Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Unsloth Studio (AMD/Windows): import PyTorch before Triton to avoid WinError 126 (#10932)

boobcactusf96409d08da7
diff --git a/studio/backend/core/_msvc_env.py b/studio/backend/core/_msvc_env.py
index eaa538855eb..7b2d375fb09 100644
--- a/studio/backend/core/_msvc_env.py
+++ b/studio/backend/core/_msvc_env.py
@@ -276,9 +276,25 @@ def _triton_windows_spec() -> str:
return "triton-windows<3.7"
+def _torch_is_rocm_build() -> bool:
+ # Off disk, no torch import: AMD's untagged Windows wheel (2.8.0a0+git...) is only told by version.py's hip field.
+ try:
+ from core._torchao_stub import _installed_torch_is_rocm # noqa: PLC0415
+ return _installed_torch_is_rocm() is True
+ except Exception: # noqa: BLE001
+ logger.debug("Could not read the installed torch's ROCm signals", exc_info = True)
+ return False
+
+
def gate_torch_compile_on_windows(log: logging.Logger) -> None:
if sys.platform != "win32":
return
+ if _torch_is_rocm_build():
+ try:
+ # triton-windows >= 3.8 imported first loads System32 amdhip64_7.dll, then torch_hip.dll fails with WinError 126.
+ import torch # noqa: F401, PLC0415
+ except Exception: # noqa: BLE001
+ logger.debug("The torch preload failed; probing triton anyway", exc_info = True)
try:
import triton # noqa: F401, PLC0415
except ImportError:
diff --git a/studio/backend/tests/test_msvc_env_7595.py b/studio/backend/tests/test_msvc_env_7595.py
index 9a0f4d189cc..c78dcbe4130 100644
--- a/studio/backend/tests/test_msvc_env_7595.py
+++ b/studio/backend/tests/test_msvc_env_7595.py
@@ -8,10 +8,13 @@ import sys
import types
import logging
+import pytest
+
_backend = os.path.join(os.path.dirname(__file__), "..")
sys.path.insert(0, _backend)
from core import _msvc_env
+import core._torchao_stub as torchao_stub
def _fake_triton(
@@ -182,9 +185,15 @@ def test_unreachable_discovery_does_not_gate_a_visual_studio_box(tmp_path, monke
assert _msvc_env.crt_headers_reachable() is True
+def _no_compile_probe(monkeypatch):
+ # These judge the discovery fallback; a real clang-cl on PATH (GitHub's Windows image) would answer first.
+ monkeypatch.setattr(_msvc_env, "_compiles_a_trivial_translation_unit", lambda cc, d: None)
+
+
def test_a_search_that_ran_and_found_nothing_still_gates(tmp_path, monkeypatch):
"""The other half: without this, the #7595 crash is no longer prevented at all."""
monkeypatch.setattr(sys, "platform", "win32")
+ _no_compile_probe(monkeypatch)
monkeypatch.delenv("INCLUDE", raising = False)
_fake_triton(monkeypatch, [], cc = "clang-cl.exe")
assert _msvc_env.crt_headers_reachable() is False
@@ -241,6 +250,7 @@ def test_tinycc_is_not_gated_on_a_release_without_is_clang_cl(tmp_path, monkeypa
def test_clang_cl_is_still_gated_without_is_clang_cl(tmp_path, monkeypatch):
"""The other half: dropping the predicate must not drop the detection."""
monkeypatch.setattr(sys, "platform", "win32")
+ _no_compile_probe(monkeypatch)
monkeypatch.delenv("INCLUDE", raising = False)
_fake_triton(monkeypatch, [], cc = "clang-cl.exe", with_is_clang_cl = False)
assert _msvc_env._needs_msvc_headers() is True
@@ -286,6 +296,7 @@ def test_the_get_cc_rename_is_followed(tmp_path, monkeypatch):
def test_the_get_cc_rename_still_gates_clang_cl(tmp_path, monkeypatch):
monkeypatch.setattr(sys, "platform", "win32")
+ _no_compile_probe(monkeypatch)
monkeypatch.delenv("INCLUDE", raising = False)
_fake_triton_38(monkeypatch, [], cc = "clang-cl.exe")
assert _msvc_env._needs_msvc_headers() is True
@@ -353,6 +364,7 @@ def test_a_split_toolchain_is_judged_over_the_union(tmp_path, monkeypatch):
def test_the_union_still_gates_when_neither_half_is_there(tmp_path, monkeypatch):
"""The other direction: a union that is still incomplete must gate, or #7595 is unprevented."""
monkeypatch.setattr(sys, "platform", "win32")
+ _no_compile_probe(monkeypatch)
ucrt = tmp_path / "ucrt"
ucrt.mkdir()
(ucrt / "stdlib.h").write_text("/* stub */")
@@ -474,6 +486,7 @@ def test_gate_survives_a_probe_that_raises(monkeypatch):
monkeypatch.setenv("TORCHDYNAMO_DISABLE", "")
monkeypatch.delenv("TORCHDYNAMO_DISABLE")
monkeypatch.setitem(sys.modules, "triton", types.ModuleType("triton"))
+ monkeypatch.setattr(_msvc_env, "_torch_is_rocm_build", lambda: False)
def boom():
raise RuntimeError("sysconfig has no platlib on this scheme")
@@ -492,6 +505,8 @@ def _gate(monkeypatch, *, triton_importable, headers_ok):
monkeypatch.setitem(
sys.modules, "triton", types.ModuleType("triton") if triton_importable else None
)
+ # Live, this would read the test host's own torch and answer for the wrong machine.
+ monkeypatch.setattr(_msvc_env, "_torch_is_rocm_build", lambda: False)
monkeypatch.setattr(_msvc_env, "crt_headers_reachable", lambda: headers_ok)
records = []
logger = logging.getLogger("test_gate_7595")
@@ -510,6 +525,11 @@ def test_gate_is_noop_off_win32(monkeypatch):
"crt_headers_reachable",
lambda: (_ for _ in ()).throw(AssertionError("gate ran off win32")),
)
+ monkeypatch.setattr(
+ _msvc_env,
+ "_torch_is_rocm_build",
+ lambda: (_ for _ in ()).throw(AssertionError("ROCm check ran off win32")),
+ )
_msvc_env.gate_torch_compile_on_windows(logging.getLogger("test_gate_7595"))
assert "TORCHDYNAMO_DISABLE" not in os.environ
@@ -538,6 +558,7 @@ def test_gate_does_not_disable_where_triton_already_compiles(monkeypatch, tmp_pa
monkeypatch.setenv("TORCHDYNAMO_DISABLE", "")
monkeypatch.delenv("TORCHDYNAMO_DISABLE")
monkeypatch.delenv("INCLUDE", raising = False)
+ monkeypatch.setattr(_msvc_env, "_torch_is_rocm_build", lambda: False)
_fake_triton(monkeypatch, _sdk_dirs(tmp_path, with_toolset = True))
_msvc_env.gate_torch_compile_on_windows(logging.getLogger("test_gate_7595"))
@@ -548,6 +569,8 @@ def _no_headers(monkeypatch, tcc, builds):
monkeypatch.setattr(_msvc_env, "_crt_headers_reachable", lambda: False)
monkeypatch.setattr(_msvc_env, "_bundled_tcc", lambda: tcc)
monkeypatch.setattr(_msvc_env, "_triton_driver_builds", lambda: builds)
+ # The real one reaches the host's triton, which on Apple Silicon is a stub that raises on cache_clear.
+ monkeypatch.setattr(_msvc_env, "_clear_compiler_cache", lambda: None)
# setenv first so teardown removes a CC the code under test writes straight into os.environ.
monkeypatch.setenv("CC", "")
monkeypatch.delenv("CC")
@@ -577,3 +600,97 @@ def test_an_explicit_cc_is_never_replaced(monkeypatch):
monkeypatch.setenv("CC", r"C:\llvm\bin\clang-cl.exe")
assert _msvc_env.crt_headers_reachable() is False
assert os.environ["CC"] == r"C:\llvm\bin\clang-cl.exe"
+
+
+def test_torch_is_rocm_build_reads_the_wheel_version(monkeypatch):
+ import importlib.metadata as md
+
+ monkeypatch.setattr(md, "version", lambda dist: "2.11.0+rocm7.13.0")
+ assert _msvc_env._torch_is_rocm_build() is True
+ # Stubbed: left real, the hip field would be read off the test host's own torch.
+ monkeypatch.setattr(md, "version", lambda dist: "2.11.0+cu128")
+ monkeypatch.setattr(torchao_stub, "_hip_field_is_set", lambda: False)
+ assert _msvc_env._torch_is_rocm_build() is False
+
+ def boom(dist):
+ raise RuntimeError("metadata is unreadable")
+
+ monkeypatch.setattr(md, "version", boom)
+ assert _msvc_env._torch_is_rocm_build() is False
+
+
+def _fake_torch_version_py(tmp_path, monkeypatch, version_text):
+ # find_spec is patched, not sys.path, so a real torch on the host is never imported.
+ import importlib.util
+
+ pkg = tmp_path / "torch"
+ pkg.mkdir()
+ init = pkg / "__init__.py"
+ init.write_text("")
+ (pkg / "version.py").write_text(version_text)
+ spec = importlib.util.spec_from_file_location("torch", str(init))
+ real_find_spec = importlib.util.find_spec
+ monkeypatch.setattr(
+ importlib.util,
+ "find_spec",
+ lambda name, *a, **k: spec if name == "torch" else real_find_spec(name, *a, **k),
+ )
+
+
+ "version, hip, expected",
+ [("2.8.0a0+gitfc14c65", '"6.2.4"', True), ("2.11.0+cu128", "None", False)],
+)
+def test_torch_is_rocm_build_reads_the_hip_field(tmp_path, monkeypatch, version, hip, expected):
+ # AMD's Windows torch 2.8 carries no rocm tag; only version.py's hip field identifies it.
+ import importlib.metadata as md
+
+ _fake_torch_version_py(
+ tmp_path, monkeypatch, f'__version__ = "{version}"\nhip: Optional[str] = {hip}\n'
+ )
+ monkeypatch.setattr(md, "version", lambda dist: version)
+ assert _msvc_env._torch_is_rocm_build() is expected
+
+
+def _gate_imports(
+ monkeypatch,
+ *,
+ rocm,
+ torch_raises = False,
+):
+ import builtins
+
+ monkeypatch.setattr(sys, "platform", "win32")
+ monkeypatch.setenv("TORCHDYNAMO_DISABLE", "")
+ monkeypatch.delenv("TORCHDYNAMO_DISABLE")
+ monkeypatch.setattr(_msvc_env, "_torch_is_rocm_build", lambda: rocm)
+ monkeypatch.setattr(_msvc_env, "crt_headers_reachable", lambda: True)
+ imported = []
+ real_import = builtins.__import__
+
+ # `import torch` of an already-imported module is silent, so the import call itself is recorded.
+ def recording(name, *args, **kwargs):
+ if name in ("torch", "triton"):
+ imported.append(name)
+ if name == "torch" and torch_raises:
+ raise RuntimeError("torch_hip.dll: WinError 126")
+ return types.ModuleType(name)
+ return real_import(name, *args, **kwargs)
+
+ monkeypatch.setattr(builtins, "__import__", recording)
+ _msvc_env.gate_torch_compile_on_windows(logging.getLogger("test_gate_7595"))
+ assert "TORCHDYNAMO_DISABLE" not in os.environ
+ return imported
+
+
+def test_gate_imports_torch_before_triton_on_a_rocm_build(monkeypatch):
+ # Measured on gfx1151 Windows, torch 2.11.0+rocm7.13.0 + triton-windows 3.8: triton first -> WinError 126.
+ assert _gate_imports(monkeypatch, rocm = True) == ["torch", "triton"]
+
+
+def test_gate_does_not_preload_torch_on_a_cuda_or_cpu_build(monkeypatch):
+ assert _gate_imports(monkeypatch, rocm = False) == ["triton"]
+
+
+def test_gate_survives_the_torch_preload_raising(monkeypatch):
+ assert _gate_imports(monkeypatch, rocm = True, torch_raises = True) == ["torch", "triton"]