diff --git a/tests/test_uma_safetensors_load.py b/tests/test_uma_safetensors_load.py
new file mode 100644
index 0000000000..c6d304ab4f
--- /dev/null
+++ b/tests/test_uma_safetensors_load.py
@@ -0,0 +1,229 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
+
+"""Unit tests for the UMA safetensors clone-then-move fast load.
+
+The module loads in isolation with a fake ``transformers.modeling_utils``. The
+CUDA correctness check needs a GPU; gating, passthrough, idempotency and opt-out
+are GPU-free. The gate is lazy (wrapper-time), so the wrapper installs
+everywhere and passes through when it's off.
+"""
+
+from __future__ import annotations
+
+import importlib.util
+import sys
+import types
+from pathlib import Path
+
+import pytest
+
+torch = pytest.importorskip("torch")
+safetensors_torch = pytest.importorskip("safetensors.torch")
+import safetensors  # noqa: E402
+
+_MODULE_PATH = Path(__file__).resolve().parent.parent / "unsloth" / "models" / "_uma_safetensors.py"
+
+
+def _load_module():
+    spec = importlib.util.spec_from_file_location("uma_safetensors_under_test", _MODULE_PATH)
+    assert spec is not None and spec.loader is not None
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+@pytest.fixture()
+def uma():
+    return _load_module()
+
+
+@pytest.fixture()
+def force_uma(uma, monkeypatch):
+    """Force the UMA gate on (or off) and keep the lru_cache from sticking."""
+
+    def _set(on):
+        monkeypatch.setenv("UNSLOTH_FORCE_UMA", "1" if on else "0")
+        uma.is_integrated_unified_memory_gpu.cache_clear()
+
+    yield _set
+    uma.is_integrated_unified_memory_gpu.cache_clear()
+
+
+@pytest.fixture()
+def tiny_safetensors(tmp_path):
+    tensors = {
+        "w": torch.arange(32, dtype = torch.float32).reshape(4, 8),
+        "b": torch.tensor([1.0, 2.0, 3.0, 4.0], dtype = torch.float32),
+    }
+    path = tmp_path / "model.safetensors"
+    safetensors_torch.save_file(tensors, str(path))
+    return path, tensors
+
+
+def _install_fake_modeling_utils(monkeypatch, safe_open_fn):
+    fake_transformers = types.ModuleType("transformers")
+    fake_mu = types.ModuleType("transformers.modeling_utils")
+    fake_mu.safe_open = safe_open_fn
+    fake_transformers.modeling_utils = fake_mu
+    monkeypatch.setitem(sys.modules, "transformers", fake_transformers)
+    monkeypatch.setitem(sys.modules, "transformers.modeling_utils", fake_mu)
+    return fake_mu
+
+
+# --- detection / gate ---
+
+
+def test_force_uma_on(uma, monkeypatch):
+    monkeypatch.setenv("UNSLOTH_FORCE_UMA", "1")
+    uma.is_integrated_unified_memory_gpu.cache_clear()
+    assert uma.is_integrated_unified_memory_gpu() is True
+
+
+def test_force_uma_off(uma, monkeypatch):
+    monkeypatch.setenv("UNSLOTH_FORCE_UMA", "0")
+    uma.is_integrated_unified_memory_gpu.cache_clear()
+    assert uma.is_integrated_unified_memory_gpu() is False
+
+
+@pytest.mark.parametrize(
+    "device,expected",
+    [
+        (0, True),
+        ("cuda", True),
+        ("cuda:0", True),
+        ("cpu", False),
+        ("disk", False),
+        (None, False),
+        (True, False),  # a bool is not a device index
+    ],
+)
+def test_is_cuda_target(uma, device, expected):
+    assert uma._is_cuda_target(device) is expected
+
+
+def test_is_cuda_target_torch_device(uma):
+    assert uma._is_cuda_target(torch.device("cuda", 0)) is True
+    assert uma._is_cuda_target(torch.device("cpu")) is False
+
+
+# --- patch gating ---
+
+
+def test_wrapper_passes_through_off_uma(uma, force_uma, monkeypatch):
+    """Gate OFF: every call -- including CUDA targets -- passes straight through
+    to the real safe_open (the gate is evaluated lazily inside the wrapper)."""
+    force_uma(False)
+    sentinel = object()
+    calls = []
+
+    def fake_safe_open(*args, **kwargs):
+        calls.append((args, kwargs))
+        return sentinel
+
+    fake_mu = _install_fake_modeling_utils(monkeypatch, fake_safe_open)
+    assert uma.patch_unified_memory_safetensors_load() is True
+    assert getattr(fake_mu.safe_open, "_unsloth_uma_clone", False) is True
+    out = fake_mu.safe_open("shard.safetensors", "pt", "cuda:0")
+    assert out is sentinel
+    assert calls == [(("shard.safetensors", "pt", "cuda:0"), {})]
+
+
+def test_patch_install_does_not_evaluate_gate(uma, monkeypatch):
+    """Installing the wrapper must NOT query the integrated-GPU property -- that
+    would init CUDA at ``import unsloth`` (fork-unsafe, and before the Spark
+    allocator config is set)."""
+
+    def _boom():
+        raise AssertionError("gate must not be evaluated at install time")
+
+    _install_fake_modeling_utils(monkeypatch, safetensors.safe_open)
+    monkeypatch.setattr(uma, "is_integrated_unified_memory_gpu", _boom)
+    assert uma.patch_unified_memory_safetensors_load() is True
+
+
+def test_patch_noop_when_opted_out(uma, force_uma, monkeypatch):
+    force_uma(True)
+    monkeypatch.setenv("UNSLOTH_DISABLE_UMA_CLONE_LOAD", "1")
+    real = object()
+    fake_mu = _install_fake_modeling_utils(monkeypatch, real)
+    assert uma.patch_unified_memory_safetensors_load() is False
+    assert fake_mu.safe_open is real
+
+
+def test_patch_installs_and_is_idempotent(uma, force_uma, monkeypatch):
+    force_uma(True)
+    fake_mu = _install_fake_modeling_utils(monkeypatch, safetensors.safe_open)
+    assert uma.patch_unified_memory_safetensors_load() is True
+    wrapped = fake_mu.safe_open
+    assert getattr(wrapped, "_unsloth_uma_clone", False) is True
+    # second call must not double-wrap
+    assert uma.patch_unified_memory_safetensors_load() is True
+    assert fake_mu.safe_open is wrapped
+
+
+# --- correctness ---
+
+
+def test_cpu_target_is_passthrough(uma, force_uma, monkeypatch, tiny_safetensors):
+    path, tensors = tiny_safetensors
+    force_uma(True)
+    fake_mu = _install_fake_modeling_utils(monkeypatch, safetensors.safe_open)
+    uma.patch_unified_memory_safetensors_load()
+    # device="cpu" must NOT be intercepted -> identical data, still on CPU.
+    with fake_mu.safe_open(str(path), framework = "pt", device = "cpu") as f:
+        for key, expected in tensors.items():
+            got = f.get_slice(key)[:]
+            assert got.device.type == "cpu"
+            assert torch.equal(got, expected)
+
+
+@pytest.mark.skipif(
+    not (hasattr(torch, "cuda") and torch.cuda.is_available()),
+    reason = "needs a GPU for the host->device clone-and-move path",
+)
+def test_cuda_target_clones_and_moves(uma, force_uma, monkeypatch, tiny_safetensors):
+    path, tensors = tiny_safetensors
+    force_uma(True)
+    fake_mu = _install_fake_modeling_utils(monkeypatch, safetensors.safe_open)
+    uma.patch_unified_memory_safetensors_load()
+    # device="cuda" IS intercepted -> tensors land on cuda, byte-identical.
+    with fake_mu.safe_open(str(path), framework = "pt", device = "cuda") as f:
+        for key, expected in tensors.items():
+            got = f.get_slice(key)[:]
+            assert got.device.type == "cuda"
+            assert torch.equal(got.cpu(), expected)
+            got_full = f.get_tensor(key)
+            assert got_full.device.type == "cuda"
+            assert torch.equal(got_full.cpu(), expected)
+
+
+@pytest.mark.skipif(
+    not (hasattr(torch, "cuda") and torch.cuda.is_available()),
+    reason = "needs a GPU for the low-memory fallback path",
+)
+def test_low_memory_falls_back_to_direct_move(uma, force_uma, monkeypatch, tiny_safetensors):
+    path, tensors = tiny_safetensors
+    force_uma(True)
+    fake_mu = _install_fake_modeling_utils(monkeypatch, safetensors.safe_open)
+    uma.patch_unified_memory_safetensors_load()
+    # Clone OOMs (transient CPU doubling on a constrained UMA box): the wrapper
+    # must fall back to the direct move and still succeed.
+    real_clone = torch.Tensor.clone
+
+    def _oom_clone(self, *a, **k):
+        raise RuntimeError("[enforce fail] not enough memory")
+
+    monkeypatch.setattr(torch.Tensor, "clone", _oom_clone)
+    try:
+        with fake_mu.safe_open(str(path), framework = "pt", device = "cuda") as f:
+            for key, expected in tensors.items():
+                got = f.get_slice(key)[:]
+                assert got.device.type == "cuda"
+                got_full = f.get_tensor(key)
+                assert got_full.device.type == "cuda"
+    finally:
+        monkeypatch.setattr(torch.Tensor, "clone", real_clone)
+    for key, expected in tensors.items():
+        with fake_mu.safe_open(str(path), framework = "pt", device = "cuda") as f:
+            assert torch.equal(f.get_tensor(key).cpu(), expected)
diff --git a/unsloth/models/_uma_safetensors.py b/unsloth/models/_uma_safetensors.py
new file mode 100644
index 0000000000..38d8b7d33a
--- /dev/null
+++ b/unsloth/models/_uma_safetensors.py
@@ -0,0 +1,169 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
+
+"""Faster safetensors weight loading on unified-memory (integrated) GPUs.
+
+A direct ``safe_open(..., device=<cuda>)`` on CUDA/HIP UMA GPUs (AMD APUs,
+NVIDIA GB10 Spark) misses torch's fast pinned-DMA path: the mmap-backed
+safetensors buffers aren't recognized, so it falls to a slow per-tensor copy
+with page faults. Cloning each tensor into a normal torch CPU allocation before
+moving it restores the fast path; outputs are bit-identical.
+
+CUDA/HIP only, and only for loads that pass a CUDA device to ``safe_open``
+directly: Intel XPU iGPUs and the CPU-open + later ``.to()`` flows (e.g. bnb /
+HQQ quantized loads) keep the stock path until they can be validated on real
+hardware.
+"""
+
+import os
+import functools
+
+import torch
+
+__all__ = [
+    "is_integrated_unified_memory_gpu",
+    "patch_unified_memory_safetensors_load",
+]
+
+
+@functools.lru_cache(maxsize = None)
+def is_integrated_unified_memory_gpu():
+    """True only when EVERY visible CUDA/HIP device is integrated (UMA).
+
+    Discrete and mixed discrete+iGPU boxes return False (pinned-DMA already
+    works there). Test override: ``UNSLOTH_FORCE_UMA=1`` / ``=0``.
+    """
+    _force = os.environ.get("UNSLOTH_FORCE_UMA")
+    if _force == "1":
+        return True
+    if _force == "0":
+        return False
+    try:
+        if not (hasattr(torch, "cuda") and torch.cuda.is_available()):
+            return False
+        count = torch.cuda.device_count()
+        if count == 0:
+            return False
+        for index in range(count):
+            props = torch.cuda.get_device_properties(index)
+            if not getattr(props, "is_integrated", 0):
+                return False
+        return True
+    except Exception:
+        return False
+
+
+def _is_cuda_target(device):
+    """Does a ``safe_open`` ``device=`` arg name a CUDA/HIP device?"""
+    if isinstance(device, bool):
+        return False
+    if isinstance(device, int):
+        return True
+    if isinstance(device, str):
+        return device == "cuda" or device.startswith("cuda:")
+    try:
+        return isinstance(device, torch.device) and device.type == "cuda"
+    except Exception:
+        return False
+
+
+def patch_unified_memory_safetensors_load():
+    """Wrap ``transformers.modeling_utils.safe_open`` so CUDA-target shard loads
+    open on CPU then clone+``.to(device)``, restoring the UMA fast path.
+
+    Gated to integrated GPUs (no-op on discrete/CPU/XPU/MLX), ``framework="pt"``
+    CUDA targets only, idempotent. Opt out: ``UNSLOTH_DISABLE_UMA_CLONE_LOAD=1``.
+
+    The gate runs lazily inside the wrapper, never here: probing device
+    properties at install would init CUDA during ``import unsloth`` -- breaking
+    fork multiprocessing and preempting ``patch_dgx_spark_memory_config``'s
+    allocator config. Returns ``True`` if the wrapper was installed.
+    """
+    if os.environ.get("UNSLOTH_DISABLE_UMA_CLONE_LOAD") == "1":
+        return False
+    try:
+        from transformers import modeling_utils as _mu
+    except Exception:
+        return False
+    real_safe_open = getattr(_mu, "safe_open", None)
+    if real_safe_open is None:
+        return False
+    if getattr(real_safe_open, "_unsloth_uma_clone", False):
+        return True
+
+    def _clone_move(tensor, device):
+        # Clone into a regular CPU allocation to restore fast pinned-DMA, then
+        # move. The clone transiently doubles the tensor's CPU footprint and can
+        # OOM a low-memory UMA box; fall back to the direct, allocation-free move
+        # (a genuine non-memory error re-raises identically from it).
+        try:
+            return tensor.clone().to(device, non_blocking = False)
+        except (MemoryError, RuntimeError):
+            return tensor.to(device, non_blocking = False)
+
+    class _ClonedSlice:
+        """Proxy over a safetensors ``PySafeSlice`` that clones+moves on read."""
+
+        __slots__ = ("_real", "_device")
+
+        def __init__(self, real, device):
+            self._real = real
+            self._device = device
+
+        def __getattr__(self, name):
+            if name in ("_real", "_device"):
+                raise AttributeError(name)
+            return getattr(self._real, name)
+
+        def __getitem__(self, key):
+            return _clone_move(self._real[key], self._device)
+
+    class _ClonedSafeOpen:
+        """Safetensors-handle proxy: load on CPU, clone+move tensors to CUDA."""
+
+        __slots__ = ("_real", "_device")
+
+        def __init__(self, args, kwargs):
+            self._device = kwargs.get("device", args[2] if len(args) > 2 else "cpu")
+            # Open on CPU; move ourselves.
+            if len(args) > 2:
+                args = args[:2] + ("cpu",) + tuple(args[3:])
+            else:
+                kwargs = dict(kwargs)
+                kwargs["device"] = "cpu"
+            self._real = real_safe_open(*args, **kwargs)
+
+        def __enter__(self):
+            self._real.__enter__()
+            return self
+
+        def __exit__(self, *exc):
+            return self._real.__exit__(*exc)
+
+        def __getattr__(self, name):
+            if name in ("_real", "_device"):
+                raise AttributeError(name)
+            return getattr(self._real, name)
+
+        def get_slice(self, name):
+            return _ClonedSlice(self._real.get_slice(name), self._device)
+
+        def get_tensor(self, name):
+            return _clone_move(self._real.get_tensor(name), self._device)
+
+    @functools.wraps(real_safe_open)
+    def _uma_safe_open(*args, **kwargs):
+        framework = kwargs.get("framework", args[1] if len(args) > 1 else None)
+        device = kwargs.get("device", args[2] if len(args) > 2 else "cpu")
+        # Device check first: non-CUDA loads must not trigger the CUDA-init gate.
+        if (
+            framework in ("pt", "pytorch")
+            and _is_cuda_target(device)
+            and is_integrated_unified_memory_gpu()
+        ):
+            return _ClonedSafeOpen(args, kwargs)
+        return real_safe_open(*args, **kwargs)
+
+    _uma_safe_open._unsloth_uma_clone = True
+    _mu.safe_open = _uma_safe_open
+    return True
diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index 57169fa3de..f9ac879de6 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -1670,6 +1670,13 @@ except:
 from transformers.modeling_utils import logger as transformers_logger
 
 
+# Faster safetensors loads on UMA (integrated) GPUs; lazy gate keeps this import
+# fork-safe (no CUDA init). No-op off-UMA. Opt out: UNSLOTH_DISABLE_UMA_CLONE_LOAD=1.
+from ._uma_safetensors import patch_unified_memory_safetensors_load
+
+patch_unified_memory_safetensors_load()
+
+
 def _all_missing_keys_are_position_ids(record_str):
     """True only when EVERY key in the 'newly initialized: [...]' list is a position_ids
     buffer.