Two more things only macOS and Windows could tell us

Ran the branch on macos-14 and windows-latest, which no workflow in this repo
does. Two failures, both in the tests rather than in saving_utils.

test_a_local_4bit_copy_still_merges_when_the_config_fetch_is_offline uses the same
`Outputs/MyModel` over an on-disk `outputs/mymodel` shape as its sibling file, and
for the same reason: that case difference is the only way to make
os.path.exists(name) miss while check_local_model_exists(name) hits. On APFS and
NTFS os.path.exists answers True, the local branch is taken, and the Hub round trip
never happens. macOS failed on the path comparison and Windows on `resolved[1] is
True`, which is the same cause wearing two symptoms. Probe and skip with the reason,
as the sibling now does.

test_a_local_base_still_merges_with_the_hub_down died in MPS:

  RuntimeError: MPS backend out of memory (MPS allocated: 8.00 MiB, ... max
  allowed: 7.93 GiB). Tried to allocate 256 bytes on shared pool.

Refusing 256 bytes with 7.93 GiB allowed is the macos-14 runner's virtualized MPS,
not this branch, and _active_merge_device prefers MPS on purpose. But it exposed a
claim in that file that was not true: the docstring said CPU-only while
set_offline_cpu_env only sets UNSLOTH_ALLOW_CPU, which permits CPU rather than
requiring it. On this box the merge math was running on cuda. These cells assert
control flow, whether the merge raises and whether it writes, so the device is
incidental and letting the host pick it makes the result depend on hardware the
assertions have no opinion about. Pinned to CPU and said so.

Verified the pin is real rather than decorative: _active_merge_device() answers
cuda here unpatched, and the merged tensors come back on cpu with the fixture.
Behavioral gate 564 passed, 2 skipped.

Author: danielhanchen

tests/test_gated_and_offline_config_classification.py

@@ -79,6 +79,28 @@
 _ON_DISK = ("outputs", "mymodel")
 
 
+def _require_a_case_sensitive_filesystem(tmp_path):
+    """Same precondition, same reason as in that sibling file, stated not assumed.
+
+    A case difference is the only thing that makes `os.path.exists(name)` miss
+    while `check_local_model_exists(name)` hits, so the shape cannot be written
+    filesystem independently. On macOS APFS and Windows NTFS `os.path.exists`
+    answers True for `Outputs/MyModel`, the local branch is taken, and the
+    resolved path comes back in the requested casing instead of the on-disk one.
+
+    macos-14 and windows-latest both fail here without this: macOS on the path
+    comparison, Windows on `resolved[1] is True`.
+    """
+    probe = tmp_path / "case_probe"
+    probe.mkdir(exist_ok = True)
+    if os.path.exists(str(tmp_path / "CASE_PROBE")):
+        pytest.skip(
+            "case insensitive filesystem: os.path.exists would resolve "
+            f"{_REQUESTED!r} locally and skip the Hub branch under test"
+        )
+    pass
+
+
 class _StubResponse:
     """The attributes `HfHubHTTPError.__init__` actually reads on 1.x."""
     status_code = 403
@@ -284,6 +306,7 @@ def test_a_local_4bit_copy_still_merges_when_the_config_fetch_is_offline(
 ):
     """The new raise must still reach the local fallback, exactly as a 429 does.
     Both 4bit merges fold LoRA into weights already in memory."""
+    _require_a_case_sensitive_filesystem(tmp_path)
     directory = _make_local_model(tmp_path.joinpath(*_ON_DISK), quant_config = _NF4)
     monkeypatch.chdir(tmp_path)
     _patch_ls_present(monkeypatch)

tests/test_merge_e2e_hub_unreachable.py

@@ -30,10 +30,18 @@
 things a user experiences: an exception is raised, and no output directory is left
 behind that could be mistaken for a successful export.
 
-Uses the tiny synthetic models from `_merge_e2e_helpers`, so this is CPU-only and
-sub-second. `set_offline_cpu_env()` there already forbids the network; the Hub
-entry points are monkeypatched on top so a failure is deterministic rather than
-dependent on being genuinely offline.
+Uses the tiny synthetic models from `_merge_e2e_helpers`, so this is sub-second.
+`set_offline_cpu_env()` there already forbids the network; the Hub entry points are
+monkeypatched on top so a failure is deterministic rather than dependent on being
+genuinely offline.
+
+The merge math is pinned to CPU. `set_offline_cpu_env()` sets `UNSLOTH_ALLOW_CPU`,
+which permits CPU rather than requiring it, so `_active_merge_device()` still
+prefers whatever accelerator is present: CUDA on a training box, MPS on a Mac. What
+is under test here is control flow, whether the merge raises and whether it writes,
+so the device is incidental, and leaving it to the host makes the result depend on
+hardware these assertions have no opinion about. macos-14 showed why: its
+virtualized MPS reports 7.93 GiB available and then refuses a 256 byte allocation.
 """
 
 from __future__ import annotations
@@ -50,6 +58,16 @@
 FAMILY = "llama"
 
 
+@pytest.fixture(autouse = True)
+def _merge_on_the_cpu(monkeypatch):
+    """Pin `_merge_lora`'s device, for the reason in the module docstring.
+
+    `_active_merge_device` is `lru_cache`d, so replacing the module attribute is
+    both what takes effect and what gets undone cleanly by monkeypatch.
+    """
+    monkeypatch.setattr(saving_utils, "_active_merge_device", lambda: "cpu")
+
+
 def _skip_if_missing():
     if not H.family_available(FAMILY):
         pytest.skip(f"{FAMILY} unavailable in this transformers")