SGLang - Nemotron
Migrated Llama3.3-Nemontron-Super-49B support from vLLM to SGLang and submitted PR for it
0 Support in vLLM
Deci team recently added Nemotron reasoning model support for the super 48B model. This model is based on DeciLMForCausalLM
class. This is prunning from Llama 3.3 model which NAS(Network Architecture Search), so it is different from Nano version which is still LlamaCausal LM.
1 Adding model support in SGLang
- SGLang dev was recommended in the container environment based on
lmsysorg/sglang:dev
. The structure of the container is as following/sgl-workspace/sglang # source code dir /sgl-workspace/work-dir # Mounted for test code /root/.cach/huggingface # Mounted for model cache
- The model directory is
/sglang/python/sglang/srt/models
- The migration guide is listed here