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

Unify the model type checking (#1905)

Browse Source
This commit is contained in:
Lianmin Zheng
2024-11-03 12:25:39 -08:00
committed by GitHub
parent c17c578108
commit 0abbf289a8
13 changed files with 146 additions and 160 deletions
Download Patch File Download Diff File Expand all files Collapse all files

50
python/sglang/srt/utils.py
View File

@@ -204,56 +204,6 @@ def is_port_available(port):
return False
def is_multimodal_model(model_architectures):
if (
"LlavaLlamaForCausalLM" in model_architectures
or "LlavaQwenForCausalLM" in model_architectures
or "LlavaMistralForCausalLM" in model_architectures
or "LlavaVidForCausalLM" in model_architectures
or "MllamaForConditionalGeneration" in model_architectures
or "Qwen2VLForConditionalGeneration" in model_architectures
):
return True
else:
return False
def is_attention_free_model(model_architectures):
return False
def model_has_inner_state(model_architectures):
return False
def is_embedding_model(model_architectures):
if (
"LlamaEmbeddingModel" in model_architectures
or "MistralModel" in model_architectures
or "LlamaForSequenceClassification" in model_architectures
or "LlamaForSequenceClassificationWithNormal_Weights" in model_architectures
):
return True
else:
return False
def is_generation_model(model_architectures, is_embedding: bool = False):
# We have two ways to determine whether a model is a generative model.
# 1. Check the model architectue
# 2. check the `is_embedding` server args
if (
"LlamaEmbeddingModel" in model_architectures
or "MistralModel" in model_architectures
or "LlamaForSequenceClassification" in model_architectures
or "LlamaForSequenceClassificationWithNormal_Weights" in model_architectures
):
return False
else:
return not is_embedding
def decode_video_base64(video_base64):
from PIL import Image