#NPU
mldi-lab/Kairos_23m (KairosModel, 23M)main (model.safetensors 87.76 MB, 23,000,965 参数 float32)past_target [B, context] 输出 prediction_outputs [B, 9, prediction_length] 9分位数 (0.1~0.9), median 取 0.5tsfm==0.1.0 (foundation-model-research/Kairos) + transformers 4.57.6, 自定义 KairosModel 非 transformers 内置, 需 trust_remote_code=True + AutoModel.register(kairos)torch.abs(torch.fft.rfft(complex)) 在 Ascend 上报错 AclNN Abs complex, 已在推理侧打补丁: 若 device.type=="npu" 则 rfft+abs 走 CPU 再回 NPU, 其余路径 (polar/view_as_complex/DRoPE) 原生支持| 项 | 值 |
|---|---|
| 轴顺序 | [B, T] (严格, past_target 二维 [batch, time], 内部加 s_features via FFT) |
| 形状 | batch=2, context=2048 → 输出 [2, 9, 64] (9 quantiles × 64 horizon) |
| 频率 | 无显式 freq token; 依赖 DRoPE (instance spectral features 128-d via FFT+LayerNorm) |
| 缩放 | 模型内部 log scale + per-sample 归一 (context 内计算) |
| 缺失/掩码 | mask = isnan(context).logical_not(), 验证用 dense 无 NaN |
| 数据 | 确定性合成 (seed 42): randn[2,2048] + sin(2πt/96)*0.5, mean 0.0044 std 1.0546, preview batch0 last5 [1.01,-0.27,0.44,-0.37,0.096], SHA256 78e1beea |
| 输出 | prediction_outputs float32 [B, Q=9, P=64] , 9分位数单调递增, median 为 q4=0.5, stats mean -0.196 std 0.808 |
| dtype | float32 (config "dtype": "float32") |
| 预处理 | 无外部 scaling; 模型内 FFT→s_feature→ParamNet→DRoPE |
Ascend910_9362 ×2, 目标 npu:0torch.npu.is_available()==True, device_count==2tsfm.model.kairos.AutoModel.register → AutoModel.from_pretrained(local_dir, trust_remote_code=True) → model.to("npu:0"), 输入 Tensor 显式 to(npu:0), FFT 补丁仅在 NPU 路径触发pip install -r requirements.txt
# requirements 已包含 tsfm 依赖 (einops/jaxtyping/matplotlib/pandas); tsfm 源码由 inference.py 自动 bootstrap (git clone Kairos 若本地无 tsfm)python inference.py加载 working/kairos/model 或自动从 HF 镜像下载至 ./.cache/kairos-23m → npu:0 → warmup 3 → 同步计时 → 打印真实分位数预测。
关键日志(npu:0):
[INFO] model_name: mldi-lab/Kairos_23m
[INFO] task: probabilistic forecasting (quantiles) | backend: torch_npu | device: npu:0
[INFO] first param device before NPU move: cpu
[INFO] first param device after to(npu:0): npu:0
[INFO] input NPU device past:npu:0
[INFO] output summary: shape=(2, 9, 64) mean=-0.196054 std=0.808401
[INFO] median quantile idx 4 value 0.5 preview batch0 median first5: [0.2244,0.2475,0.2381,0.2283,0.1864]
[SUCCESS] NPU inference completed on npu:0 with Kairos_23m probabilistic forecastinginput [2,2048] mean 0.0044 std 1.0546 SHA256 78e1beea, preview batch0 last5 [1.01,-0.27,0.44,-0.37,0.096]output [2,9,64] device npu:0 dtype float32, mean -0.196054 std 0.808401 min -1.758 max 1.697median (q=0.5) batch0 first5 [0.2244,0.2475,0.2381,0.2283,0.1864] | q=0.1 first5 [-0.988,-1.036,-1.092,-1.122,-1.153] | q=0.9 first5 [1.348,1.418,1.429,1.444,1.398]quantile monotonic True min_diff 0.2189 (9分位数单调递增)finite True shape_ok True fallback_detected Falseweight 87.76 MB, 23M params, d_model 384, 4+4 layers, num_heads 8python .opencode/skills/npu-adapt-skills/scripts/compare_outputs.py --cpu working/kairos/cpu_output.npy --npu working/npu_output.npy --task regression --atol 1e-4 --rtol 1e-3| 输出 | max_abs | mean_abs | passed |
|---|---|---|---|
forecast [2,9,64] | 8.67e-05 | 2.70e-05 | ✅ (atol 1e-4) |
阈值说明: FP32 默认 1e-4 对 23M Transformer+MoE+DRoPE 足够严格 (实测 max 8.6e-05); mean 2.7e-05 远低于阈值; 同权重同预处理同随机生成输入,
torch.npu.synchronize()后落盘npy直接比较; NPU 仅 FFT 的 abs 走 CPU (数值等价, 误差 source 仅 float32 累加顺序)。
torch.npu.synchronize() 前后计时, 纯模型 forward)batch2 ctx2048 → out [2,9,64] (含 Mixture-of-Size 编码 + DRoPE + 4层编码+4层解码 + quantiles head)sync elapsed 66.05 ms (含编译)model(past_target, prediction_length, generation=True) 同步计时

由
scripts/render_xterm_evidence.mjs --style raw根据本次真实日志生成,提示符atomgit@pod-a94f8701860f4700b161b00e290de466:~$仅为展示标签,不证明服务器身份。
working/kairos/model。[B, context] 二维, context 默认 2048 (最大 2048, 过长截断最后 2048), prediction_length 默认 64 (支持 64/96, 需与 config 一致)。Cannot create tensor with interal format warning 无害已保留; past_key_values Tuple 弃用 warning 来自 transformers 已忽略。torch.fft.rfft → torch.abs(complex) 通过 CPU fallback 等价实现, 已记录并验证误差 <1e-4, 无静默 CPU fallback (仅该 FFT 算子, 其余均在 NPU)。submission/
├── inference.py
├── readme.md
├── requirements.txt
└── assets/
├── agent_workflow.png
├── npu_device_call.png
└── model_result.png