datasets>=4.0.0
megatron-fsdp>=0.2.3
mistral-common[audio,hf-hub,sentencepiece]
pybind11
pyyaml
tiktoken
torch>=2.6.0
torchdata
transformers==5.8.1
wandb>=0.28.0
torchao
mlflow
flashoptim>=0.1.3

[all]
nemo_automodel[cli]
nemo_automodel[cuda]
nemo_automodel[delta-databricks]
nemo_automodel[diffusion]
nemo_automodel[extra]
nemo_automodel[fla]
nemo_automodel[s3]
nemo_automodel[vlm]

[cli]
pyyaml
nemo-run

[cuda]
causal-conv1d
mamba-ssm
nv-grouped-gemm
onnxscript>=0.5.6
transformer-engine[pytorch]>=2.14.1
tilelang>=0.1.11
tile-kernels==1.0.0
apache-tvm-ffi<=0.1.11

[cuda_source]
bitsandbytes

[delta-databricks]
deltalake>=1.0.0
databricks-sql-connector>=3.0.0

[diffusion]
diffusers>=0.37.0
ftfy
imageio
torchvision

[diffusion-media]
imageio-ffmpeg
opencv-python-headless==4.10.0.84

[diffusion_kernels]
nemo_automodel[diffusion]
kernels

[extra]
perceptron
sentencepiece

[fa]
flash-attn<=2.8.3

[fla]
flash-linear-attention>=0.4.2

[media]
nemo_automodel[vlm-media]
nemo_automodel[diffusion-media]

[moe]
nemo_automodel[cuda]
nemo_automodel[fla]
deep_ep

[msc]
multi-storage-client>=0.13

[s3]
boto3>=1.34

[vlm]
backoff
mistral_common>=1.11.0
numpy
numba
open-clip-torch
pillow
soundfile
timm<=1.0.22

[vlm-media]
opencv-python-headless==4.10.0.84
qwen-omni-utils
qwen-vl-utils

[vlm-media:platform_machine == "x86_64" and platform_system != "Darwin"]
decord
