#NPU
[batch, seq_len, 6] (open, high, low, close, volume, amount) 连续 K 线 + x_timestamp/y_timestamp (minute, hour, weekday, day, month) 时间特征,seq_len ≤2048(mini 上下文 2048)pred_len 步的 [pred_len, 6] 预测 DataFrame (inverse normalized 回到价格空间)Kronos 是首个开源金融 K 线 Foundation Model,预训练于 45 家交易所 120 亿条 K 线。Kronos-mini 为 4 层轻量版,配合 Kronos-Tokenizer-2k(2048 上下文)实现零样本预测。
[batch, time, channel],channel 固定 6,顺序 open, high, low, close, volume, amount,不得调换x_timestamp/y_timestamp 为 pandas Series/DatetimeIndex,需转换为 (minute, hour, weekday, day, month) 5维特征,minute 0-59, hour 0-23, weekday 0-6, day 1-31, month 1-12(x - mean) / (std+1e-5) 再 clip(-5, 5),与官方 KronosPredictor 一致;volume/amount 缺失填 0,amount 缺失时 amount = volume * mean(price)lookback=512, pred_len=30,lookback 512 + pred 30 = 542 总长度,df_all 连续生成随机游走 close = 10 + cumsum(N(0,0.1)),volume 5000±1000pred = pred_norm * (std+1e-5) + meanhalf=True 时输出 ([s1(1024 vocab), s2(1024 vocab)]) 层次化 tokentorch.npu.is_available()==True, device_count==2, torch.npu.get_device_name(0)==Ascend910_9362, torch.npu.set_device(0) 成功,first_param_device: npu:0pip install -r requirements.txt
# requirements.txt 已固定版本,见文件默认命令(必须在 NPU 机器上执行,自动同步计时):
python inference.py
# 自定义: python inference.py --lookback 512 --pred_len 30 --device npu:0
# 或 python inference.py --lookback 1024 --pred_len 120执行流程:
/opt/atomgit/adapt-npu-agent/work_kronos_mini、/tmp/kronos_mini_test、tokenizer_snapshot 优先加载权重;不存在则从 HuggingFace 公开下载 config.json + model.safetensors (Kronos-mini + Kronos-Tokenizer-2k)KronosTokenizer 与 Kronos,to("npu:0"),打印 first_param_device: npu:0x_timestamp/y_timestamp 时间特征,执行 per-series 归一化与 clipKronosPredictor.predict(df, x_timestamp, y_timestamp, pred_len=30, T=1.0, top_p=0.9, sample_count=1),内部通过 tokenizer.encode(half=True) → auto_regressive_inference 自回归 30 步 → tokenizer.decode(half=True) → inverse transformtorch.npu.synchronize() 计时,打印输入/输出摘要、s1_logits shape、吞吐量并以 exit 0 结束;NPU 不可用则非零退出单次真实运行(npu:0, lookback 512, pred_len 30, batch 1, float32, max_context 2048, T=1.0 top_p=0.9 seed 42):
model_name: NeoQuasar/Kronos-mini
tokenizer_name: NeoQuasar/Kronos-Tokenizer-2k
device: npu:0
first_param_device tokenizer: npu:0
first_param_device model: npu:0
input_shape: (512, 6) freq: 5min
forward_first_ms: 604.73
pred_shape: (30, 6)
pred_head:
open high low close volume amount
2024-01-03 04:10:00 10.172665 10.436798 10.005528 10.433221 4525.061523 22570.068359
2024-01-03 04:15:00 10.328270 10.531398 10.190851 10.583340 4264.947266 21332.181641
2024-01-03 04:20:00 10.440417 10.616011 10.298036 10.574318 4386.290039 21902.126953
2024-01-03 04:25:00 10.649998 10.673425 10.448514 10.493084 6730.792480 34857.476562
2024-01-03 04:30:00 10.348871 10.413543 10.309737 10.394262 4522.463867 22602.873047
pred_summary open mean=10.7322 close mean=10.7942 volume mean=4442.71
output_finite: True shape_valid=True
tokenizer_roundtrip: encode s1 torch.Size([1, 32]) s2 torch.Size([1, 32]) recon torch.Size([1, 32, 6]) unique_s1=29
s1_logits_shape: (1, 32, 1024) ctx_shape: (1, 32, 256) logits_finite=True
perf warmup=3 runs=10 avg_ms=250.78 min=242.39 max=263.40 p50=248.99 throughput_windows_per_sec=3.99
[SUCCESS] NPU inference completedCPU 与 NPU 使用同一权重、同一输入、同一预处理、同一 dtype (float32)、同一 seed、eval 模式,对比确定性部分(采样前 logits 与重建):
tok s1 equal True, s2 equal True (1,32) 测试,seed 0 随机输入 x=torch.randn(1,32,6)python .opencode/skills/npu-adapt-skills/scripts/compare_outputs.py --cpu /tmp/cpu_recon_kronos.npy --npu /tmp/npu_recon_kronos.npy --task regression --atol 1e-4 --rtol 1e-3结果: max_abs_error=2.38e-06 mean_abs_error=2.82e-07 PASSED (FP32 atol=1e-4 rtol=1e-3)
python .opencode/skills/npu-adapt-skills/scripts/compare_outputs.py --cpu /tmp/cpu_s1_logits.npy --npu /tmp/npu_s1_logits.npy --task regression --atol 1e-4 --rtol 1e-3结果: max_abs_error=1.34e-05 mean_abs_error=2.21e-06 PASSED
torch.multinomial RNG 实现差异会有统计波动(CPU mean open 10.407 vs NPU 10.732),属预期现象;通过 sample_logits=False (greedy argmax) 可获得完全一致的输出,已验证 logits 一致性蕴含 greedy 一致性。单样本 smoke consistency 仅证明算子对齐,不代表全数据集精度评测。设备 npu:0, batch 1, lookback 512, pred_len 30, max_context 2048, float32,计时含 torch.npu.synchronize(),不含数据合成与 pandas 预处理(归一化已计入 generate 前):

三图由 scripts/render_xterm_evidence.mjs 从真实日志 xterm.js 渲染,固定提示符 atomgit@pod-a94f8701860f4700b161b00e290de466:~$ 仅为展示标签。
torch_npu 与 npu:0 强依赖,CPU 仅用于一致性对比max_context (mini 2048) 限制,超长输入会被截断到最后 2048 tokenT=1.0 top_p="0.9" 有随机性;需复现请固定 torch.manual_seed 与 torch.npu.manual_seed_allclip((x-mean)/(std+1e-5), -5, 5)sample_count 或 pred_len > 1024 时的显存溢出,需分批