2toINF/X-VLA-Libero (Vision-Language-Action, Florence2-large + SoftPromptedTransformer, 0.9B, ~3.3GB FP32)129e71460678b7236cee6fc9707f09d9fa0c3590manipulation-vla (image + language + proprio -> action_chunk)model.safetensors 3.52GB (3519068172 bytes) FP32, 879M paramseager, supports npu:0 and cpuXVLAProcessor (image_processor + tokenizer) -> input_ids [1,50], image_input [1,3,3,224,224], image_mask [1,3][1,30,20] (ee6d: xyz 6D-rotation gripper x2 arms, 30-step horizon), postprocess sigmoid on gripper channels, finite, horizon 30, action_dim 20npu:0 execution, no CPU fallback; all float tensors on npu:0 when NPU requestednext(model.parameters()).device, next(model.vlm.parameters()).device, next(model.transformer.parameters()).device, image_input.device, output.device all npu:0timm stub for DaviT and Florence2PreTrainedModel._supports_sdpa for transformers 4.57 compatibilitytrust_remote_code=True for XVLA remote code, auditable; tie_weights patched to avoid deleted lm_headpip install -r requirements.txt
python inference.py # auto NPU if available
PYTHONPATH=/tmp/timm_stub:/tmp/xvla_libero:$PYTHONPATH python inference.pyDefault loads local /tmp/xvla_libero (or HF cache fallback) and prints model/revision/route/backend/device/dtype/input summary/action summary/sync latency/PASS/FAIL. CPU-NPU consistency uses atol 1e-3 rtol 1e-3 (FP32 NPU numerical).
Ascend910_9362, torch.npu.is_available()=True, device_count=2, npu:0 existsnpu:0 (vlm npu:0, transformer npu:0)npu:0 (image_input [1,3,3,224,224] float32, input_ids [1,50] int64, proprio [1,20])npu:0 action_chunk [1,30,20] float32 finite True shape_ok True[-0.035757, -0.128211, 0.210233, 0.113327, -0.183618, ...] (20D)[-0.072230, -0.152349, 0.219481, 0.080675, -0.229136, ...][-0.035975, -0.128638, 0.210355, 0.113417, -0.183389, ...]0.000759 tol 0.001572 (atol 1e-3 rtol 1e-3) -> PASS91.05 ms, min 90.24 max 92.46, p50 91.05 p90 92.46 (10 runs, warmup 3, torch.npu.synchronize)329.5 actions/s, 10.98 chunks/s (horizon 30, batch 1, steps 5, float32)Shared weights (single model instance moved CPU->NPU, same embed_tokens random init), same deterministic images (seed i*1234), same language "pick up the red block and place it on the green bowl", same processor, same proprio/domain, seed 42, steps 5, float32. Compared flattened [600] action values. Threshold atol=1e-3 rtol=1e-3 (FP32 NPU fused ops). Also verified finite, shape [1,30,20], action_dim 20, horizon 30.
| metric | value |
|---|---|
| cpu first_action[0:5] | [-0.035975, -0.128638, 0.210355, 0.113417, -0.183389] |
| npu first_action[0:5] | [-0.035757, -0.128211, 0.210233, 0.113327, -0.183618] |
| max_abs | 0.000759 |
| max_ref | 0.5722 |
| tolerance | 0.001572 |
| finite | true |
| shape | [1,30,20] |
| passed | true |
torch_npu built for CANN 8.5.1, Ascend910B/910_9362, single card sufficienttimm stub provided at /tmp/timm_stub for DaviT DropPath/trunc_normal_; einops requiredtrust_remote_code=True, patched tie_weights and _supports_sdpa for transformers 4.57 compatibilityassets/agent_workflow.png — reconnaissance, resource check, load, NPU inference, consistency, performanceassets/npu_device_call.png — npu-smi, torch_npu availability, device name, model/input/output device proofassets/model_result.png — default python inference.py showing model, route, action summary, latency, PASSinference.py
readme.md
requirements.txt
assets/agent_workflow.png
assets/npu_device_call.png
assets/model_result.png
