graph TD
A[Hugging Face 模型与 Tokenizer] --> B[模型适配层：结构自动识别]
B --> C[并行 Recipe 解析器]
C --> D{"并行策略组合"}
D -->|"数据并行"| E[FSDP2 fully_shard]
D -->|"张量并行"| F[DTensor 按维切分]
D -->|"流水线并行"| G[层块划分与调度]
D -->|"上下文并行"| H[序列维度切分]
E --> I[DeviceMesh 多维拓扑]
F --> I
G --> I
H --> I
I --> J[统一训练循环与优化器]
J --> K[分布式检查点]
K --> L[回写 HF 标准权重格式]
技术难点集中在两处。一是异构并行的组合正确性：TP 与 FSDP 叠加时，参数的分片顺序、梯度 reduce 的时机、optimizer state 的切分方式必须严格对齐，否则会出现静默的数值错误而非显式崩溃，调试成本极高。二是检查点格式的双向转换：DTensor 分片后的 state_dict 与 Hugging Face 的标准权重之间需要一套可逆的映射，还要支持在训练中途改变并行度（resharding）。这两点恰是很多“自研 FSDP 封装”最终放弃的地方。
[11/04/2026]REINFORCE-Zero We added a train-inference consistency skill: the trainer calls the rollout engine’s own kernels, so its logprobs match vLLM’s bitwise and the train-inference mismatch is exactly 0 (mismatch 0, vllm_kl 0). labs-molt has the end-to-end implementation and an example. It follows REINFORCE-Zero by @hijkzzz: with aligned kernels, asynchronous RL stays stable with plain importance sampling, no clipping.
[11/01/2026]Muown optimizer We now support the Muown optimizer for dense and MoE models with FSDP2 and expert parallelism. Check out the Qwen3-30B-A3B recipe.
[10/21/2026]EAGLE-4 We now support training EAGLE-4 drafters: the EAGLE-3.1 draft with a sliding-window layer and a low-rank candidate head over the full vocabulary, trained in two stages and exported to the LlamaForCausalLMEagle4 serving layout, many thanks to @Khazic. Check out the Qwen3-8B draft and candidate-head recipes and the speculative decoding overview.
[10/16/2026]Beam-1.0-501B-A23B We now support LoRA and full-parameter fine-tuning for Reflection AI’s 501B-A23B sparse MoE with mixed sliding-window and full attention, using FSDP2, expert parallelism, and packed sequences. Check out the Tulu3 LoRA EP8 and Tulu3 SFT EP64 8-node recipes and the model coverage page.
[10/12/2026]FLUX 3 Image Klein Base We added text-to-image and image-editing LoRA fine-tuning for Black Forest Labs’ base image model. See the T2I LoRA recipe and model guide for checkpoint setup and the 100-step validation scope.
[10/04/2026]HunyuanImage-3.0 We now support text-to-image full fine-tuning and LoRA for Tencent’s 80B-A13B MoE multimodal decoder with FSDP2 and expert parallelism, many thanks to @Khazic. Check out the full fine-tuning and LoRA recipes and the model coverage page.
[09/28/2026]Qwen-Image-2.1 We now support text-to-image full fine-tuning and LoRA for Qwen’s 7B single-stream DiT conditioned on Qwen3-VL. Check out the full fine-tuning and LoRA recipes and the model coverage page.
[09/26/2026]MiMo-V2.5-Pro We now support full-parameter fine-tuning for Xiaomi’s hybrid full and sliding-window attention Mixture-of-Experts model from its FP8 checkpoint. Check out the HellaSwag PP18/EP8 recipe and model coverage page.
[08/26/2026]Qwen3.8-Flash-Next We now support full-parameter fine-tuning for the 180B hybrid-attention Engram MoE with FlexAttention sparse QSA, packed (THD) sequences, and model-owned context parallelism. Check out the HellaSwag EP64 recipe and model coverage page.
[08/21/2026]DFlash 2 We now support training DFlash 2 drafters: the DFlash block-diffusion backbone plus a two-tap dynamic convolution and a pairwise path selector for higher acceptance lengths. Check out the Qwen3-8B recipe and the speculative decoding overview.
[08/10/2026]MuseGlimmer We now support SFT/LoRA the dense 30B MuseGlimmer vision-language model, including TP/CP packed sequence recipes. Check out the recipes and model coverage page.
[07/29/2026]Kimi K3 We now support full-parameter fine-tuning for Moonshot AI’s 2.8T-parameter MoE model on NVIDIA GB200. Check out the HellaSwag EP32/PP8 recipe on 256 GB200.
[06/27/2026] Speculative decoding in NeMo AutoModel Train target-aligned drafters end to end: EAGLE-1/2/3, P-EAGLE, DFlash, and DeepSeek’s newly released DSpark. Huge thanks to @Khazic for the speculative decoding stack and @kashif for DSpark support.
[06/21/2026]GLM-5.2 We now support finetuning zai-org/GLM-5.2 with IndexShare DSA, optional TileLang sparse kernels, and long-context CP recipes. Check out our 32K long-context recipe and model coverage page.
[06/15/2026]DiffusionGemma. We now support fine-tuning the google/diffusiongemma-26B-A4B-it model. Check out our recipe and guide.
[06/12/2026]MiniMax M3 We now support finetuning MiniMax’s MiniMax-M3. Check out our recipe and guide.
[06/04/2026]Nemotron-3 Ultra We now support finetuning NVIDIA’s Nemotron 3 Ultra 550B A55B. Check out our recipe and guide.
[06/03/2026]Gemma 4 12B We now support finetuning the dense google/gemma-4-12B model. Check out our recipe.
[05/27/2026]Step-3.7-Flash We added model coverage for Stepfun AI’s 198B-A13B MoE vision-language model, targeting image/video agentic developer workflows with a 256k context language backbone and 1.8B ViT vision tower. See the model coverage page.
[05/19/2026]Ling 2.0 We now support finetuning the inclusionAI Ling 2.0 MoE family (inclusionAI/Ling-mini-2.0, inclusionAI/Ling-flash-2.0, and inclusionAI/Ling-1T), thanks to @Hayden727. Check out our recipes.
[05/17/2026]ERNIE 4.5 and MiMo-V2-Flash We now support finetuning baidu/ERNIE-4.5-0.3B-PT, baidu/ERNIE-4.5-21B-A3B-PT, and XiaomiMiMo/MiMo-V2-Flash. Check out our ERNIE dense recipe, ERNIE MoE recipe, and MiMo recipe.
[04/29/2026]Mistral Medium 3.5 We now support finetuning Mistral AI’s 128B FP8-native VLM Mistral Medium 3.5. Check out our recipe and guide.
[04/28/2026]Nemotron-3-Nano-Omni We now support finetuning nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16, NVIDIA’s 30B-A3B omnimodal MoE (text · image · audio) with NemotronH hybrid Mamba+Attention backbone. Check out our SFT recipe, LoRA recipe, and guide.
[04/28/2026]Hy3-preview We now support finetuning tencent/Hy3-preview, thanks to @Khazic. Check out our recipe.
[04/16/2026]Qwen3.6 MoE We now support finetuning Qwen/Qwen3.6-35B-A3B. Check out our recipe.
[04/16/2026]LLaVA-OneVision-1.5 We now support finetuning lmms-lab/LLaVA-OneVision-1.5-4B-Instruct, thanks to @vgauraha62. Check out our recipe.
[04/12/2026]MiniMax-M2.7 We now support finetuning MiniMaxAI/MiniMax-M2.7. Check out our recipe.
[04/07/2026]GLM-5.1 We now support finetuning zai-org/GLM-5.1. GLM-5.1 is Zhipu AI’s latest open-source MoE model featuring MLA + DeepSeek Sparse Attention. Check out our recipe and discussion.
[04/02/2026]Gemma 4 We support fine-tuning for Gemma4 (2B, 4B, 31B, 26BA4B)! Check out our recipes.
[03/30/2026]NeMo AutoModel ships with agent-friendly skills in skills/ to help you with common development tasks (e.g., running a recipe, model onboarding, development) across the repo. We welcome PRs that improve existing skills or add new ones.
[03/16/2026]Mistral Small 4 We support fine-tuning for Mistral4 119B! Check out our recipe.
[03/11/2026]Nemotron Super v3 We support fine-tuning for nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16. Check out our recipe.
[03/11/2026]GLM-5 We now support finetuning zai-org/GLM-5. Check out our recipe.
Nemo AutoModel is a Pytorch DTensor‑native SPMD open-source training library under NVIDIA NeMo Framework, designed to streamline and scale training and finetuning for LLMs, VLMs, diffusion models, and retrieval models. Designed for flexibility, reproducibility, and scale, NeMo AutoModel enables both small-scale experiments and massive multi-GPU, multi-node deployments for fast experimentation in research and production environments.
What you can expect:
Hackable with a modular design that allows easy integration, customization, and quick research prototypes.
Minimal ceremony: YAML-driven recipes; override any field using CLI.
High performance and flexibility with custom kernels and DTensor support.
Seamless integration with Hugging Face for day-0 model support, ease of use, and wide range of supported models.
Efficient resource management using Kubernetes and Slurm, enabling scalable and flexible deployment across configurations.
Documentation with step-by-step guides and runnable examples.
Why PyTorch Distributed and SPMD
One program, any scale: The same training script runs on 1 GPU or 1000+ by changing the mesh.
PyTorch Distributed native: Partition model/optimizer states with DeviceMesh + placements (Shard, Replicate).
SPMD first: Parallelism is configuration. No model rewrites when scaling up or changing strategy.
Decoupled concerns: Model code stays pure PyTorch; parallel strategy lives in config.
Composability: Mix tensor, sequence, and data parallel by editing placements.
Portability: Fewer bespoke abstractions; easier to reason about failure modes and restarts.
✅ PyTorch DTensor-native SPMD training Same training script can scale from 1 GPU to large multi-node jobs by changing the device mesh/config.
✅ Composable Parallelism - PyTorch native FSDP2, HSDP, TP, CP, SP and PP for distributed training.
✅ Optimized kernels - Uses NVIDIA-oriented kernel paths such as Transformer Engine, DeepEP, FlexAttn, TorchSDPA, fused attention, rotary embeddings, Triton, and optional kernel patches.
✅ MoE acceleration - Includes MoE routing and DeepEP integration, plus expert-parallel configurations used in DeepSeek, Qwen MoE, GPT-OSS, and Nemotron MoE benchmarks.
✅ FP8 and mixed precision - FP8 support with torchao and Transformer Engine.
✅ MXFP8 MoE training - Transformer Engine and torchao MXFP8 grouped-expert training on GB200.
✅ Activation checkpointing - Trades recomputation for lower activation memory, especially useful with FSDP and memory-efficient losses.
✅ Memory-efficient loss - Linear-Cut / fused linear cross entropy avoids materializing full logits for the loss, reducing output-layer memory pressure.
✅ Sequence packing - Packs variable-length examples together to reduce padding compute and improve GPU utilization.
✅ FlashAttention packed-sequence support - Packed masks can feed variable-length FlashAttention paths using per-document cu_seqlens.
✅ DCP - Supports PyTorch DCP and SafeTensors, sharded and consolidated layouts, merge/reshard utilities, and Hugging Face-compatible outputs.
✅ Async checkpointing - Can write checkpoints in the background to reduce training stalls caused by I/O.
✅ Dion, Muon, and Muown optimizers - Distributed optimizer integrations with typed recipe configuration.
✅ Environment Support - SLURM, interactive, SkyPilot, and Kubernetes (via SkyPilot) launchers.
SOTA algorithms
✅ Pre-training - Support for model pre-training, including DeepSeekV3.
The model and HellaSwag dataset are public; no Hugging Face token is required. Checkpoints are saved to the mounted checkpoints directory.
For source installation, use uv for reproducible Python environments.
# Setup environment before running any recipes
uv venv
# Choose ONE:
uv sync --frozen # LLM recipes (default)
# uv sync --frozen --extra vlm --extra vlm-media # VLM recipes (Qwen/Mistral/Omni need vlm-media for video/vision; fixes: ImportError: qwen_vl_utils is not installed)
# uv sync --frozen --extra cuda # Optional CUDA deps (e.g., Transformer Engine, Mamba SSM)
# uv sync --frozen --extra cuda_source # Optional bitsandbytes dependency
# uv sync --frozen --extra all # Most optional deps (includes `vlm` and `cuda`; NOTE: excludes media — add --extra media for video/image decode)
# uv sync --frozen --all-extras # Everything (includes `fa`, `moe`, `media`, etc.)
# One-off runs (examples):
# uv run --extra vlm <command>
# uv run --extra cuda <command>
uv run python -c "import nemo_automodel; print('NeMo AutoModel ready')"
Run a Recipe
All recipes are launched via the automodel CLI (or its short alias am). Each YAML config specifies the recipe class and all training parameters:
# LLM example: multi-GPU fine-tuning of Qwen3-0.6B with FSDP2 (no HF token required)
automodel examples/llm_finetune/qwen/qwen3_0p6b_hellaswag.yaml --nproc-per-node 8
# VLM example: single-GPU fine-tuning (Gemma-3-VL) with LoRA
automodel examples/vlm_finetune/gemma3/gemma3_vl_4b_cord_v2_peft.yaml
# Both commands also work with uv run:
uv run automodel examples/llm_finetune/qwen/qwen3_0p6b_hellaswag.yaml --nproc-per-node 8
Tip 提示
NeMo-Run submission: The cli extra adds NeMo Run to the base package: uv pip install "nemo-automodel[cli]". It is additive; the base package still installs its core training dependencies, including PyTorch.
LLM Pre-training
LLM Pre-training Single Node
We provide an example SFT experiment using the FineWeb dataset with a nano-GPT model, ideal for quick experimentation on a single node.
All cluster-specific settings (nodes, GPUs, partition, container, mounts) live in your sbatch script.
NeMo-Run (nemo_run:) sections are also supported – see our
cluster guide for details.
# Memory-efficient SFT with LoRA
automodel examples/llm_finetune/qwen/qwen3_0p6b_hellaswag_peft.yaml
# Override any YAML parameter via the command line:
automodel examples/llm_finetune/qwen/qwen3_0p6b_hellaswag_peft.yaml \
--step_scheduler.local_batch_size 16
Note 注意
Launching a multi-node PEFT example uses the same sbatch slurm.sub workflow as the SFT case above.
VLM Supervised Fine-Tuning (SFT)
We provide a VLM SFT example using Qwen2.5-VL for end-to-end fine-tuning on image-text data.
VLM SFT Single Node
# Qwen2.5-VL on 8 GPUs
automodel examples/vlm_finetune/qwen2_5/qwen2_5_vl_3b_rdr.yaml --nproc-per-node 8
VLM Parameter-Efficient Fine-Tuning (PEFT)
We provide a VLM PEFT (LoRA) example for memory-efficient adaptation with Gemma3 VLM.
VLM PEFT Single Node
# Gemma-3-VL PEFT on 8 GPUs
automodel examples/vlm_finetune/gemma3/gemma3_vl_4b_medpix_peft.yaml --nproc-per-node 8
Supported Models
NeMo AutoModel provides native support for a wide range of models available on the Hugging Face Hub, enabling efficient fine-tuning for various domains. Below is a small sample of ready-to-use families (train as-is or swap any compatible 🤗 causal LM), you can specify nearly any LLM/VLM model available on 🤗 hub:
Check out more LLM and VLM examples. Any causal LM on Hugging Face Hub can be used with the base recipe template, just overwrite --model.pretrained_model_name_or_path <model-id> in the CLI or in the YAML config.
Performance
NeMo AutoModel achieves great training performance on NVIDIA GPUs. Below are highlights from our benchmark results:
Model
#GPUs
Seq Length
Model TFLOPs/sec/GPU
Tokens/sec/GPU
Kernel Optimizations
DeepSeek V3 671B
256
4096
250
1,002
TE + DeepEP
GPT-OSS 20B
8
4096
279
13,058
TE + DeepEP + FlexAttn
Qwen3 MoE 30B
8
4096
212
11,842
TE + DeepEP
For complete benchmark results including configuration details, see the Performance Summary.
🔌 Interoperability
NeMo RL: Use AutoModel checkpoints directly as starting points for DPO/RM/GRPO pipelines.
评论