fix: address Dependabot security alerts for mlflow, transformers, setuptools

- mlflow: bump floor to >=3.11.1 (drops the <3.0 cap), closing CVEs across
  the 2.x/early-3.x line. The cap existed for a Python 3.12 packaging bug
  in mlflow.utils.uv_utils that is no longer present in current releases.
  mlflow>=3.13 also turns the local file-store warning into a hard error
  (#8891); MLFlowHandler now sets MLFLOW_ALLOW_FILE_STORE=true by default
  since it documents and relies on that local store. Documented the
  side effect in the class docstring and added tests covering both the
  unset-defaults-to-true and existing-value-is-preserved cases.
- transformers: bump floor to >=5.5.0 (drops the <5.0 cap), closing two
  HIGH severity CVEs. The cap existed because transformers>=5.x broke
  Transchex: BertConfig was previously a bare ad-hoc class missing
  `_attn_implementation`, and BertLayer's forward() return type changed
  from a tuple to a bare Tensor (the latter was already handled). Fixed
  both in transchex.py and verified against transformers 4.36-4.40 and
  5.5-5.14. The previous <5.0 cap's stated reason (torch.float8_e8m0fnu
  missing from the nv25.03 Docker image's PyTorch 2.7 build) is unrelated
  to transchex.py and should be re-verified against the current NGC base
  image before merging, since it wasn't reproducible against a stock
  PyPI torch>=2.8.0 install.
- setuptools: bump requirements-min.txt floor to >=78.1.1, closing one
  HIGH severity CVE. Still capped at <=79.0.1 because setuptools>=80
  breaks MONAI's own setup.py CLI usage (#8439); a MEDIUM severity CVE
  fixed in 83.0.0 remains open until that's resolved. Also drop
  requirements-dev.txt's separate `setuptools<71` cap, which conflicted
  with that floor and broke CI dependency installation (mypy, hyena-dep,
  full-dep): it was added for MetricsReloaded's legacy pkg_resources-based
  setup.py, but the `monai-support` branch already has that import
  commented out, and the pinned segment-anything commit never used
  pkg_resources either, so the cap is no longer needed.

Verified via targeted venv testing against the actual pinned versions
(transformers==5.5.0, mlflow==3.11.1): tests/networks/nets/test_transchex.py
and tests/handlers/test_handler_mlflow.py both pass.

CodeRabbit flagged MultiModal.__init__ (transchex.py) for allegedly calling
transformers' PreTrainedModel.__init__() without a config, which 5.5.0
requires. That's a false positive: MultiModal subclasses transchex.py's own
local `BertPreTrainedModel(nn.Module)` shim, not transformers' class, so
HF's config-in-super().__init__() requirement doesn't apply. Confirmed by
running the test suite unchanged against transformers 5.5.0 (3 passed).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Signed-off-by: R. Garcia-Dias <rafaelagd@gmail.com>

Author: garciadias

docs/requirements.txt

@@ -20,8 +20,8 @@ sphinxcontrib-serializinghtml
 sphinx-autodoc-typehints==1.11.1
 pandas
 einops
-transformers>=4.53.0
-mlflow>=2.12.2,<3.13
+transformers>=5.5.0
+mlflow>=3.11.1  # see requirements-dev.txt for why the previous <3.0/<3.13 caps are no longer needed
 clearml>=1.10.0rc0
 tensorboardX
 imagecodecs; platform_system == "Linux" or platform_system == "Darwin"

monai/handlers/mlflow_handler.py

@@ -60,6 +60,10 @@ class MLFlowHandler:
           ``engine.state.metrics`` in MLFlow.
         - When ITERATION_COMPLETED, track expected item in
           ``self.output_transform(engine.state.output)`` in MLFlow, default to `Loss`.
+        - On construction, sets the ``MLFLOW_ALLOW_FILE_STORE`` environment variable to
+          ``"true"`` if it is not already set, since ``MLFlowHandler`` defaults to (and
+          documents) tracking to the local filesystem store, which mlflow>=3.13 otherwise
+          refuses to use. Any value the user has already set is left untouched.
 
     Usage example is available in the tutorial:
     https://github.com/Project-MONAI/tutorials/blob/master/3d_segmentation/unet_segmentation_3d_ignite.ipynb.
@@ -156,6 +160,11 @@ def __init__(
         self.experiment_param = experiment_param
         self.artifacts = ensure_tuple(artifacts)
         self.optimizer_param_names = ensure_tuple(optimizer_param_names)
+        # mlflow>=3.13 raises instead of warning when the tracking URI resolves to the local
+        # filesystem store (e.g. the default `mlruns` directory), see
+        # https://github.com/Project-MONAI/MONAI/issues/8891. MLFlowHandler documents and relies
+        # on this local file store as its default, so opt back into it unless the user overrode it.
+        os.environ.setdefault("MLFLOW_ALLOW_FILE_STORE", "true")
         self.client = mlflow.MlflowClient(tracking_uri=tracking_uri if tracking_uri else None)
         self.run_finish_status = mlflow.entities.RunStatus.to_string(mlflow.entities.RunStatus.FINISHED)
         self.close_on_complete = close_on_complete

monai/networks/nets/transchex.py

@@ -23,6 +23,7 @@
 transformers = optional_import("transformers")
 load_tf_weights_in_bert = optional_import("transformers", name="load_tf_weights_in_bert")[0]
 cached_file = optional_import("transformers.utils", name="cached_file")[0]
+BertConfig = optional_import("transformers", name="BertConfig")[0]
 BertEmbeddings = optional_import("transformers.models.bert.modeling_bert", name="BertEmbeddings")[0]
 BertLayer = optional_import("transformers.models.bert.modeling_bert", name="BertLayer")[0]
 
@@ -219,7 +220,11 @@ def __init__(
 
         """
         super().__init__()
-        self.config = type("obj", (object,), bert_config)
+        self.config = BertConfig(**bert_config)
+        # explicitly select the eager attention path: transformers>=4.48 dispatches attention
+        # implementations via `config._attn_implementation`, which is otherwise left unset since
+        # `bert_config` above does not come from a `from_pretrained` call.
+        self.config._attn_implementation = "eager"
         self.embeddings = BertEmbeddings(self.config)
         self.language_encoder = nn.ModuleList([BertLayer(self.config) for _ in range(num_language_layers)])
         self.vision_encoder = nn.ModuleList([BertLayer(self.config) for _ in range(num_vision_layers)])

requirements-dev.txt

@@ -18,7 +18,6 @@ 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
@@ -34,8 +33,15 @@ tifffile; platform_system == "Linux" or platform_system == "Darwin"
 pandas
 requests
 einops
-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)
+transformers>=5.5.0  # 4.x/early-5.x are vulnerable (GHSA-fgcw-684q-jj6r, GHSA-29pf-2h5f-8g72); the
+# previous <5.0 cap was needed because the nv25.03 Docker image's PyTorch 2.7 build lacked
+# `torch.float8_e8m0fnu`; the PyPI `torch>=2.8.0` floor this repo declares has it, and
+# `monai/networks/nets/transchex.py` has been updated for transformers>=5's BertLayer/BertConfig
+# API changes. Re-verify against the current NGC base image before merging.
+mlflow>=3.11.1  # <3.4.0rc0 through <=3.10.1 versions have several CVEs; the previous <3.0 cap was
+# for a Python 3.12 packaging bug in mlflow.utils.uv_utils, no longer present in current releases.
+# mlflow>=3.13 also hard-errors on the local file-store backend unless MLFLOW_ALLOW_FILE_STORE=true
+# (see monai/handlers/mlflow_handler.py and https://github.com/Project-MONAI/MONAI/issues/8891).
 clearml>=1.10.0rc0
 matplotlib>=3.6.3
 tensorboardX

requirements-min.txt

@@ -1,7 +1,10 @@
 # Requirements for minimal tests
 -r requirements.txt
-setuptools>=50.3.0,<66.0.0,!=60.6.0 ; python_version < "3.12"
-setuptools>=70.2.0,<=79.0.1; python_version >= "3.12"
+# <78.1.1 is vulnerable to GHSA-5rjg-fvgr-3xxf (path traversal / arbitrary file write). The
+# upper bound is unrelated to the CVE: setuptools>=80 dropped the legacy `setup.py` CLI
+# invocation MONAI's own build script relies on (Project-MONAI/MONAI#8439), so this can't yet
+# go as high as 83.0.0, which would also fix GHSA-h35f-9h28-mq5c (MANIFEST.in exclusion bypass).
+setuptools>=78.1.1,<=79.0.1
 coverage>=5.5
 parameterized
 packaging

tests/handlers/test_handler_mlflow.py

@@ -106,6 +106,31 @@ def _update_metric(engine):
             # the run count should equal to the times of creating engine
             self.assertEqual(create_engine_times, run_cnt)
 
+    def test_allow_file_store_env_var_defaults_true(self):
+        original = os.environ.pop("MLFLOW_ALLOW_FILE_STORE", None)
+        try:
+            with tempfile.TemporaryDirectory() as tempdir:
+                MLFlowHandler(tracking_uri=path_to_uri(os.path.join(tempdir, "mlflow_test")))
+            self.assertEqual(os.environ["MLFLOW_ALLOW_FILE_STORE"], "true")
+        finally:
+            if original is None:
+                os.environ.pop("MLFLOW_ALLOW_FILE_STORE", None)
+            else:
+                os.environ["MLFLOW_ALLOW_FILE_STORE"] = original
+
+    def test_allow_file_store_env_var_preserves_existing(self):
+        original = os.environ.get("MLFLOW_ALLOW_FILE_STORE")
+        os.environ["MLFLOW_ALLOW_FILE_STORE"] = "false"
+        try:
+            with tempfile.TemporaryDirectory() as tempdir:
+                MLFlowHandler(tracking_uri=path_to_uri(os.path.join(tempdir, "mlflow_test")))
+            self.assertEqual(os.environ["MLFLOW_ALLOW_FILE_STORE"], "false")
+        finally:
+            if original is None:
+                os.environ.pop("MLFLOW_ALLOW_FILE_STORE", None)
+            else:
+                os.environ["MLFLOW_ALLOW_FILE_STORE"] = original
+
     def test_metrics_track(self):
         experiment_param = {"backbone": "efficientnet_b0"}
         with tempfile.TemporaryDirectory() as tempdir: