YuankaiLuo/SimVLA-LIBERO (SimVLA LIBERO Edition, 0.5B VLM 811M total params, Vision-Language-Action baseline)观测→动作块 契约config.json 453 B + model.safetensors 3.1 GB (FP32, 3245529028 bytes, 786 tensors), state.json 23 B (global_step 150000) — 无需外挂 SiMLA 权重,vlm+transformer 已融合[B,3,3,384,384] + proprio 8维 [B,8] (libero_joint: ee_pos 3 + ee_ori 3 + gripper 2) + 语言指令 token [B,50] (Tokenizer HuggingFaceTB/SmolVLM-500M-Instruct, vocab 49280)action_chunk [B,10,7] (horizon 10, action_dim 7: delta_xyz 3 + delta_euler 3 + gripper 1, libero_joint)manipulation-vla (图像/状态/语言 → 连续动作块) — 已通过 inspect_robotics_model.py 判定HF_ENDPOINT=https://hf-mirror.com (国内镜像下载)/usr/local/Ascend/cann-8.5.1, 测试设备 npu:0 (0000:0A:00.0, HBM 65536 MB, 当前 18709 MB used)/opt/atomgit/adapt-npu-agent/model_cache/SimVLA-LIBERO (本地加载, 失败回退 snapshot_download via hf-mirror) + SmolVLM base /opt/atomgit/adapt-npu-agent/model_cache/SmolVLM-500M-Instructpip install -r requirements.txtrequirements.txt:
torch==2.9.0
torch_npu==2.9.0.post1
transformers==4.57.6
safetensors
huggingface_hub
Pillow
numpy
accelerate默认命令 (CPU 基准 + NPU 推理 + 一致性 + 性能, 单步 horizon 10, 真实 npu:0):
python inference.py显式镜像/离线:
HF_ENDPOINT=https://hf-mirror.com python inference.py日志至少打印: model_name, revision, route, backend, npu_available, device_name, dtype, 首参数 device, 输入摘要 (image_input/proprio/language), 固定 noise, CPU/NPU action shape/有限值/首尾动作, 同步耗时, 一致性与基准结果。
来自 logs/model_result.log 的真实 NPU 执行 (npu:0, float32, 同步计时, 固定 noise):
torch.Size([1,10,7]) first [-0.65429443, 0.09992165, -0.13238291, -0.07856778, -0.09159766, 0.92313379, 1.05058360] last [-1.04619765, -0.01803930, -0.74430179, -0.15692564, 0.06776998, 0.84471583, -0.90212231] mean -0.072222npu:0 Ascend910_9362, dtype torch.float32, horizon 10 action_dim 7 batch 1, image_input (1,3,3,384,384)action_chunk [1,10,7] (flow matching 去噪后反归一化, LIBERO joint 7维)fixed_noise [1,10,7] seed 123, 相同 proprio/image_input/input_ids, 同权重, 同采样步数 10, 同 Euler 积分 dt=-0.1, 手动同步流程 flow_generate_actions (非随机 generate_actions) 以消除 CPU/NPU RNG 分流差异max_abs/mae 计算 (FP32 atol 1e-4 rtol 1e-3, BF16 atol 5e-3 rtol 5e-2)generate_actions (直接 torch.randn on device), CPU vs NPU 差值可达 1.97 (随机性主导), 故以确定性 fixed_noise 为最终一致性标准, 符合 manipulation-vla 路由阈值torch.npu.synchronize()assets/agent_workflow.png — 完整工作流日志 (环境检查→获取模型→分析→NPU验证→性能与校验)assets/npu_device_call.png — npu-smi、NPU availability、设备名、模型/输入/输出 device 证据assets/model_result.png — 默认推理输出与一致性/性能结果 (CPU/NPU 首动作、耗时、PASS)model_cache/SmolVLM-500M-Instruct# 环境检查
npu-smi info
python -c "import torch; import torch_npu; print(torch.npu.is_available()); print(torch.npu.get_device_name(0))"
# 分析模型
python .opencode/skills/npu_adapt_robotics/scripts/inspect_robotics_model.py ./model_cache/SimVLA-LIBERO
# 推理验证
HF_ENDPOINT=https://hf-mirror.com python inference.py
# 校验提交
python .opencode/skills/npu_adapt_robotics/scripts/validate_submission.py YuankaiLuo-SimVLA-LIBERO-NPU