Compact vision-language-action model for LIBERO manipulation, adapted to Huawei Ascend 910B NPU.
Base model: lerobot/smolvla_libero https://huggingface.co/lerobot/smolvla_libero (revision 31d453f7edd78c839a8bbc39744a292686daf0de)
Route: manipulation-vla | Task: language-conditioned manipulation (LIBERO)
VLM backbone: HuggingFaceTB/SmolVLM2-500M-Video-Instruct (tokenizer 48, num_vlm_layers 16, cross_attn)
Weights: 865 MB model.safetensors + 500M VLM, served via HuggingFace Hub
Input (batch=1):
observation.state 6-dim (joint + gripper, normalized MEAN_STD, padded to max_state_dim 32)observation.images.camera1/2/3 each 3×256×256 (resized with padding to 512×512, IDENTITY norm)task language instruction tokenized to 48 tokens (observation.language.tokens + observation.language.attention_mask bool)Output:
action_chunk shape (1, 50, 7) float32, horizon 50, action_dim 7 (delta joint + gripper), normalized with MEAN_STDnpu:0 Ascend910_9362, torch_npu 2.9.0.post1, torch 2.9.0npu:0 via .to("npu:0"); first param device npu:0, input images/state/tokens on npu:0, sampler noise on npu:0, output on npu:0Device proof:
npu:0 (state, 3 cameras, tokens, attention_mask)npu:0, sampler noise on npu:0, output on npu:0, torch.npu.synchronize() timedCPU vs NPU consistency (same weights, same observation, same fixed noise 1×50×32, seed 0):
Performance (10 timed runs after 3 warmup, with synchronize):
pip install -r requirements.txt
python inference.pyDefault prints model/revision, route, backend, device, dtype, input summary, CPU/NPU action summaries, sync latency, PASS/FAIL.
