[staging CI] unslothai/unsloth#9858

.github/workflows/staging-9858-macos-15.yml

@@ -0,0 +1,31 @@
+name: "staging-9858 macos-15"
+on:
+  push:
+    branches: ["pr-9858-xplat-ci"]
+    paths-ignore:
+      - '**/*.md'
+  workflow_dispatch:
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+permissions:
+  contents: read
+defaults:
+  run:
+    shell: bash
+jobs:
+  test:
+    runs-on: macos-15
+    timeout-minutes: 30
+    env:
+      UNSLOTH_COMPILE_DISABLE: '1'
+      UNSLOTH_IS_PRESENT: '1'
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          persist-credentials: false
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.12'
+      - run: python -m pip install -r studio/backend/requirements/studio.txt pytest pytest-asyncio pytest-timeout python-multipart
+      - run: rc=0; PYTHONPATH=studio/backend python -m pytest studio/backend/tests/test_torch_cpu_build_on_nvidia_host.py tests/python/test_install_python_stack.py tests/studio/install/test_cuda_repair.py tests/studio/install/test_gpu_detection_followups.py tests/studio/install/test_pr5940_followups.py tests/studio/install/test_rocm_arch_table_parity.py tests/studio/install/test_rocm_support.py tests/studio/install/test_windows_torch_flavor_invariant.py tests/studio/test_xpu_triton_swap.py -q --tb=short -k 'not llama_cpp_load_progress_live and not TestGpuAutoSelection and not TestPreSpawnGpuResolution and not TestPerGpuFitGuardAllCounts and not TestTransformersIntrospection and not test_returns_cuda_when_cuda_available and not test_calls_cuda_cache_when_cuda' > pytest_out.txt 2>&1 || rc=$?; cat pytest_out.txt; if [ "$rc" = "5" ]; then echo "no tests ran (deps absent on this runner)"; rc=0; fi; if [ "$rc" = "2" ] && grep -qE "No module named .(torch|unsloth_zoo|transformers)." pytest_out.txt && ! grep -qE "^(FAILED|ERROR) " pytest_out.txt; then echo "collection needs a dep this runner does not ship"; rc=0; fi; if [ "$rc" = "1" ] && ! grep -qE "^ERROR " pytest_out.txt; then tot=$(grep -cE "^FAILED " pytest_out.txt); dep=$(grep -cE "^FAILED .* No module named .(torch|unsloth_zoo|transformers).$" pytest_out.txt); if [ "$tot" -gt 0 ] && [ "$tot" = "$dep" ]; then echo "only tests needing a dep this runner does not ship failed"; rc=0; fi; fi; exit "$rc"

.github/workflows/staging-9858-ubuntu-latest.yml

@@ -0,0 +1,31 @@
+name: "staging-9858 ubuntu-latest"
+on:
+  push:
+    branches: ["pr-9858-xplat-ci"]
+    paths-ignore:
+      - '**/*.md'
+  workflow_dispatch:
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+permissions:
+  contents: read
+defaults:
+  run:
+    shell: bash
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    env:
+      UNSLOTH_COMPILE_DISABLE: '1'
+      UNSLOTH_IS_PRESENT: '1'
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          persist-credentials: false
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.12'
+      - run: python -m pip install -r studio/backend/requirements/studio.txt pytest pytest-asyncio pytest-timeout python-multipart
+      - run: rc=0; PYTHONPATH=studio/backend python -m pytest studio/backend/tests/test_torch_cpu_build_on_nvidia_host.py tests/python/test_install_python_stack.py tests/studio/install/test_cuda_repair.py tests/studio/install/test_gpu_detection_followups.py tests/studio/install/test_pr5940_followups.py tests/studio/install/test_rocm_arch_table_parity.py tests/studio/install/test_rocm_support.py tests/studio/install/test_windows_torch_flavor_invariant.py tests/studio/test_xpu_triton_swap.py -q --tb=short -k 'not llama_cpp_load_progress_live and not TestGpuAutoSelection and not TestPreSpawnGpuResolution and not TestPerGpuFitGuardAllCounts and not TestTransformersIntrospection and not test_returns_cuda_when_cuda_available and not test_calls_cuda_cache_when_cuda' > pytest_out.txt 2>&1 || rc=$?; cat pytest_out.txt; if [ "$rc" = "5" ]; then echo "no tests ran (deps absent on this runner)"; rc=0; fi; if [ "$rc" = "2" ] && grep -qE "No module named .(torch|unsloth_zoo|transformers)." pytest_out.txt && ! grep -qE "^(FAILED|ERROR) " pytest_out.txt; then echo "collection needs a dep this runner does not ship"; rc=0; fi; if [ "$rc" = "1" ] && ! grep -qE "^ERROR " pytest_out.txt; then tot=$(grep -cE "^FAILED " pytest_out.txt); dep=$(grep -cE "^FAILED .* No module named .(torch|unsloth_zoo|transformers).$" pytest_out.txt); if [ "$tot" -gt 0 ] && [ "$tot" = "$dep" ]; then echo "only tests needing a dep this runner does not ship failed"; rc=0; fi; fi; exit "$rc"

.github/workflows/staging-9858-windows-latest.yml

@@ -0,0 +1,31 @@
+name: "staging-9858 windows-latest"
+on:
+  push:
+    branches: ["pr-9858-xplat-ci"]
+    paths-ignore:
+      - '**/*.md'
+  workflow_dispatch:
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+permissions:
+  contents: read
+defaults:
+  run:
+    shell: bash
+jobs:
+  test:
+    runs-on: windows-latest
+    timeout-minutes: 30
+    env:
+      UNSLOTH_COMPILE_DISABLE: '1'
+      UNSLOTH_IS_PRESENT: '1'
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          persist-credentials: false
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.12'
+      - run: python -m pip install -r studio/backend/requirements/studio.txt pytest pytest-asyncio pytest-timeout python-multipart
+      - run: rc=0; PYTHONPATH=studio/backend python -m pytest studio/backend/tests/test_torch_cpu_build_on_nvidia_host.py tests/python/test_install_python_stack.py tests/studio/install/test_cuda_repair.py tests/studio/install/test_gpu_detection_followups.py tests/studio/install/test_pr5940_followups.py tests/studio/install/test_rocm_arch_table_parity.py tests/studio/install/test_rocm_support.py tests/studio/install/test_windows_torch_flavor_invariant.py tests/studio/test_xpu_triton_swap.py -q --tb=short -k 'not llama_cpp_load_progress_live and not TestGpuAutoSelection and not TestPreSpawnGpuResolution and not TestPerGpuFitGuardAllCounts and not TestTransformersIntrospection and not test_returns_cuda_when_cuda_available and not test_calls_cuda_cache_when_cuda' > pytest_out.txt 2>&1 || rc=$?; cat pytest_out.txt; if [ "$rc" = "5" ]; then echo "no tests ran (deps absent on this runner)"; rc=0; fi; if [ "$rc" = "2" ] && grep -qE "No module named .(torch|unsloth_zoo|transformers)." pytest_out.txt && ! grep -qE "^(FAILED|ERROR) " pytest_out.txt; then echo "collection needs a dep this runner does not ship"; rc=0; fi; if [ "$rc" = "1" ] && ! grep -qE "^ERROR " pytest_out.txt; then tot=$(grep -cE "^FAILED " pytest_out.txt); dep=$(grep -cE "^FAILED .* No module named .(torch|unsloth_zoo|transformers).$" pytest_out.txt); if [ "$tot" -gt 0 ] && [ "$tot" = "$dep" ]; then echo "only tests needing a dep this runner does not ship failed"; rc=0; fi; fi; exit "$rc"

install.sh

@@ -4559,6 +4559,15 @@ while [ -n "$_torch_index_leaf" ] && [ "${_torch_index_leaf%/}" != "$_torch_inde
 done
 _torch_index_leaf="${_torch_index_leaf##*/}"
 _torch_index_leaf=$(printf '%s' "$_torch_index_leaf" | tr '[:upper:]' '[:lower:]')
+# Whether the caller had already STATED a backend before the assignment below overwrites it.
+# setup.sh documents UNSLOTH_TORCH_BACKEND=cpu as the way to keep a deliberate CPU install,
+# and on a GPU-less host the resolved value is cpu too, so without this the manifest cannot
+# tell a stated choice from the automatic answer.
+if [ -n "${UNSLOTH_TORCH_BACKEND:-}" ]; then
+    _torch_backend_was_stated=true
+else
+    _torch_backend_was_stated=false
+fi
 case "$_torch_index_leaf" in
     rocm*|gfx*) export UNSLOTH_TORCH_BACKEND="rocm" ;;
     cpu)        export UNSLOTH_TORCH_BACKEND="cpu"  ;;
@@ -4568,6 +4577,15 @@ case "$_torch_index_leaf" in
     *)          unset UNSLOTH_TORCH_BACKEND ;;
 esac
 
+# Derived from the index this script RESOLVED, which on a GPU-less machine is "cpu" whether
+# or not anyone asked. Without the marker every ordinary Linux CPU install is recorded as a
+# deliberate choice, and a machine that later gains a GPU is never offered the repair.
+if [ -n "${UNSLOTH_TORCH_BACKEND:-}" ] && [ "$_torch_backend_was_stated" != true ]; then
+    export UNSLOTH_TORCH_BACKEND_SOURCE="resolved"
+else
+    unset UNSLOTH_TORCH_BACKEND_SOURCE
+fi
+
 # Whether TORCH_INDEX_URL names an actual pip ROCm family (rocm<digit>* / gfx*), gating the
 # ROCm-only side effects below (AMD bitsandbytes, ROCm-torch repair). Digit-gated so a leaf
 # merely STARTING with "rocm" isn't force-repaired from the wrong path.

studio/backend/core/inference/llama_cpp.py

@@ -10610,6 +10610,10 @@ def _probe_or_none():
     # only when an axis is actually quantized.
     _tensor_quant_kv_unsupported_binaries: set[tuple[str, int]] = set()
 
+    # Binary dirs already reported by _warn_missing_windows_cuda_runtime. The env is rebuilt
+    # for every launch and every --list-devices probe, so one line per binary is enough.
+    _missing_cuda_runtime_warned: set[str] = set()
+
     @classmethod
     def _binary_key(cls, binary: Optional[str]) -> Optional[tuple[str, int]]:
         """(path, mtime_ns); ns mtime re-probes a same-second binary swap."""
@@ -10733,6 +10737,47 @@ def _add(path: Path) -> None:
         _add(site_packages / "torch" / "lib")
         return out
 
+    @classmethod
+    def _warn_missing_windows_cuda_runtime(cls, binary_dir: str, path_dirs: list[str]) -> None:
+        """Say so when a CUDA llama-server has no cudart to load. Diagnostic only.
+
+        The CUDA prebuilt links ``cudart64_*.dll`` / ``cublas64_*.dll`` and takes them
+        from the managed venv -- ``torch/lib`` or the ``nvidia/*`` wheels, per
+        _windows_pip_nvidia_dll_dirs. A 2.11.0+cpu torch ships neither, so ggml cannot
+        load its CUDA backend and ``llama-server.exe --list-devices`` prints
+        ``Available devices: (none)`` while UNSLOTH_PREBUILT_INFO.json still says
+        ``backend cuda`` (#8473, HF discussion 87). Today that is entirely silent.
+
+        Changes nothing about the launch: the process still starts, still falls back to
+        CPU, and a custom build with the DLLs somewhere else is not second-guessed.
+        Never raises -- a diagnostic must not be able to stop a load.
+        """
+        try:
+            if binary_dir in cls._missing_cuda_runtime_warned:
+                return
+            # Same identification _installed_ggml_backends uses: the official prebuilts are
+            # single-backend, so the ggml CUDA lib beside llama-server IS the build.
+            ggml_cuda = os.path.join(binary_dir, "ggml-cuda.dll")
+            if not os.path.isfile(ggml_cuda):
+                return
+            for directory in path_dirs:
+                try:
+                    names = os.listdir(directory)
+                except OSError:
+                    continue
+                if any(name.lower().startswith("cudart64_") for name in names):
+                    return
+            cls._missing_cuda_runtime_warned.add(binary_dir)
+            logger.warning(
+                "llama.cpp is the CUDA build (%s) but no cudart64_*.dll was found on its "
+                "DLL search path. The CUDA ggml backend will not load and llama-server "
+                "will report no devices. This is what a CPU-only PyTorch in the managed "
+                "environment looks like; repair the installation to restore GPU support.",
+                ggml_cuda,
+            )
+        except Exception as e:
+            logger.debug(f"CUDA runtime DLL diagnostic failed: {e}")
+
     @staticmethod
     def _build_windows_path_dirs(binary_dir: str, prefix: str, cuda_path: str) -> list[str]:
         """Ordered PATH entries prepended so llama-server.exe resolves cudart /
@@ -10771,6 +10816,13 @@ def _llama_server_env_for_binary(
             )
             existing_path = env.get("PATH", "")
             env["PATH"] = ";".join(path_dirs) + ";" + existing_path
+            # Warn against the FULL search path, inherited entries included: a hand-installed CUDA
+            # toolkit puts cudart64_*.dll on PATH without the venv or CUDA_PATH knowing, and warning
+            # on the prepended directories alone told working custom setups to repair a fine install.
+            LlamaCppBackend._warn_missing_windows_cuda_runtime(
+                binary_dir,
+                path_dirs + [d for d in existing_path.split(";") if d],
+            )
 
             # ROCm: the prebuilt bundles rocblas.dll but NOT the Tensile
             # kernel files (rocblas/library/*.dat + *.hsaco); the DLL searches

studio/backend/main.py

@@ -1561,11 +1561,15 @@ def _hardware_snapshot() -> Optional[tuple[bool, Optional[str], Optional[str]]]:
         generation = _hw_module.DETECTION_GENERATION
         device = _hw_module.DEVICE
         chat_only = bool(_hw_module.CHAT_ONLY)
-        reason = getattr(_hw_module, "CHAT_ONLY_REASON", None)
-        # Inside the guarded read, with the reason it belongs to. Read after it, a forced
-        # re-detect starting in between would pair this reply's reason with a detail from
-        # a different pass, or with none at all.
-        detail = getattr(_hw_module, "CHAT_ONLY_DETAIL", None)
+        # Refreshed, not the frozen global: the three inventory-sensitive verdicts can change
+        # after startup (an eGPU attached, a driver that finished restarting). Reason and detail
+        # come back together, or a forced re-detect starting in between would pair this reply's
+        # reason with a detail from a different pass.
+        try:
+            reason, detail = _hw_module.current_chat_only_verdict()
+        except Exception:
+            reason = getattr(_hw_module, "CHAT_ONLY_REASON", None)
+            detail = getattr(_hw_module, "CHAT_ONLY_DETAIL", None)
         if (
             device is not None
             and _hw_module.DETECTION_COMPLETE.is_set()
@@ -2057,8 +2061,9 @@ def _get_cached_system_gpu_info(logger) -> tuple[dict[str, Any], dict[str, Any]]
             logger.debug(f"Could not resolve gpu_ids support: {e}")
             llama_uses_vulkan = False
             gpu_ids_supported = True
-        # Preserve backend/index metadata from the visibility probe: a CPU training host can expose
-        # a Vulkan inference GPU, and the UI must label it Vulkan, not the top-level CPU backend.
+        # The spread also carries `physical_devices` and `mismatch`: GPUs the OS sees that this PyTorch
+        # cannot open (#8473). They stay their own fields, because `devices` below is the runtime-usable
+        # list that model fit budgets against and the training device picker pins from.
         gpu_info = {
             **visibility_info,
             "available": visibility_info.get("available", False),

studio/backend/tests/conftest.py

@@ -942,3 +942,28 @@ def real_prequant_safe_globals(monkeypatch):
     monkeypatch.setattr(pq, "_SAFE_GLOBALS_REGISTERED", None)
     monkeypatch.setattr(pq, "_RESOLVED_SAFE_GLOBALS", set())
     return resolver
+
+
+@pytest.fixture(autouse = True)
+def _no_carried_over_hardware_measurements():
+    """Both hardware caches start empty for every test, as they do in a fresh process.
+
+    The torch build snapshot and the physical GPU inventory are module globals with a
+    60 second TTL, so one test's host -- a suite that makes `import torch` fail, say --
+    would otherwise answer for every test that ran within a minute of it. Cleared
+    afterwards as well, so a test that warms one deliberately does not leak either.
+    """
+    from utils.hardware import hardware as _hw
+
+    def _clear():
+        # Under the locks: a non-blocking read hands the refresh to a daemon thread that holds
+        # these while it writes, so clearing without waiting lets a previous test's REAL host land
+        # in the cache a moment later. Torch lock FIRST, then the inventory lock, because that is
+        # the order the background refresh takes them in.
+        with _hw._torch_build_snapshot_lock, _hw._physical_gpu_inventory_lock:
+            _hw._torch_build_snapshot_cache = None
+            _hw._physical_gpu_inventory_cache = None
+
+    _clear()
+    yield
+    _clear()