test: centralise HF Hub skip guard to boost CI resiliency against 429 errors

test/backends/test_huggingface.py

@@ -47,6 +47,7 @@
 from mellea.stdlib.components import Intrinsic, Message
 from mellea.stdlib.context import ChatContext, SimpleContext
 from mellea.stdlib.requirements import ALoraRequirement, LLMaJRequirement
+from test.conftest import hf_skip
 
 
 @pytest.fixture(scope="module")
@@ -59,15 +60,18 @@ def backend():
     against this granite-3.3 backend will fail once revision pinning is wired through
     in phase-2.2 (#1141). Other intrinsics are not affected.
     """
-    backend = LocalHFBackend(
-        model_id=model_ids.IBM_GRANITE_4_1_3B, cache=SimpleLRUCache(5)
-    )
-    backend.add_adapter(
-        IntrinsicAdapter("requirement-check", base_model_name=backend.base_model_name)
-    )
-    backend.add_adapter(
-        IntrinsicAdapter("answerability", base_model_name=backend.base_model_name)
-    )
+    with hf_skip():
+        backend = LocalHFBackend(
+            model_id=model_ids.IBM_GRANITE_4_1_3B, cache=SimpleLRUCache(5)
+        )
+        backend.add_adapter(
+            IntrinsicAdapter(
+                "requirement-check", base_model_name=backend.base_model_name
+            )
+        )
+        backend.add_adapter(
+            IntrinsicAdapter("answerability", base_model_name=backend.base_model_name)
+        )
     yield backend
 
     from test.conftest import cleanup_gpu_backend

test/backends/test_huggingface_tools.py

@@ -25,14 +25,16 @@
 from mellea.backends.cache import SimpleLRUCache
 from mellea.backends.huggingface import LocalHFBackend
 from mellea.stdlib.context import ChatContext
+from test.conftest import hf_skip
 
 
 @pytest.fixture(scope="module")
 def backend():
     """Shared HuggingFace backend for all tests in this module."""
-    backend = LocalHFBackend(
-        model_id=model_ids.MISTRALAI_MISTRAL_0_3_7B, cache=SimpleLRUCache(5)
-    )
+    with hf_skip():
+        backend = LocalHFBackend(
+            model_id=model_ids.MISTRALAI_MISTRAL_0_3_7B, cache=SimpleLRUCache(5)
+        )
     # add_granite_aloras(backend)
     yield backend
 

test/conftest.py

@@ -1,3 +1,4 @@
+import contextlib
 import gc
 import os
 import subprocess
@@ -9,6 +10,30 @@
 
 from mellea.core import MelleaLogger
 
+# ============================================================================
+# HuggingFace Hub Skip Helper
+# ============================================================================
+
+
+@contextlib.contextmanager
+def hf_skip():
+    """Skip the current test/fixture when a HuggingFace Hub call fails.
+
+    Catches OSError (covers HfHubHTTPError, LocalEntryNotFoundError, and
+    transformers-wrapped Hub errors) plus requests network errors.  Use this
+    in every fixture that downloads model weights or YAML files from the Hub
+    so a 429 or transient network blip results in a clean skip rather than an
+    ERROR that masks unrelated failures.
+    """
+    try:
+        yield
+    except (OSError, requests.exceptions.RequestException) as e:
+        pytest.skip(
+            f"HuggingFace Hub not accessible: {type(e).__name__}: {e}",
+            allow_module_level=True,
+        )
+
+
 # Try to import optional dependencies for system detection
 try:
     import psutil

test/formatters/granite/test_intrinsics_formatters.py

@@ -21,7 +21,6 @@
 
 torch = pytest.importorskip("torch", reason="torch not installed — install mellea[hf]")
 import yaml
-from huggingface_hub.errors import LocalEntryNotFoundError
 
 # First Party
 from mellea.formatters.granite import (
@@ -32,6 +31,7 @@
 )
 from mellea.formatters.granite.base import util as base_util
 from mellea.formatters.granite.intrinsics import json_util, util as intrinsics_util
+from test.conftest import hf_skip
 
 
 def _read_file(name):
@@ -147,15 +147,13 @@ def _resolve_yaml(self):
         object. Called at fixture creation (execution time) to prevent collection time errors.
         """
         if not self.yaml_file:
-            try:
+            with hf_skip():
                 self.yaml_file = intrinsics_util.obtain_io_yaml(
                     self.task,
                     self.base_model_id,
                     self.repo_id,
                     revision=self.revision,  # type: ignore
                 )
-            except (LocalEntryNotFoundError, requests.exceptions.RequestException) as e:
-                pytest.skip(f"HuggingFace Hub not accessible: {type(e).__name__}: {e}")
         return self
 
 
@@ -537,12 +535,10 @@ def test_read_yaml():
     IntrinsicsRewriter(config_file=_INPUT_YAML_DIR / "answerability.yaml")
 
     # Read from Hugging Face hub.
-    try:
+    with hf_skip():
         local_path = intrinsics_util.obtain_io_yaml(
             "answerability", "granite-4.0-micro", _RAG_INTRINSICS_REPO_NAME
         )
-    except (LocalEntryNotFoundError, requests.exceptions.RequestException) as e:
-        pytest.skip(f"HuggingFace Hub not accessible: {type(e).__name__}: {e}")
     IntrinsicsRewriter(config_file=local_path)
 
 

test/stdlib/components/docs/test_richdocument.py

@@ -11,6 +11,7 @@
 import mellea
 from mellea.core import TemplateRepresentation
 from mellea.stdlib.components.docs.richdocument import RichDocument, Table
+from test.conftest import hf_skip
 from test.predicates import require_gpu
 
 pytestmark = pytest.mark.integration
@@ -20,9 +21,10 @@
 def rd() -> RichDocument:
     # Use a specific document so we can test some of the functionality
     # related to extracting and transforming text.
-    return RichDocument.from_document_file(
-        "https://arxiv.org/pdf/1906.04043", do_ocr=False
-    )
+    with hf_skip():
+        return RichDocument.from_document_file(
+            "https://arxiv.org/pdf/1906.04043", do_ocr=False
+        )
 
 
 def test_richdocument_basics(rd: RichDocument):

test/stdlib/components/intrinsic/test_core.py

@@ -17,7 +17,7 @@
 from mellea.stdlib.components.intrinsic import core
 from mellea.stdlib.components.intrinsic._util import call_intrinsic
 from mellea.stdlib.context import ChatContext
-from test.conftest import cleanup_gpu_backend
+from test.conftest import cleanup_gpu_backend, hf_skip
 from test.predicates import require_gpu
 from test.stdlib.components.intrinsic.test_rag import (
     _read_input_json as _read_rag_input_json,
@@ -48,7 +48,8 @@ def _backend():
     # Prevent thrashing if the default device is CPU
     torch.set_num_threads(4)
 
-    backend_ = LocalHFBackend(model_id=BASE_MODEL)
+    with hf_skip():
+        backend_ = LocalHFBackend(model_id=BASE_MODEL)
     yield backend_
 
     # Code after yield is cleanup code.

test/stdlib/components/intrinsic/test_guardian.py

@@ -14,7 +14,7 @@
 from mellea.stdlib.components import Document, Message
 from mellea.stdlib.components.intrinsic import guardian
 from mellea.stdlib.context import ChatContext
-from test.conftest import cleanup_gpu_backend
+from test.conftest import cleanup_gpu_backend, hf_skip
 from test.predicates import require_gpu
 
 # Skip entire module in CI since all tests are qualitative
@@ -37,7 +37,8 @@ def _backend():
     """Backend used by the tests in this file. Module-scoped to avoid reloading the model for each test."""
     torch.set_num_threads(4)
 
-    backend_ = LocalHFBackend(model_id=IBM_GRANITE_4_MICRO_3B.hf_model_name)  # type: ignore
+    with hf_skip():
+        backend_ = LocalHFBackend(model_id=IBM_GRANITE_4_MICRO_3B.hf_model_name)  # type: ignore
     yield backend_
 
     cleanup_gpu_backend(backend_, "test_guardian")

test/stdlib/components/intrinsic/test_rag.py

@@ -15,6 +15,7 @@
 from mellea.stdlib.components import Document, Message
 from mellea.stdlib.components.intrinsic import rag
 from mellea.stdlib.context import ChatContext
+from test.conftest import hf_skip
 from test.predicates import require_gpu
 
 # Skip entire module in CI since all 7 tests are qualitative
@@ -42,7 +43,8 @@ def _backend():
     torch.set_num_threads(4)
 
     # No adapters for hybrid version.
-    backend_ = LocalHFBackend(model_id=IBM_GRANITE_4_1_3B.hf_model_name)
+    with hf_skip():
+        backend_ = LocalHFBackend(model_id=IBM_GRANITE_4_1_3B.hf_model_name)
     yield backend_
 
     from test.conftest import cleanup_gpu_backend
@@ -57,7 +59,8 @@ def _backend_4_0():
     torch.set_num_threads(4)
 
     # No adapters for hybrid version.
-    backend_ = LocalHFBackend(model_id=IBM_GRANITE_4_MICRO_3B.hf_model_name)
+    with hf_skip():
+        backend_ = LocalHFBackend(model_id=IBM_GRANITE_4_MICRO_3B.hf_model_name)
     yield backend_
 
     from test.conftest import cleanup_gpu_backend

test/stdlib/requirements/test_groundedness_requirement.py

@@ -9,13 +9,15 @@
 from mellea.stdlib.components import Document, Message
 from mellea.stdlib.context import ChatContext
 from mellea.stdlib.requirements.rag import GroundednessRequirement
+from test.conftest import hf_skip
 from test.predicates import require_gpu
 
 
 @pytest.fixture
 def backend():
     """Provide HuggingFace backend for tests."""
-    return LocalHFBackend(model_id="ibm-granite/granite-4.0-micro")
+    with hf_skip():
+        return LocalHFBackend(model_id="ibm-granite/granite-4.0-micro")
 
 
 @pytest.fixture

test/stdlib/requirements/test_groundedness_requirement_e2e.py

@@ -6,13 +6,15 @@
 from mellea.stdlib.components import Document, Message
 from mellea.stdlib.context import ChatContext
 from mellea.stdlib.requirements.rag import GroundednessRequirement
+from test.conftest import hf_skip
 from test.predicates import require_gpu
 
 
 @pytest.fixture(scope="session")
 def backend():
     """Provide HuggingFace backend for tests (session-scoped to avoid OOM)."""
-    return LocalHFBackend(model_id="ibm-granite/granite-4.0-micro")
+    with hf_skip():
+        return LocalHFBackend(model_id="ibm-granite/granite-4.0-micro")
 
 
 @pytest.fixture

test/stdlib/test_spans.py

@@ -10,6 +10,7 @@
 from mellea.core import CBlock
 from mellea.stdlib.components import SimpleComponent
 from mellea.stdlib.session import MelleaSession, start_session
+from test.conftest import hf_skip
 from test.predicates import require_gpu
 
 # Module-level markers for all tests using Granite 4.1 3B model
@@ -19,11 +20,12 @@
 # We edit the context type in the async tests below. Don't change the scope here.
 @pytest.fixture(scope="function")
 def m_session(gh_run):
-    m = start_session(
-        "hf",
-        model_id=IBM_GRANITE_4_1_3B,
-        model_options={ModelOption.MAX_NEW_TOKENS: 64},
-    )
+    with hf_skip():
+        m = start_session(
+            "hf",
+            model_id=IBM_GRANITE_4_1_3B,
+            model_options={ModelOption.MAX_NEW_TOKENS: 64},
+        )
     yield m
 
     from test.conftest import cleanup_gpu_backend

test/telemetry/test_metrics_backend.py

@@ -16,6 +16,7 @@
 )
 from mellea.stdlib.components import Message
 from mellea.stdlib.context import SimpleContext
+from test.conftest import hf_skip
 from test.predicates import require_api_key, require_gpu
 
 # Check if OpenTelemetry is available
@@ -72,10 +73,11 @@ def hf_metrics_backend(gh_run):
     from mellea.backends.cache import SimpleLRUCache
     from mellea.backends.huggingface import LocalHFBackend
 
-    backend = LocalHFBackend(
-        model_id=IBM_GRANITE_4_1_3B.hf_model_name,  # type: ignore
-        cache=SimpleLRUCache(5),
-    )
+    with hf_skip():
+        backend = LocalHFBackend(
+            model_id=IBM_GRANITE_4_1_3B.hf_model_name,  # type: ignore
+            cache=SimpleLRUCache(5),
+        )
 
     yield backend