atlasleong/ug-cppo-finai-2025-npu
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

UG-CPPO seed 47 — Ascend NPU Inference

Model introduction

Real Ascend NPU inference adaptation for the pinned local model graceesthi/ug-cppo-finai-2025.

  • Architecture route: Stable-Baselines3 PPO (UG-CPPO seed 47)
  • Model family: Uncertainty-Gated CVaR-PPO trading agent (UG-CPPO paper v3, FinAI Contest 2025, NeurIPS 2026 submission). Author — Grace-Esther Dong, Aivancity Paris-Cachan.
  • Pinned source revision: 8be042ec04ee8811d3bc4e6a69372e61f34a72d5 (git commit 8be042e "Upload paper (PAT-corrected v3)").
  • Selected real checkpoint: ug_cppo_seed47.zip — 270607 bytes, SHA256 13be9637a0424f6dbea8b258691f1e771f938c923fd1ac806f83c5a26d90d823.
  • License: mit (as declared in the upstream model card — not guessed).
  • library_name: stable-baselines3 · pipeline_tag: reinforcement-learning · task: reinforcement-learning-policy.

The policy is a standard Stable-Baselines3 ActorCriticPolicy (obs dim 81, action dim 10, hidden [64, 64], diagonal-Gaussian log_std, no SDE). The deterministic test feeds a zero observation through the real loaded policy on physical npu:7 and emits action, value, log_prob.

NPU-specific tags

ascend, npu, torch_npu, hardware: NPU. Engine: torch_npu on physical npu:7 (Ascend910B4-1).

Environment dependencies

Captured from the final logs/npu_result.json:

Package / toolVersion
OSLinux 5.10.0-325.0.0.226.oe2203sp4.aarch64
Python3.11.14 (CPython)
torch2.9.0+cpu (image-provided, never replaced)
torch_npu2.9.0 (image-provided, never replaced)
numpy2.3.4
cloudpickle3.1.2
gymnasium1.2.0
stable-baselines32.8.0
CANN8.5.1
npu-smi (driver)25.5.1
NPU deviceAscend910B4-1

torch and torch_npu are the container's CANN-matched pair. They MUST NOT be replaced by ordinary PyPI wheels. All other non-torch dependencies are installed exclusively through the Aliyun PyPI mirror.

Exact Aliyun-only install steps

cd /root/ascend-model-pipeline/jobs/rl26/ug-cppo-finai-2025/ug-cppo-finai-2025-npu

# 1. Create a runtime outside the publishable delivery tree. It reuses the
#    image's system torch/torch_npu pair.
export UG_CPPO_RUNTIME_VENV=/work/pipeline/runtime/venvs-rl26/ug-cppo-finai-2025
python3 -m venv --system-site-packages "$UG_CPPO_RUNTIME_VENV"

# 2. Install only the necessary non-torch dependencies via the Aliyun mirror.
export PIP_INDEX_URL=https://mirrors.aliyun.com/pypi/simple/
"$UG_CPPO_RUNTIME_VENV/bin/pip" install --no-cache-dir -i https://mirrors.aliyun.com/pypi/simple/ 'numpy==2.3.4'
"$UG_CPPO_RUNTIME_VENV/bin/pip" install --no-cache-dir --no-deps -i https://mirrors.aliyun.com/pypi/simple/ 'stable-baselines3==2.8.0'

# (requirements.txt is the equivalent record; see below.)
"$UG_CPPO_RUNTIME_VENV/bin/pip" install --no-cache-dir -i https://mirrors.aliyun.com/pypi/simple/ -r requirements.txt

requirements.txt:

numpy==2.3.4
stable-baselines3==2.8.0

numpy==2.3.4 reproduces the checkpoint's own numpy 2.x environment, and stable-baselines3==2.8.0 matches the checkpoint serialization (_stable_baselines3_version = 2.8.0, JSON data format).

Inference steps

# 1. Activate the CANN environment (best-effort; already configured on the image).
source /usr/local/Ascend/ascend-toolkit/set_env.sh

# 2. Confirm the physical device is visible.
npu-smi info

# 3. Point to the selected immutable source checkpoint outside this delivery.
export UG_CPPO_CKPT=/work/pipeline/models/rl26/ug-cppo-finai-2025/ug_cppo_seed47.zip
ls -l "$UG_CPPO_CKPT"
sha256sum "$UG_CPPO_CKPT"
# expected: 13be9637a0424f6dbea8b258691f1e771f938c923fd1ac806f83c5a26d90d823

# 4. Run the real deterministic NPU inference (this is the one test command).
bash run_npu.sh

# 5. Inspect the produced evidence.
cat logs/npu_inference.log
cat logs/npu_result.json
cat logs/npu_smi_snapshot.txt

run_npu.sh uses the external runtime and pinned source checkpoint, launches inference.py on physical npu:7, waits for the [NPU_READY] marker, captures an npu-smi snapshot for the live PID, then waits for completion. Set NPU_HOLD_SECONDS to change the hold window (default 8 s).

Complete test case and output

Command:

cd /root/ascend-model-pipeline/jobs/rl26/ug-cppo-finai-2025/ug-cppo-finai-2025-npu
bash run_npu.sh

Runner transcript:

[RUNNER] launching: /work/pipeline/runtime/venvs-rl26/ug-cppo-finai-2025/bin/python inference.py --ckpt /work/pipeline/models/rl26/ug-cppo-finai-2025/ug_cppo_seed47.zip
[RUNNER] python pid=3312749
[RUNNER] ready marker detected; capturing npu-smi snapshot
[RUNNER] exit_code=0

logs/npu_inference.log:

[INFO] device=npu:7 current=7 name=Ascend910B4-1
[INFO] checkpoint ug_cppo_seed47.zip verified bytes=270607 sha256=13be9637a0424f6d...
[INFO] policy on npu:7 ok, params=19541 buffers_checked=0
[INFO] obs_space=Box(-inf, inf, (81,), float32) act_space=Box(-1.0, 1.0, (10,), float32) obs_dim=81 act_dim=10
[NPU_READY] pid=3312749 device=npu:7 ready_at=1787387255.408577 marker=/work/pipeline/jobs/rl26/ug-cppo-finai-2025/ug-cppo-finai-2025-npu/logs/npu_ready.marker
[NPU_HOLD] holding 8.0s for external npu-smi correlation
[RESULT] action=[[-0.3171851336956024, -0.6704590320587158, -0.21038641035556793, 0.5674934983253479, 0.6816909909248352, 0.6051729321479797, 0.012787358835339546, -0.35191959142684937, -0.818838357925415, -0.7145680785179138]]
[RESULT] value=[[0.001923096482641995]]
[RESULT] log_prob=[-15.180788040161133]
[RESULT] mean_abs_action=0.495050 std=[1.7112782001495361, 2.2721967697143555, 1.6971241235733032, 1.9133570194244385, 1.305491328239441, 2.009915351867676, 2.6645281314849854, 1.5622222423553467, 1.821391224861145, 1.5923500061035156]
[RESULT] forward_elapsed_s=0.084724 pid=3312749 exit_code=0
[DONE] result_json=/work/pipeline/jobs/rl26/ug-cppo-finai-2025/ug-cppo-finai-2025-npu/logs/npu_result.json

logs/npu_result.json (top-level contract):

{
  "status": "ok",
  "engine": "torch_npu",
  "exit_code": 0,
  "npu_available": true,
  "npu_device_name": "Ascend910B4-1",
  "source_revision": "8be042ec04ee8811d3bc4e6a69372e61f34a72d5",
  "input_device": "npu:7",
  "model_device": "npu:7",
  "output_device": "npu:7",
  "pid": 3312749
}

Physical npu-smi process correlation (logs/npu_smi_snapshot.txt):

| 7       0                 | 3312749       | python                   | 116                     |
	Process id:3312749 Process name:python            Process memory(MB):116

The exact live Python PID 3312749 appears on physical NPU 7 (116 MB HBM).

Outputs

  • action — [1, 10] deterministic policy mean (clipped to the action Box).
  • value — [1, 1] value-function estimate.
  • log_prob — [1] log-probability of the deterministic action.
  • policy_std — [10] per-dimension diagonal-Gaussian std (exp(log_std)).
  • q — null (PPO has no Q-function; not applicable).

Screenshots

Evidence images rendered from the real final logs/result:

  • assets/agent_workflow.png — ModelAgent pipeline steps with the real commands used.
  • assets/npu_device_call.png — device call and placement evidence on npu:7 (log + npu-smi process line).
  • assets/model_result.png — meaningful model output (action/value/log_prob) from the real log and result JSON.
  • assets/modelagent_live_session.png — live runner session, ready marker, and physical NPU process correlation.

Limitations

  • Single deterministic inference: feeds one zero observation to the loaded policy. It validates placement and the forward path; it is not a training run or a portfolio backtest.
  • PPO, no Q-function: the result JSON records q: null.
  • CPU only orchestrates: the checkpoint is deserialized on CPU (allowed), but every parameter, numeric input and tensor output used by the forward lives on npu:7. There is no CPU forward and no CPU offload/fallback.
  • No Gym/Unity/robot hardware is launched; the policy is invoked directly.
  • Determinism: deterministic=True; output is reproducible from the pinned immutable checkpoint.
  • The checkpoint bytes are not part of the published file set (SHA256SUMS covers the published artifacts; the checkpoint SHA256 is recorded in logs/npu_result.json and verified at every run).