DeepSeek-V3#
DeepSeek-V3 is a large-scale Mixture-of-Experts model with 671B total parameters and 37B activated per token. It features Multi-head Latent Attention (MLA), innovative load balancing, and Multi-Token Prediction (MTP). DeepSeek-V3.2 is an updated release with further improvements.
Moonlight by Moonshot AI also uses this architecture with 16B total / 3B activated parameters.
Task |
Text Generation (MoE) |
Architecture |
|
Parameters |
671B total / 37B active |
HF Org |
Available Models#
DeepSeek-V3: 671B total, 37B activated
DeepSeek-V3.2 (
DeepseekV32ForCausalLM): updated architectureMoonlight-16B-A3B (Moonshot AI): 16B total, 3B activated
Architectures#
DeepseekV3ForCausalLMDeepseekV32ForCausalLM
Example HF Models#
Model |
HF ID |
|---|---|
DeepSeek-V3 |
|
DeepSeek-V3-Base |
|
DeepSeek-V3.2 |
|
Moonlight 16B A3B |
Example Recipes#
Recipe |
Description |
|---|---|
SFT — DeepSeek-V3.2 on HellaSwag with pipeline parallelism |
|
SFT — Moonlight 16B with Transformer Engine |
|
SFT — Moonlight 16B with packed sequences |
Try with NeMo AutoModel#
1. Install (full instructions):
pip install nemo-automodel
2. Clone the repo to get the example recipes:
git clone https://github.com/NVIDIA-NeMo/Automodel.git
cd Automodel
Note
This recipe was validated on 32 nodes × 8 GPUs (256 H100s). See the Launcher Guide for multi-node setup.
3. Run the recipe from inside the repo:
automodel --nproc-per-node=8 examples/llm_finetune/deepseek_v32/deepseek_v32_hellaswag_pp.yaml
Run with Docker
1. Pull the container and mount a checkpoint directory:
docker run --gpus all -it --rm \
--shm-size=8g \
-v $(pwd)/checkpoints:/opt/Automodel/checkpoints \
nvcr.io/nvidia/nemo-automodel:26.02.00
2. Navigate to the AutoModel directory (where the recipes are):
cd /opt/Automodel
3. Run the recipe:
automodel --nproc-per-node=8 examples/llm_finetune/deepseek_v32/deepseek_v32_hellaswag_pp.yaml
See the Installation Guide and LLM Fine-Tuning Guide.
Fine-Tuning#
See the LLM Fine-Tuning Guide and the Large MoE Fine-Tuning Guide.