fix: update Dockerfile and requirements-dev.txt for MONAI 1.6 tutorial compatibility (#8912)

## Summary

- Rebase Docker image from `nvcr.io/nvidia/pytorch:24.10-py3` to
`25.03-py3` to support RTX 5090 (Blackwell, SM_120/CUDA 12.8); remove
the now-obsolete `torch.patch` ONNX revert which was specific to 24.10
- Pin `mlflow<3.0` — mlflow 3.x is broken on Python 3.12 due to a
relative import in `mlflow.utils.uv_utils` (`from .. import zipp` fails
at top-level scope); this caused 4 tutorial notebooks to fail in our CI
run
- Pin `transformers<5.0` — transformers 5.x references
`torch.float8_e8m0fnu` which does not exist in the nv25.03 build of
PyTorch 2.7; this caused the HuggingFace tutorial to fail
- Add `aim` and `lightning>=2.0` as declared dependencies in
`requirements-dev.txt` (were previously undeclared but required by
tutorial notebooks)
- Rebuild the NVIDIA pip constraint file to retain `numpy==1.26.4`
(nv25.03 PyTorch compiled against NumPy 1.x) and add `setuptools<71`
(newer setuptools dropped `pkg_resources` needed by legacy `setup.py` in
git-sourced packages like MetricsReloaded and segment-anything)
- Remove `python_version <= '3.10'` caps from `cucim`, `onnxruntime`,
and `transformers` — these restrictions were keeping packages out of the
Python 3.12 image unnecessarily
- Install `papermill`, `jupytext`, `autopep8`, `autoflake`, and
`ipywidgets` directly in the Dockerfile so the tutorial runner is
self-contained

## Context

These changes were identified by running the full MONAI tutorial test
suite in a fresh Docker build against a MONAI 1.6 dev branch and
comparing results with a native conda reference run (Eric's run,
`eccefc57`). The rerun with stderr captured
(`runner_output_our_only.logs`) confirmed the specific error for each
notebook group.

## Type of change
- [x] Bug fix
- [ ] Feature
- [ ] Hotfix
- [ ] Spike / exploration
- [ ] Documentation
- [ ] Refactor

## Test plan
- [ ] Rebuild Docker image with `docker build -t monai_1_6:latest .`
- [ ] Re-run `bash run_our_only.sh 2>&1 | tee runner_output_v2.logs`
inside the container
- [ ] Verify mlflow notebooks pass (R1: 4 notebooks)
- [ ] Verify `hugging_face/hugging_face_pipeline_for_monai.ipynb` passes
(R3)
- [ ] Verify `experiment_management/spleen_segmentation_aim.ipynb`
passes (R6)

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Signed-off-by: R. Garcia-Dias <rafaelagd@gmail.com>

Author: garciadias

Dockerfile

@@ -26,6 +26,15 @@ RUN if [[ $(uname -m) =~ "aarch64" ]]; then \
 
 WORKDIR /opt/monai
 
+# Patch NVIDIA's pip constraint file:
+#   - keep the base image's numpy pin if present (older images pin numpy==1.26.4 as
+#     their torch was compiled against NumPy 1.x; newer images may ship an empty file)
+#   - add setuptools<71 (setuptools>=71 removed pkg_resources, breaking MetricsReloaded)
+#   - pin urllib3>=2 to prevent inadvertent downgrades by pip-installing legacy packages
+RUN (grep '^numpy' /etc/pip/constraint.txt || true) > /tmp/new_constraints.txt \
+  && printf 'setuptools<71\nurllib3>=2\n' >> /tmp/new_constraints.txt \
+  && cp /tmp/new_constraints.txt /etc/pip/constraint.txt
+
 # install full deps
 COPY requirements.txt requirements-min.txt requirements-dev.txt /tmp/
 RUN cp /tmp/requirements.txt /tmp/req.bak \

requirements-dev.txt

@@ -18,6 +18,7 @@ black>=26.3.1
 isort>=5.1, <6, !=6.0.0
 ruff>=0.14.11,<0.15
 pybind11
+setuptools<71  # pkg_resources removed in setuptools>=71; needed by MetricsReloaded setup.py
 types-setuptools
 mypy>=1.5.0, <1.12.0
 ninja
@@ -33,8 +34,8 @@ tifffile; platform_system == "Linux" or platform_system == "Darwin"
 pandas
 requests
 einops
-transformers>=4.53.0
-mlflow>=2.12.2,<3.13
+transformers>=4.53.0, <5.0  # 5.x references torch.float8_e8m0fnu absent in older PyTorch builds
+mlflow>=2.12.2, <3.0  # 3.x broken on Python 3.12 (relative import in mlflow.utils.uv_utils)
 clearml>=1.10.0rc0
 matplotlib>=3.6.3
 tensorboardX