Use logical NPU 0 in Ascend tests

Author: Qubitium

.github/scripts/ci_tests.py

@@ -40,9 +40,6 @@ def configure_ascend_npu_test_env(env: dict[str, str], test_script: str) -> bool
         visible_devices = env.get("GPTQMODEL_TEST_ASCEND_RT_VISIBLE_DEVICES", DEFAULT_ASCEND_RT_VISIBLE_DEVICES)
         env["ASCEND_RT_VISIBLE_DEVICES"] = visible_devices
 
-    if "GPTQMODEL_TEST_NPU_DEVICE" not in env and visible_devices and "," not in visible_devices:
-        env["GPTQMODEL_TEST_NPU_DEVICE"] = "npu:0"
-
     env["CUDA_VISIBLE_DEVICES"] = ""
     return True
 
@@ -136,7 +133,6 @@ def run_tests(args: argparse.Namespace) -> int:
     ascend_npu_test = configure_ascend_npu_test_env(env, args.test_script)
     if ascend_npu_test:
         print(f"ASCEND_RT_VISIBLE_DEVICES={env.get('ASCEND_RT_VISIBLE_DEVICES', '')}")
-        print(f"GPTQMODEL_TEST_NPU_DEVICE={env.get('GPTQMODEL_TEST_NPU_DEVICE', '')}")
 
     if args.xpu_mode:
         maybe_uninstall_vllm()

tests/test_npu_linalg.py

@@ -10,19 +10,11 @@
 import torch
 
 from gptqmodel.quantization.npu_linalg import npu_inverse_cholesky_factor
-from gptqmodel.utils.torch import HAS_NPU, last_npu_device_by_pci_bus_order
+from gptqmodel.utils.torch import HAS_NPU
 
 
 pytestmark = pytest.mark.skipif(not HAS_NPU, reason="Ascend NPU is required")
-DEFAULT_ASCEND_RT_VISIBLE_DEVICES = "7"
-
-
-def _default_npu_test_device() -> str:
-    selected = last_npu_device_by_pci_bus_order()
-    return str(selected) if selected is not None else "npu:0"
-
-
-NPU_TEST_DEVICE = os.environ.get("GPTQMODEL_TEST_NPU_DEVICE", _default_npu_test_device())
+NPU_TEST_DEVICE = "npu:0"
 
 
 def _test_npu_device() -> torch.device:
@@ -32,17 +24,6 @@ def _test_npu_device() -> torch.device:
     return device
 
 
-def _default_subprocess_env() -> dict[str, str]:
-    env = os.environ.copy()
-    visible_devices = env.get("ASCEND_RT_VISIBLE_DEVICES", "").strip()
-    if not visible_devices:
-        visible_devices = env.get("GPTQMODEL_TEST_ASCEND_RT_VISIBLE_DEVICES", DEFAULT_ASCEND_RT_VISIBLE_DEVICES)
-        env["ASCEND_RT_VISIBLE_DEVICES"] = visible_devices
-    if "GPTQMODEL_TEST_NPU_DEVICE" not in env and visible_devices and "," not in visible_devices:
-        env["GPTQMODEL_TEST_NPU_DEVICE"] = "npu:0"
-    return env
-
-
 def _spd_matrix(size: int, seed: int) -> torch.Tensor:
     generator = torch.Generator(device="cpu").manual_seed(seed)
     values = torch.randn(size, size, generator=generator, dtype=torch.float32)
@@ -82,7 +63,6 @@ def test_npu_inverse_cholesky_factor_rejects_non_positive_definite_matrix():
 def test_gptq_npu_hessian_inverse_avoids_torch_npu_cpu_fallback_warnings():
     script = textwrap.dedent(
         """
-        import os
         import torch
         import torch.nn as nn
         from gptqmodel.quantization.config import QuantizeConfig
@@ -92,7 +72,7 @@ def test_gptq_npu_hessian_inverse_avoids_torch_npu_cpu_fallback_warnings():
         if not HAS_NPU:
             raise RuntimeError("Ascend NPU is not available")
 
-        npu_test_device = os.environ.get("GPTQMODEL_TEST_NPU_DEVICE", "npu:0")
+        npu_test_device = "npu:0"
         torch.npu.set_device(npu_test_device)
         torch.manual_seed(0)
 
@@ -119,11 +99,10 @@ def test_gptq_npu_hessian_inverse_avoids_torch_npu_cpu_fallback_warnings():
         """
     )
 
-    env = _default_subprocess_env()
     proc = subprocess.run(
         [sys.executable, "-c", script],
         cwd=os.getcwd(),
-        env=env,
+        env=os.environ.copy(),
         text=True,
         capture_output=True,
         timeout=60,

tests/test_npu_support.py

@@ -1,5 +1,4 @@
 import copy
-import os
 import sys
 import warnings
 
@@ -29,12 +28,7 @@
 from gptqmodel.utils.torch import HAS_NPU, last_npu_device_by_pci_bus_order
 
 
-def _default_npu_test_device() -> str:
-    selected = last_npu_device_by_pci_bus_order()
-    return str(selected) if selected is not None else "npu:0"
-
-
-NPU_TEST_DEVICE = os.environ.get("GPTQMODEL_TEST_NPU_DEVICE", _default_npu_test_device())
+NPU_TEST_DEVICE = "npu:0"
 NPU_CPU_FALLBACK_MARKERS = (
     "not currently supported on the NPU backend",
     "fall back to run on the CPU",