借助bos 实现跨设备精度对比

engineV2.py

@@ -29,6 +29,7 @@
         APITestPaddleTorchGPUPerformance,
         APITestAccuracyStable,
         APITestCustomDeviceVSCPU,
+        APITestPaddleDeviceVSGPU,
     )
     import torch
     import paddle
@@ -38,7 +39,19 @@
 os.environ["FLAGS_use_system_allocator"] = "1"
 os.environ["NVIDIA_TF32_OVERRIDE"] = "0"
 
-VALID_TEST_ARGS = {"test_amp", "test_backward", "atol", "rtol", "test_tol"}
+VALID_TEST_ARGS = {
+    "test_amp",
+    "test_backward",
+    "atol",
+    "rtol",
+    "test_tol",
+    "operation_mode",
+    "bos_path",
+    "target_device_type",
+    "random_seed",
+    "bos_conf_path",
+    "bcecmd_path",
+}
 
 DEVICE_TYPE = None
 DEVICE_TYPE_DETECTED = False
@@ -122,7 +135,7 @@ def detect_device_type() -> str:
         try:
             out = subprocess.check_output(["ixsmi"], text=True, stderr=subprocess.STDOUT)
             if any(re.match(r"^\|\s*\d+\s+Iluvatar", line) for line in out.splitlines()):
-                DEVICE_TYPE = "iluvatar"
+                DEVICE_TYPE = "iluvatar_gpu"
                 DEVICE_TYPE_DETECTED = True
                 return DEVICE_TYPE
         except Exception:
@@ -162,7 +175,7 @@ def get_device_count() -> int:
         DEVICE_COUNT = len(ids)
         return DEVICE_COUNT
 
-    if device_type == "iluvatar":
+    if device_type == "iluvatar_gpu":
         out = subprocess.check_output(["ixsmi"], text=True, stderr=subprocess.STDOUT)
         ids = set()
         for line in out.splitlines():
@@ -201,7 +214,7 @@ def _refresh_snapshot(device_type):
                         snapshot[dev_id] = (total_mib / 1024.0, used_mib / 1024.0)
                         break
 
-    elif device_type == "iluvatar":
+    elif device_type == "iluvatar_gpu":
         out = subprocess.check_output(["ixsmi"], text=True, stderr=subprocess.STDOUT)
         lines = out.splitlines()
         for i, line in enumerate(lines):
@@ -238,7 +251,7 @@ def get_memory_info(gpu_id):
         finally:
             pynvml.nvmlShutdown()
 
-    if device_type in ("xpu", "iluvatar"):
+    if device_type in ("xpu", "iluvatar_gpu"):
         _refresh_snapshot(device_type)
         if _MEM_SNAPSHOT is None or gpu_id not in _MEM_SNAPSHOT:
             raise RuntimeError(f"Failed to get memory info for {device_type} device {gpu_id}")
@@ -384,7 +397,8 @@ def pid_exists(pid):
                             APITestPaddleOnly,
                             APITestPaddleTorchGPUPerformance,
                             APITestTorchGPUPerformance,
-                            APITestCustomDeviceVSCPU)
+                            APITestCustomDeviceVSCPU,
+                            APITestPaddleDeviceVSGPU)
 
         test_classes = {
             "APIConfig": APIConfig,
@@ -395,7 +409,8 @@ def pid_exists(pid):
             "APITestTorchGPUPerformance": APITestTorchGPUPerformance,
             "APITestPaddleTorchGPUPerformance": APITestPaddleTorchGPUPerformance,
             "APITestAccuracyStable": APITestAccuracyStable,
-            "APITestCustomDeviceVSCPU": APITestCustomDeviceVSCPU
+            "APITestCustomDeviceVSCPU": APITestCustomDeviceVSCPU,
+            "APITestPaddleDeviceVSGPU": APITestPaddleDeviceVSGPU
         }
         globals().update(test_classes)
 
@@ -466,6 +481,7 @@ def run_test_case(api_config_str, options):
         "paddle_torch_gpu_performance": APITestPaddleTorchGPUPerformance,
         "accuracy_stable": APITestAccuracyStable,
         "paddle_custom_device": APITestCustomDeviceVSCPU,
+        "custom_device_vs_gpu": APITestPaddleDeviceVSGPU,
     }
     test_class = next(
         (cls for opt, cls in option_to_class.items() if getattr(options, opt, False)),
@@ -646,6 +662,42 @@ def main():
         default=0,
         help="The numpy random seed ",
     )
+    parser.add_argument(
+        "--custom_device_vs_gpu",
+        type=parse_bool,
+        default=False,
+        help="test paddle api on custom device vs GPU",
+    )
+    parser.add_argument(
+        "--operation_mode",
+        type=str,
+        choices=["upload", "download"],
+        help="Operation mode: upload or download",
+    )
+    parser.add_argument(
+        "--bos_path",
+        type=str,
+        default="",
+        help="BOS storage path (required when operation_mode is specified)",
+    )
+    parser.add_argument(
+        "--bos_conf_path",
+        type=str,
+        default="./conf",
+        help="Path for bcecmd --conf-path when using BOS",
+    )
+    parser.add_argument(
+        "--bcecmd_path",
+        type=str,
+        default="./bcecmd",
+        help="bcecmd binary path used for BOS upload/download",
+    )
+    parser.add_argument(
+        "--target_device_type",
+        type=str,
+        choices=["gpu", "xpu", "iluvatar_gpu"],
+        help="Target device type for download mode",
+    )
     parser.add_argument(
         "--bitwise_alignment",
         type=bool,
@@ -667,6 +719,7 @@ def main():
         options.paddle_torch_gpu_performance,
         options.accuracy_stable,
         options.paddle_custom_device,
+        options.custom_device_vs_gpu,
     ]
     if len([m for m in mode if m is True]) != 1:
         print(
@@ -679,10 +732,18 @@ def main():
             "--paddle_torch_gpu_performance"
             "--accuracy_stable"
             "--paddle_custom_device"
+            "--custom_device_vs_gpu"
             " to True.",
             flush=True,
         )
         return
+    if options.custom_device_vs_gpu:
+        if options.operation_mode and not options.bos_path:
+            print("--bos_path is required when --operation_mode is specified", flush=True)
+            return
+        if options.operation_mode == "download" and not options.target_device_type:
+            print("--target_device_type is required in download mode", flush=True)
+            return
     if options.test_tol and not options.accuracy:
         print(f"--test_tol takes effect when --accuracy is True.", flush=True)
     if options.test_backward and not options.paddle_cinn:

tester/__init__.py

@@ -12,6 +12,7 @@
     'APITestPaddleTorchGPUPerformance',
     'APITestAccuracyStable',
     'APITestCustomDeviceVSCPU',
+    'APITestPaddleDeviceVSGPU',
     'paddle_to_torch',
     'TensorConfig', 
     'APIConfig', 
@@ -32,6 +33,7 @@
     from .paddle_cinn_vs_dygraph import APITestCINNVSDygraph
     from .accuracy_stable import APITestAccuracyStable
     from .paddle_device_vs_cpu import APITestCustomDeviceVSCPU
+    from .paddle_device_vs_gpu import APITestPaddleDeviceVSGPU
     from . import paddle_to_torch
     from .api_config import (
         TensorConfig,
@@ -74,6 +76,9 @@ def __getattr__(name: str) -> Any:
     elif name == 'APITestCustomDeviceVSCPU':
         from .paddle_device_vs_cpu import APITestCustomDeviceVSCPU
         return APITestCustomDeviceVSCPU
+    elif name == 'APITestPaddleDeviceVSGPU':
+        from .paddle_device_vs_gpu import APITestPaddleDeviceVSGPU
+        return APITestPaddleDeviceVSGPU
     elif name == 'paddle_to_torch':
         from . import paddle_to_torch
         return paddle_to_torch