graceesthi/ug-cppo-finai-2025 是 UG-CPPO v3 论文(FinAI Contest 2025, NeurIPS 2026 提交)的 30 个训练态 Stable-Baselines3 智能体集合(PPO / CVaR-PPO / UG-CPPO × 10 seeds 42-51)。本仓库适配其中 UG-CPPO seed 47(Rachev 1.0104,top performer)为单卡昇腾 NPU(Ascend910_9362, npu:0)真实推理。
8be042ec04ee8811d3bc4e6a69372e61f34a72d5ActorCriticPolicy, MlpPolicy)Flatten -> Linear(81->64) Tanh -> Linear(64->64) Tanh -> Linear(64->10) (actor) + Linear(64->1) (critic) + log_std (10)Box(-1,1, (10,)),确定性动作为 clip(mean, -1, 1)(多头/空头)ug_cppo_seed47.zip (~270KB, 19541 params FP32),其余 29 权重同结构可替换所有参数与输入输出 Tensor 显式位于 npu:0,inference.py 从本地权重加载并执行真实前向与采样,不使用 dummy policy。
| Field | Value |
|---|---|
| Route | offline-rl (SB3 PPO trading policy) |
| Obs dim | 81 |
| Action dim | 10 continuous (-1..1) |
| Horizon | 1 (单步决策) |
| Actor | Linear 81→64 → Tanh → Linear 64→64 → Tanh → Linear 64→10 |
| Critic | Linear 64→1 |
| Sampler | deterministic (mean clipped) |
| Seed (推理对比) | 42 (固定合成观测批量) |
使用固定 seed 42 生成的合成 FinRL 兼容观测批量(5×81,首行为零基线,其余 N(0,0.5²)),在 CPU 与 NPU 上共享同一权重、观测、dtype (float32) 与确定性采样,确保唯一对比对象。投资者可替换为真实回测状态,shape 须为 [batch,81]。
npu:0), CANN 8.5.1Ascend910_9362, device_count=2pip install -r requirements.txtpython inference.py通过本地目录加载(避免重复下载):
UG_CPPO_LOCAL_DIR=/opt/atomgit/adapt-npu-agent/work/graceesthi_ug-cppo-finai-2025 python inference.py全链路在 npu:0 同步执行,5 样本批量输出:
(5,10), dtype float32, device npu:0, sample [-0.317, -0.670, -0.210, 0.567, 0.681](5,10), device npu:0, 全部在 [-1,1] 且有限(5,1), device npu:0npu:0, obs_device npu:0| Metric | Value |
|---|---|
| Max abs diff (mean) | 1.19e-06 |
| Max abs diff (value) | 1.47e-07 |
| Max abs diff (clipped action) | 5.36e-07 |
| Tolerance | atol=5e-3, rtol=5e-2 (≈0.23) |
| Finite & in bounds | True |
| Result | PASS |
| Metric | Value |
|---|---|
| Avg latency | 0.27 ms |
| Min / Max | 0.26 / 0.29 ms |
| p50 / p90 / p95 | 0.27 / 0.28 / 0.29 ms |
| Throughput | ~18344 actions/s, ~18344 decisions/s |


