def __init__(
self,
text_config: dict | DeepseekV3Config | None = None,
vision_config: dict | KimiK25VisionConfig | None = None,
# Other parameters
ignore_index: int = -100,
media_placeholder_token_id: int = 163605,
pad_token_id: int = 0,
use_unified_vision_chunk: bool = False,
video_placeholder: str = "<|kimi_k25_video_placeholder|>",
**kwargs,
):
if text_config is None:
text_config = DeepseekV3Config()
elif isinstance(text_config, dict):
text_config = DeepseekV3Config(**text_config)
if vision_config is None:
vision_config = KimiK25VisionConfig()
elif isinstance(vision_config, dict):
vision_config = KimiK25VisionConfig(**vision_config)
self.vision_config = vision_config
self.text_config = text_config
# Other config
self.ignore_index = ignore_index
self.media_placeholder_token_id = media_placeholder_token_id
self.use_unified_vision_chunk = use_unified_vision_chunk
self.video_placeholder = video_placeholder
# Propagate quantization config from text model
if getattr(self.text_config, "quantization_config", None) is not None:
self.quantization_config = self.text_config.quantization_config
super().__init__(pad_token_id=pad_token_id, **kwargs)
@property
def hidden_size(self) -> int:
"""Get hidden size from text config for compatibility."""
return self.text_config.hidden_size
@property
def vocab_size(self) -> int:
"""Get vocab size from text config for compatibility."""
return self.text_config.vocab_size

20
python/sglang/srt/configs/model_config.py
View File

@@ -391,16 +391,17 @@ class ModelConfig:
or "MistralLarge3ForCausalLM" in self.hf_config.architectures
or "PixtralForConditionalGeneration" in self.hf_config.architectures
or "MistralLarge3ForCausalLMEagle" in self.hf_config.architectures
or "KimiK25ForConditionalGeneration" in self.hf_config.architectures
):
self.head_dim = 256
self.attention_arch = AttentionArch.MLA
self.kv_lora_rank = self.hf_config.kv_lora_rank
self.qk_nope_head_dim = self.hf_config.qk_nope_head_dim
self.qk_rope_head_dim = self.hf_config.qk_rope_head_dim
self.v_head_dim = self.hf_config.v_head_dim
self.kv_lora_rank = self.hf_text_config.kv_lora_rank
self.qk_nope_head_dim = self.hf_text_config.qk_nope_head_dim
self.qk_rope_head_dim = self.hf_text_config.qk_rope_head_dim
self.v_head_dim = self.hf_text_config.v_head_dim
self.index_head_dim = (
get_nsa_index_head_dim(self.hf_config)
if is_deepseek_nsa(self.hf_config)
get_nsa_index_head_dim(self.hf_text_config)
if is_deepseek_nsa(self.hf_text_config)
else None
)
@@ -412,11 +413,11 @@ class ModelConfig:
self.scaling = 1 / math.sqrt(
self.qk_nope_head_dim + self.qk_rope_head_dim
)
if self.hf_config.rope_scaling:
mscale_all_dim = self.hf_config.rope_scaling.get(
if self.hf_text_config.rope_scaling:
mscale_all_dim = self.hf_text_config.rope_scaling.get(
"mscale_all_dim", False
)
scaling_factor = self.hf_config.rope_scaling["factor"]
scaling_factor = self.hf_text_config.rope_scaling["factor"]
mscale = yarn_get_mscale(scaling_factor, float(mscale_all_dim))
self.scaling = self.scaling * mscale * mscale
@@ -1169,6 +1170,7 @@ multimodal_model_archs = [
"PaddleOCRVLForConditionalGeneration",
"MiDashengLMModel",
"StepVLForConditionalGeneration",
"KimiK25ForConditionalGeneration",
]
if external_mm_model_arch := envs.SGLANG_EXTERNAL_MM_MODEL_ARCH.get():