#NPU
autogluon/mitra-classifierworking/models/mitra-classifier (config.json dim 512, dim_output 10, n_layers 12, n_heads 4, task CLASSIFICATION)tabular-classification (多分类)Tab2D —— 12 层 Transformer,隐藏维度 512,4 头注意力,约 75.7M 参数 (~303 MB fp32, 289 MB safetensors),由 AutoGluon 团队在纯合成混合先验上预训练,支持 in-context 学习[n_query, dim_output] 经切片到真实类别数后 softmax 得概率,argmax 得标签npu:0https://huggingface.co/autogluon/mitra-classifierrows × features,本验证使用 sklearn.datasets.load_wine,形状 [178, 13] -> 划分 train [142,13] / test [36,13] (80/20, random_state=42, stratify)['alcohol','malic_acid','ash','alcalinity_of_ash','magnesium','total_phenols','flavanoids','nonflavanoid_phenols','proanthocyanins','color_intensity','hue','od280/od315_of_diluted_wines','proline'],列顺序固定target 三分类 0,1,2Preprocessor 执行 nanmean 均值填充,单值列剔除,无缩放;特征选择关闭load_wine,SHA256 前缀 8edcf3903afd97c6npu-smi 25.5.5, 设备 Ascend910_9362, 2 设备/usr/local/Ascend/cann-8.5.1torch.npu.is_available()==True, device_count==2, 目标 npu:0float32pip install -r requirements.txtpython inference.py[36,3], probs [36,3]mean 6.2054, std 3.3782, min 0.9608, max 11.4540accuracy 1.0000, macro-F1 1.0000 (36 测试样本)python scripts/compare_outputs.py --cpu /tmp/cpu_logits.npy --npu npu_logits.npy --task classification --atol 0.01 --rtol 0.01max_abs 0.00581, mean_abs 0.00162, argmax_agreement 1.0, passed Truetorch.npu.synchronize()avg 25.8 ms, min 25.4 ms, max 26.4 ms, p50 25.8 ms, p90 26.3 ms, p95 26.4 ms1392.9 rows/s三张 PNG 均由 scripts/render_xterm_evidence.mjs --style raw 根据本次真实日志生成:
assets/agent_workflow.pngassets/npu_device_call.pngassets/model_result.png

dim_output=10 为最大类别数,实际需 ≤10support 上限 8192,query 上限 1024