File size: 1,626 Bytes
1a18f22 ed138b4 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 | # Isolated env for U-Mamba reference baseline (nnU-Net v2.1.1 + Mamba).
# VALIDATED on the A100 server (2026-06-05): mamba CUDA kernel + UMambaBot trainer
# run end-to-end. mamba_ssm/causal-conv1d are installed from PREBUILT WHEELS to
# avoid local CUDA compilation (system nvcc 12.8 != torch cu118).
#
# Reproduce with these exact commands (NOT `conda env create -f`; pip-driven):
#
# conda create -n umamba python=3.10 -y && conda activate umamba
# pip install torch==2.0.1 torchvision==0.15.2 --index-url https://download.pytorch.org/whl/cu118
# pip install https://github.com/Dao-AILab/causal-conv1d/releases/download/v1.2.0.post2/causal_conv1d-1.2.0.post2+cu118torch2.0cxx11abiFALSE-cp310-cp310-linux_x86_64.whl
# pip install https://github.com/state-spaces/mamba/releases/download/v1.2.0.post1/mamba_ssm-1.2.0.post1+cu118torch2.0cxx11abiFALSE-cp310-cp310-linux_x86_64.whl
# pip install -e sota/U-Mamba/umamba
# # critical pins (otherwise: transformers drops GreedySearchDecoderOnlyOutput; opencv/numpy clash):
# pip install "numpy<2" "transformers==4.38.2" "opencv-python-headless==4.9.0.80"
#
# Train (uses nnU-Net format from framework/nnunet_convert.py; A100 only — bf16/sm_80):
# export CUDA_VISIBLE_DEVICES=4
# nnUNetv2_plan_and_preprocess -d <ID> -c 2d # umamba env does its OWN 2.1.1 preprocess
# cp <raw>/Dataset<ID>_*/splits_final.json $nnUNet_preprocessed/Dataset<ID>_*/splits_final.json
# nnUNetv2_train <ID> 2d 0 -tr nnUNetTrainerUMambaBot
# # AMP can NaN in Mamba: if so use -tr nnUNetTrainerUMambaEncNoAMP
name: umamba
channels: [conda-forge]
dependencies:
- python=3.10
|