feat(research): support CUDA convergence runs

Allow the sealed ResNet and YOLO convergence protocol to run unchanged on the remote NVIDIA accelerator while retaining CPU and MPS support.

Constraint: Preserve fixed seeds, exact workloads, and one-device-per-run semantics

Rejected: Reuse mixed MPS and CUDA results | invalid cross-device production matrix

Confidence: high

Scope-risk: narrow

Not-tested: Full production matrix on RTX 3080 Ti

test/post_training_model_convergence.py

@@ -128,8 +128,8 @@ class StudyConfig:
             raise ProtocolError("residual bounds and initialization radius are fixed")
         if tuple(self.objective_checkpoints) != OBJECTIVE_CHECKPOINTS:
             raise ProtocolError("objective checkpoints are fixed")
-        if self.device not in {"cpu", "mps"}:
+        if self.device not in {"cpu", "mps", "cuda"}:
-            raise ProtocolError("device must be cpu or mps")
+            raise ProtocolError("device must be cpu, mps, or cuda")
         if not self.workload_ids:
             raise ProtocolError("at least one workload must be registered")
 
@@ -1303,7 +1303,11 @@ def persist_state(run_root: str | os.PathLike[str], state_machine: StudyStateMac
 def build_cli_parser() -> argparse.ArgumentParser:
     parser = argparse.ArgumentParser(description="Post-training ResNet/YOLO convergence protocol")
     parser.add_argument("--phase", choices=["prepare", "smoke", "develop", "confirm", "publish", "all"], required=True)
-    parser.add_argument("--device", choices=["mps", "cpu"], default="cpu")
+    parser.add_argument(
+        "--device",
+        choices=["cuda", "mps", "cpu"],
+        default="cpu",
+    )
     parser.add_argument("--data-root", type=Path, default=Path("result/cache"))
     parser.add_argument("--run-root", type=Path, required=True)
     parser.add_argument("--allow-download", action="store_true")

test/post_training_resnet_convergence.py

@@ -790,8 +790,10 @@ class ResNetConvergenceAdapter:
         self.data_root = Path(data_root)
         self.device = torch.device(device)
         self.allow_download = bool(allow_download)
-        if str(self.device) not in {"cpu", "mps"}:
+        if str(self.device) not in {"cpu", "mps", "cuda"}:
-            raise ProtocolError("ResNet device must be cpu or mps")
+            raise ProtocolError(
+                "ResNet device must be cpu, mps, or cuda"
+            )
 
     def _result_path(self) -> Path:
         return self.run_root / "workloads" / self.workload_id / "result.json"

test/post_training_yolo_convergence.py

@@ -1047,9 +1047,10 @@ class StrictScratchTrainer:
         batch: int = 16,
         epochs: int = 100,
     ) -> None:
-        if device not in {"cpu", "mps"}:
+        if device not in {"cpu", "mps", "cuda"}:
             raise YoloProtocolError(
-                "device must remain cpu or mps for the complete run"
+                "device must remain cpu, mps, or cuda for the "
+                "complete run"
             )
         if epochs not in {2, 100}:
             raise YoloProtocolError("trainer epochs must be smoke 2 or production 100")
@@ -1519,8 +1520,10 @@ class YoloConvergenceAdapter:
     def __init__(self, *, workload_id: str, config: StudyConfig, run_root: str | os.PathLike[str], data_root: str | os.PathLike[str], device: str | torch.device, allow_download: bool) -> None:
         if workload_id != WORKLOAD_ID:
             raise YoloProtocolError(f"unsupported workload id: {workload_id}")
-        if str(device) not in {"cpu", "mps"}:
+        if str(device) not in {"cpu", "mps", "cuda"}:
-            raise YoloProtocolError("device must be cpu or mps")
+            raise YoloProtocolError(
+                "device must be cpu, mps, or cuda"
+            )
         self.workload_id, self.config = workload_id, config
         self.run_root, self.data_root, self.device = Path(run_root), Path(data_root), str(device)
         self.allow_download = bool(allow_download)