This repository adapts the official graceesthi/ug-cppo-finai-2025 Stable-Baselines3 UG-CPPO trading policy for a direct PyTorch smoke run on Ascend NPU. The official model card identifies the MIT license and the upstream implementation at graceesthi/ug_cppo.
8be042ec04ee8811d3bc4e6a69372e61f34a72d5ug_cppo_seed42.zip5b9283ef8a19343842141191fac90c8ab4422af8a642a78b7f29a188c3d69814The official policy.pth was loaded from the Stable-Baselines3 ZIP with a strict 13-key and shape check. The reconstructed deterministic actor is 81 -> 64 -> 64 -> 10 with Tanh activations; the value branch is 81 -> 64 -> 64 -> 1. The action output is clipped to the official continuous action range [-1, 1] for the smoke call. model-lock.json records the source hash and architecture.
The same deterministic [1,81] float32 input was executed on CPU and a real Ascend910 npu:0. This is an execution and finite-output smoke gate; it does not claim a full financial-environment return or trading-accuracy result.
[1,81] -> [1,10] action and [1,1] value, float3219,531; strict policy keys: 130.003444596193730831 s0.1290899938903749 s0.99999999999999868.940696716308594e-08 / 2.458691596984863e-081.0; value max absolute difference: 4.470348358154297e-08torch_npu reported npu_available=true; NPU output was finite on npu:0The run is intentionally smoke-only: it proves that the official policy checkpoint can execute on the real Ascend NPU and that the deterministic CPU/NPU outputs agree closely. It does not claim full financial-environment return or trading-accuracy parity.
Place the official artifact under model/ug_cppo_seed42.zip, source the CANN environment, and run:
./run.shCPU smoke can be selected with DEVICE=cpu ./run.sh. The script writes machine-readable JSON and NumPy artifacts under validation/.


