EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[Feature] add support kimi vl model (#5383)

Browse Source 
Co-authored-by: wenju.li <wenju.li@deepctr.cn>
This commit is contained in:
liwenju0
2025-04-30 12:31:19 +08:00
committed by GitHub
parent 403b855a22
commit 8fefdd32c7
13 changed files with 1189 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
python/sglang/srt/configs/model_config.py
View File

@@ -176,6 +176,13 @@ class ModelConfig:
self.attention_arch = AttentionArch.MLA
self.kv_lora_rank = self.hf_text_config.kv_lora_rank
self.qk_rope_head_dim = self.hf_text_config.qk_rope_head_dim
elif "KimiVLForConditionalGeneration" in self.hf_config.architectures:
self.head_dim = 256
self.attention_arch = AttentionArch.MLA
self.kv_lora_rank = self.hf_text_config.kv_lora_rank
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.qk_nope_head_dim = self.hf_text_config.qk_nope_head_dim
else:
self.attention_arch = AttentionArch.MHA
@@ -530,6 +537,7 @@ multimodal_model_archs = [
"Qwen2VLForConditionalGeneration",
"Qwen2_5_VLForConditionalGeneration",
"CLIPModel",
"KimiVLForConditionalGeneration",
]