凤
weixin_44093752/smolvlm-256m-ascend-npu
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

SmolVLM-256M-Instruct Ascend NPU smoke adaptation

This delivery adapts the official HuggingFaceTB/SmolVLM-256M-Instruct checkpoint for a real Ascend NPU compatibility smoke. The official model card and source project are:

  • https://huggingface.co/HuggingFaceTB/SmolVLM-256M-Instruct
  • https://github.com/huggingface/smollm

The verified Hugging Face revision is 7e3e67edbbed1bf9888184d9df282b700a323964. The official Idefics3ForConditionalGeneration model and Idefics3Processor were loaded on npu:0. A text-only compatibility input with shape [1, 6] produced finite bfloat16 logits with shape [1, 6, 49280]. The real NPU forward completed in about 0.295 seconds.

This is an execution compatibility gate only. It does not claim image-task quality, captioning accuracy, or full precision. The official weight file is excluded from Git and must be downloaded into the model directory before running:

python inference.py --model-dir /tmp/model-agent/smolvlm --output validation/npu-runtime.json

The expected file is model.safetensors. Its verified size and SHA-256 are recorded in model-lock.json. The smoke uses the official Transformers model class and a local-only model load; no weight conversion is applied.

Evidence

Agent workflow

NPU device call

Model result