Instructions to use ThakiCloud/PowerProof-P2-ZeroShot-Baseline with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Scikit-learn
How to use ThakiCloud/PowerProof-P2-ZeroShot-Baseline with Scikit-learn:
from huggingface_hub import hf_hub_download import joblib model = joblib.load( hf_hub_download("ThakiCloud/PowerProof-P2-ZeroShot-Baseline", "sklearn_model.joblib") ) # only load pickle files from sources you trust # read more about it here https://skops.readthedocs.io/en/stable/persistence.html - Notebooks
- Google Colab
- Kaggle
PowerProof P2 zero-shot baselines (Ridge / GBDT): failed transfer to B200
These models failed zero-shot transfer to an unseen GPU. Trained only on H100 and H200 measurements and evaluated once on B200, the primary Ridge model had a mean-power WAPE of 30.02% and the secondary GBDT model 43.33%; for p95 power the WAPE was 23.71% (Ridge) and 43.04% (GBDT). All 240 B200 runs were under-predicted, for both models and for both mean and p95 power. Both models are released only as experimental reproduction baselines for that negative result. They are not general-purpose GPU power predictors and should not be used to budget power for any GPU.
π°π· νκ΅μ΄ μΉ΄λ: MODEL_CARD.ko.md
Intended use
- Reproducing the reported B200 zero-shot failure and serving as a baseline that new methods (for example per-device calibration) must beat.
- Not intended for: power budgeting, rack or facility planning, capacity planning, or any production decision on any GPU, including the two GPUs the models were trained on.
Models
| File | Role | Estimator | Outputs |
|---|---|---|---|
model.joblib |
Primary (fixed before evaluation) | StandardScaler (numeric) + OneHotEncoder(handle_unknown="ignore") (categorical) + Ridge(alpha=1.0) |
mean power (W), p95 power (W) |
model_secondary.joblib |
Secondary (pre-registered diagnostic, never substituted) | numeric passthrough + OneHotEncoder + HistGradientBoostingRegressor (max_iter 200, learning_rate 0.05, min_samples_leaf 8, l2_regularization 1.0, random_state 2026) |
mean power (W), p95 power (W) |
Each file is a Python dict {"name", "models": {"mean_power": Pipeline, "p95_power": Pipeline}, "rule", "n_rows"}.
There was no hyperparameter search.
Inputs (see feature_schema.json)
Numeric: requested_input_tokens, requested_output_tokens, max_concurrency, target_rate_rps, burstiness,
power_limit_w. Categorical: workload_family (balanced, decode_heavy, prefill_heavy), loop (closed, open).
Preprocessing (implemented in predict.py, identical to the training code):
- The only device feature is the power limit:
source_gpu_typeis mapped topower_limit_w= H100 400 W, H200 600 W, B200 1000 W. Other GPU types are rejected. - For closed-loop rows,
target_rate_rpsandburstinessare set to 0; theloopone-hot distinguishes them. - Anything only known after running the workload (measured power, energy, throughput, temperature, clocks, latency, power-cap fractions) is excluded from the inputs.
Training data (not released)
504 valid runs: H100 186 + H200 318. Serving stack: vLLM 0.28.0 serving Qwen3-8B in bf16 on a single GPU, three synthetic workload shapes, closed-loop and open-loop load. B200 data was never used for training or model selection. The measurement data itself is not released.
What is measured
Labels are NVML GPU board power over the request window (first request start to last request end): mean power (energy-counter difference divided by window length) and p95 power (95th percentile of 0.2 s samples; NVML power is itself roughly a 1 s moving average, so this is not an instantaneous peak). This is GPU board power only, not server, node, or rack power.
Evaluation procedure
- Model selection contract frozen (2026-11-09).
- Models trained on H100 + H200 and frozen (2026-11-10).
- B200 labels kept encrypted; sealed-data integrity audit passed 240/240.
- One evaluation run on B200 (2026-11-11). No primary swap, refit on B200, ensembling, or post-hoc calibration afterwards.
Results on B200 (240 runs, WAPE, lower is better)
| Model | Mean power WAPE | p95 power WAPE | Signed bias (mean / p95) |
|---|---|---|---|
| Ridge (primary) | 30.02% | 23.71% | β30.02% / β23.71% |
| GBDT (secondary) | 43.33% | 43.04% | β43.33% / β43.04% |
| Ridge, by workload | Mean power WAPE | p95 power WAPE |
|---|---|---|
| balanced | 29.96% | 23.96% |
| decode_heavy | 32.27% | 24.53% |
| prefill_heavy | 27.87% | 22.71% |
- Runs with measured mean power at or above 80% of the power limit (107 runs): Ridge mean-power WAPE 28.89%.
- A 6-block bootstrap gives 29.82% to 30.28% for Ridge mean-power WAPE. With only 6 blocks the coverage of this interval is not validated.
- Full aggregates:
EVAL_B200.json. Per-run values are not released.
Limitations
- Three GPU types and one model/serving stack. Nothing here says anything about GPU power prediction in general.
- The cause of the failure is not identified. Candidate explanations (device power characteristics, what the power limit means on each device, VM vs bare-metal hosts, workload or serving-environment differences) were not separated.
- The 10% / 15% WAPE values that appear in the project records were a rule for choosing the number of training replicates, not a B200 acceptance threshold.
- The 240 B200 runs are designated as development data for follow-up work and should not be treated as an independent test set again.
License and provenance
- License: Apache-2.0 for everything in this repository, including the two model weight files
(
model.joblib,model_secondary.joblib). SeeLICENSEandNOTICE. - The weights were trained by ThakiCloud only on ThakiCloud's own GPU power measurements. No third-party dataset was used, and no measurement data is included here.
- Qwen3-8B (Apache-2.0) and vLLM (Apache-2.0) were the workload being served while power was measured. No Qwen weights, no model outputs, and no vLLM code are included.
- The pickles reference scikit-learn / NumPy classes (BSD-3-Clause); those libraries are not redistributed here.
- scikit-learn version: loading and reproduction were verified with scikit-learn 1.9.0 only. Other versions may fail to unpickle or may silently give different predictions; they were not tested.
Security: pickle files
model.joblib and model_secondary.joblib are joblib/pickle files. Unpickling can execute arbitrary code. Load
them only if you trust where you got them, and check SHA256SUMS first. predict.py refuses to load a file whose
SHA-256 differs from the pinned value.
Usage
pip install joblib "scikit-learn==1.9.0" pandas # the only scikit-learn version these files were verified with
sha256sum -c SHA256SUMS
python predict.py example_input.json # both models
python predict.py example_input.json --model primary
example_input.json contains synthetic, hand-written rows (not measured runs). Output is a JSON list with
mean_power_w and p95_power_w per row and model. The files were loaded and reproduced in the environment recorded
in RELEASE_CHECK.md; that is not a statement about the training environment.
Citation / attribution
ThakiCloud, PowerProof P2 (2026). If you use these baselines, please state that they failed zero-shot transfer to B200.
- Downloads last month
- -