BREAKING
NVIDIA Releases NeMo AutoModel
0
x
min speedup
0
x
max speedup
Day-0 HF Support, No Conversion
1
Load HF model
↓
2
Apply EP + DeepEP + TE
↓
3
Scale to thousands of GPUs
Throughput Benchmarks
DeepSeek V3 671B
1002
Qwen3 MoE 30B
12040
GPT-OSS 20B
13058
A Hybrid Approach
HF + FSDP
Easy
●
Simple to use
●
Limited optimization
NeMo AutoModel
Hybrid
●
Day-0 HF support
●
NVIDIA kernels
Low Barrier, Real Scalability
AI NEWS BLITZ
NVIDIA has launched NeMo AutoModel, an open-source training library for MoE models.