This delivery tracks the live qox/vn-address-normalizer checkpoint. The
model card describes a 26M-parameter character-level Transformer for
Vietnamese address normalization:
The Hugging Face revision was checked live on 2026-08-23:
5d0de6d07356774e861f72684e4807bdbc137351. The public snapshot provides
model.safetensors at the repository root (26,368,100 bytes) with SHA-256
4fcaae4f7a2e725aad234533e579884daa227c761b82c1c6a82c7ff73bbc2e3a.
The upstream inference.py currently expects additional vocabulary files and
a nested weight path that are not present in the live snapshot. This delivery
therefore includes a self-contained inference.py with the same official
S2S architecture and strict checkpoint loading, so the execution gate is
reproducible without claiming address-normalization accuracy.
Download the files listed in model-lock.json into /tmp/model-agent/vnaddr,
then run:
python inference.py --model-path /tmp/model-agent/vnaddr --device npu --output validation/runtime.jsonThe script uses bounded deterministic token tensors, executes the official
encoder/decoder architecture on npu:0, and fails when an explicitly
requested Ascend device is unavailable. On the verified Ascend910, strict
loading succeeded for 6,589,197 parameters, output [1,269] was finite, and
the forward took about 0.124086 seconds. CANN reported that
aten::_transformer_encoder_layer_fwd fell back to CPU; this is recorded as a
runtime compatibility boundary, not as a full NPU-kernel precision claim.


