Refine device operations into a unified DeviceManager and ARDevice abstraction

[lvliang-intel]

Description

Refine device operations into a unified DeviceManager / ARDevice abstraction so that (1) adding a new device is cheap and
(2) every device PyTorch supports works automatically with no per-device code.

a) Discovery is now data-driven. Backends are found via torch.accelerator, the PrivateUse1 registration, and a generic is_available() probe, and fall back to a preferred-order hint on older PyTorch. AR_DEVICE_BACKENDS= is an escape hatch for brand-new backends with zero code change.
b) Capability flags replace device-type branches. ARDevice now declares visible_devices_env_var, supports_pipeline_parallel, supports_multi_card_tuning, supports_empty_cache, supports_device_map_dispatch, and auto_select_by_default.
c) New APIs: register_ar_device() and get_active_device(), and a CudaARDevice/XpuARDevice handle for the built-in backends.
Route remaining torch.cuda.* call sites (hadamard, spinquant, calibration, convert_model) through the device manager.

Type of Change

Enhancement

Related Issues

#1788

Checklist Before Submitting

• My code has been tested locally.
• Documentation has been updated as needed.
• New or updated tests are included where applicable.
• The CUDA CI has passed. You can trigger it by commenting /azp run Unit-Test-CUDA-AutoRound.

[AutoRoundBot]

CI Failure Analysis Report (Unit-Test-AutoRound)

• Commit: 8460229
• Run: https://dev.azure.com/lpot-inc/neural-compressor/_build/results?buildId=75834
• Totals: 5 failed test(s), 3 cluster(s) (0 known / 3 new)

New Issues (3)

No.1 — 3 occurrence(s)

🔍 AssertionError: assert None == '0'

📝 Basic info

• Affected tests (3): TestSetCudaVisibleDevices.test_without_existing_cuda_visible_devices, TestSetCudaVisibleDevicesExtra.test_combined_with_spaces_in_index, TestSetCudaVisibleDevicesExtra.test_no_existing_env_var
• Logs (1): unittest_test_common_utils.log

🖥️ Log excerpt

unit/common/utils/test_device.py:486: in test_without_existing_cuda_visible_devices
    assert os.environ.get("CUDA_VISIBLE_DEVICES") == "0"
E   AssertionError: assert None == '0'
E    +  where None = get('CUDA_VISIBLE_DEVICES')
E    +    where get = environ({'UT_MODE': '0', 'HOSTNAME': 'e9b04053db06', 'FILTER_CHANGED_TESTS': 'true', 'COVERAGE_RCFILE': '/auto-round/.azure-pipelines/scripts/ut/coveragerc/cpu.coveragerc', 'FORCE_BF16': '1', 'PWD': '/auto-round/test', 'TZ': 'Asia/Shanghai', 'HOME': '/home/hostuser', 'LANG': 'C.UTF-8', 'VIRTUAL_ENV': '/home/hostuser/.venv', 'UV_LINK_MODE': 'copy', 'SYSTEM_JOBATTEMPT': '1', 'SYSTEM_PULLREQUEST_TARGETBRANCH': 'main', 'ZE_AFFINITY_MASK': '', 'NUMA_CPUSET': '0-15', 'TQDM_MININTERVAL': '120', 'UV_NO_PROGRESS': '1', 'BUILD_REASON': 'PullRequest', 'HF_HUB_DISABLE_PROGRESS_BARS': '1', 'SHLVL': '1', 'LD_LIBRARY_PATH': '/home/hostuser/.venv/lib/:', 'NUMA_NODE': '0', 'PATH': '/home/hostuser/.venv/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin', 'OLDPWD': '/auto-round', '_': '/usr/bin/numactl', 'PYTEST_VERSION': '9.1.1', 'TORCHINDUCTOR_CACHE_DIR': '/tmp/torchinductor_hostuser', 'KMP_DUPLICATE_LIB_OK': 'True', 'KMP_INIT_AT_FORK': 'FALSE', 'TOKENIZERS_PARALLELISM': 'false', 'CUBLAS_WORKSPACE_CONFIG': ':4096:8', 'PT_HPU_LAZY_MODE': '1', 'PYTEST_CURRENT_TEST': 'unit/common/utils/test_device.py::TestSetCudaVisibleDevices::test_without_existing_cuda_visible_devices (call)'}).get
E    +      where environ({'UT_MODE': '0', 'HOSTNAME': 'e9b04053db06', 'FILTER_CHANGED_TESTS': 'true', 'COVERAGE_RCFILE': '/auto-round/.azure-pipelines/scripts/ut/coveragerc/cpu.coveragerc', 'FORCE_BF16': '1', 'PWD': '/auto-round/test', 'TZ': 'Asia/Shanghai', 'HOME': '/home/hostuser', 'LANG': 'C.UTF-8', 'VIRTUAL_ENV': '/home/hostuser/.venv', 'UV_LINK_MODE': 'copy', 'SYSTEM_JOBATTEMPT': '1', 'SYSTEM_PULLREQUEST_TARGETBRANCH': 'main', 'ZE_AFFINITY_MASK': '', 'NUMA_CPUSET': '0-15', 'TQDM_MININTERVAL': '120', 'UV_NO_PROGRESS': '1', 'BUILD_REASON': 'PullRequest', 'HF_HUB_DISABLE_PROGRESS_BARS': '1', 'SHLVL': '1', 'LD_LIBRARY_PATH': '/home/hostuser/.venv/lib/:', 'NUMA_NODE': '0', 'PATH': '/home/hostuser/.venv/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin', 'OLDPWD': '/auto-round', '_': '/usr/bin/numactl', 'PYTEST_VERSION': '9.1.1', 'TORCHINDUCTOR_CACHE_DIR': '/tmp/torchinductor_hostuser', 'KMP_DUPLICATE_LIB_OK': 'True', 'KMP_INIT_AT_FORK': 'FALSE', 'TOKENIZERS_PARALLELISM': 'false', 'CUBLAS_WORKSPACE_CONFIG': ':4096:8', 'PT_HPU_LAZY_MODE': '1', 'PYTEST_CURRENT_TEST': 'unit/common/utils/test_device.py::TestSetCudaVisibleDevices::test_without_existing_cuda_visible_devices (call)'}) = os.environ

✨ AI analysis

• Category: Code Regression

• Confidence: high

• Root cause: The refactor changed set_cuda_visible_devices into an alias that does not explicitly select CUDA. For numeric inputs such as "0" or "5", set_visible_devices infers the current backend; in this CI environment that resolves to HPU or CPU, so CUDA_VISIBLE_DEVICES is never set.

• Suggested fix: Pass device_type="cuda" from the backward-compatible set_cuda_visible_devices wrapper so it preserves its original CUDA-specific behavior.

• Patch:

  diff --git a/auto_round/utils/device.py b/auto_round/utils/device.py
  --- a/auto_round/utils/device.py
  +++ b/auto_round/utils/device.py
  @@ -369,7 +369,7 @@ def set_visible_devices(device: str, device_type: Optional[str] = None) -> None:
   def set_cuda_visible_devices(device: str):
       """Backward-compatible alias of :func:`set_visible_devices`."""
  -    set_visible_devices(device)
  +    set_visible_devices(device, device_type="cuda")

No.2 — 1 occurrence(s)

🔍 AssertionError: assert '0,1,2,3' == '0,2'

📝 Basic info

• Affected tests (1): TestSetCudaVisibleDevices.test_multiple_devices
• Logs (1): unittest_test_common_utils.log

🖥️ Log excerpt

unit/common/utils/test_device.py:434: in test_multiple_devices
    assert os.environ.get("CUDA_VISIBLE_DEVICES") == "0,2"
E   AssertionError: assert '0,1,2,3' == '0,2'
E
E     - 0,2
E     + 0,1,2,3

✨ AI analysis

• Category: Code Regression

• Confidence: high

• Root cause: The PR replaced the CUDA-specific implementation with a generic set_visible_devices helper, but the backward-compatible set_cuda_visible_devices alias no longer forces device_type="cuda". On the CPU CI runner, numeric input resolves to the current CPU backend, whose visibility variable is unset, so the function becomes a no-op and does not remap or validate CUDA_VISIBLE_DEVICES.

• Suggested fix: Pass device_type="cuda" from set_cuda_visible_devices so the compatibility wrapper retains its original CUDA-specific behavior.

• Patch:

  diff --git a/auto_round/utils/device.py b/auto_round/utils/device.py
  --- a/auto_round/utils/device.py
  +++ b/auto_round/utils/device.py
  @@ -367,7 +367,7 @@ def set_visible_devices(device: str, device_type: Optional[str] = None) -> None:
   def set_cuda_visible_devices(device: str):
       """Backward-compatible alias of :func:`set_visible_devices`."""
  -    set_visible_devices(device)
  +    set_visible_devices(device, device_type="cuda")

No.3 — 1 occurrence(s)

🔍 Failed: DID NOT RAISE ValueError

📝 Basic info

• Affected tests (1): TestSetCudaVisibleDevices.test_invalid_device_index_raises
• Logs (1): unittest_test_common_utils.log

🖥️ Log excerpt

unit/common/utils/test_device.py:470: in test_invalid_device_index_raises
    with pytest.raises(ValueError):
         ^^^^^^^^^^^^^^^^^^^^^^^^^
E   Failed: DID NOT RAISE ValueError

✨ AI analysis

• Category: Code Regression

• Confidence: high

• Root cause: The PR changed set_cuda_visible_devices to delegate without specifying CUDA. On CPU CI, get_current_device_type() returns cpu, whose backend has no visibility environment variable, so the function returns without inspecting CUDA_VISIBLE_DEVICES and does not raise for index 5.

• Suggested fix: Preserve the backward-compatible CUDA-specific behavior by explicitly passing device_type="cuda" from set_cuda_visible_devices.

• Patch:

  diff --git a/auto_round/utils/device.py b/auto_round/utils/device.py
  --- a/auto_round/utils/device.py
  +++ b/auto_round/utils/device.py
  @@ -368,7 +368,7 @@
   def set_cuda_visible_devices(device: str):
       """Backward-compatible alias of :func:`set_visible_devices`."""
  -    set_visible_devices(device)
  +    set_visible_devices(device, device_type="cuda")

Notes

• Only top-3 issues receive AI analysis.
• To request an additional fix from Copilot, use "Quote reply" on the PR comment and @mention Copilot.

[AutoRoundBot]

CI Failure Analysis Report (Unit-Test-XPU-AutoRound)

• Commit: 8460229
• Run: https://dev.azure.com/lpot-inc/neural-compressor/_build/results?buildId=75840
• Totals: 0 failed test(s), 0 cluster(s) (0 known / 0 new)