Published package validation: 0.1.0¶
This page preserves the checks of the 0.1.0 wheel. They do not verify the 0.2.0 ONNX default or its installation dependencies.
The public gemmadecision 0.1.0 wheel passed functional checks in fresh
Modal CPU and GPU environments on September 27, 2026. Both environments
installed gemmadecision==0.1.0 from PyPI. No repository package code or local
model files were mounted; only the test harness was supplied.
| Environment | Local decide() / rank() |
Native PydanticAI locally | Granian HTTP + SDK + native PydanticAI |
|---|---|---|---|
| 4 CPU cores, 8 GiB RAM | Passed | Passed | Passed |
| NVIDIA H100 80 GB, 2 CPU cores, 16 GiB RAM | Passed | Passed | Passed |
The CPU image first installed PyTorch 2.14.0+cpu from the official PyTorch
CPU index. The GPU image used the package's normal dependency resolution:
PyTorch 2.14.0 with CUDA 13.0. Both used Python 3.11.12, Transformers 5.17.0,
PydanticAI 2.51.0, Granian 2.8.3 and FastAPI 0.136.3.
The checks verified installation under site-packages, public wheel origin,
and the immutable Hugging Face model revision
785d530221c990671f29976902540101bb9c7647. Each environment downloaded the model
without a Hugging Face token and checked the encoder, decision head and
configuration hashes. Local calls reused one loaded engine. That process
exited before the real Granian server started, preventing two model copies
from occupying memory together. The HTTP check made one inference request
through AsyncDecisionClient and native PydanticAI, returning a typed
Literal/bool result.
Both installations used this published wheel:
These are installation and integration checks using short, two-option examples. They do not measure accuracy, steady-state latency, throughput or CPU-versus-GPU performance. The single-call timings retained in the reports include different initialization effects and should not be used for those comparisons.
Evidence: CPU report,
GPU report, and
provenance and file hashes.
Only the temporary Hugging Face cache directory was replaced with $HF_HOME
in the published reports; the measured values are unchanged.