This delivery adapts the official gtca/alphagenome_pytorch checkpoint for a
real Ascend NPU compatibility smoke. The current official sources are:
The Hugging Face revision was checked live on 2026-08-23:
b01c0ffa73e07c053491f3b5ea8bcf67d93b9920. The model card identifies a 450M
parameter AlphaGenome PyTorch port and recommends model_all_folds.safetensors.
The official PyTorch source HEAD checked the same day is
473d3d226b3a0aec754fb47cb6fcab1028829435.
The recommended checkpoint is 920,635,044 bytes with SHA-256
8c1cde573ba9f462e8aa91fd9c6e13ea6bd004857c429932d6c64db386e13595.
The model card's other license applies: the weights remain subject to the
Google DeepMind AlphaGenome Model Terms, while the code is Apache 2.0.
Download the exact model_all_folds.safetensors file into a local directory,
then run:
python inference.py --model-path /tmp/model-agent/alpha/model_all_folds.safetensors --device npu --output validation/npu-runtime.jsonThe script supports --device cpu|npu|auto. Explicit npu fails when an
Ascend device is unavailable; it never silently falls back to CPU.
The fixed smoke uses the official required input length of 1,048,576 bp with a
one-hot [1,1048576,4] sequence, selects only the atac head at 128 bp to
keep the compatibility gate bounded, and returns [1,8192,256] finite output.
On the verified Ascend910 npu:0, the model has 450,452,613 parameters and
the measured forward time was approximately 2.312768 seconds.
This is an execution compatibility gate only. It does not claim genomic prediction quality or a full CPU/NPU precision evaluation.


