m
mengsihan/borzoi-mouse
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

multimolecule/borzoi-mouse on Ascend NPU

模型与任务

  • 模型:multimolecule/borzoi-mouse
  • Revision:2a03d813289ccb49f98879c1c91f88888c94ca7a
  • 适配路线:long-context-tracks
  • 设备:npu:0

环境依赖

{
  "torch": "2.9.0+cpu",
  "torch_npu": "2.9.0.post1+gitee7ba04",
  "python": "3.11.14",
  "npu_name": "Ascend910_9362",
  "device_count": 2,
  "cann": "8.5.1",
  "multimolecule": "0.2.1",
  "transformers": "5.9.0"
}

安装:pip install -r requirements.txt

NPU 推理

默认执行:python inference.py

真实输入

{
  "sequence_preview": "CATTTGCTTCTGACACAACTGTGTTCACTAGCAACCTCAAACAGACACCATGGTGCATCTGACTCCTGAGGAGAAGTCTGCCGTTACTGCCCTGTGGGGCAAGGTGAACGTGGATGAAGTTGGTGGTGAGGCCCTGGGCAGG...(repeated)",
  "sequence_length": 524288,
  "sha256": "547ffa5481bb2a7824f8fdf216c321d1cc1cd520760988c60dd78c4febff98f5",
  "vocab": [
    "A",
    "C",
    "G",
    "T",
    "N"
  ],
  "tokenizer": "DnaTokenizer",
  "target_length": 6144,
  "num_tracks": 2608,
  "bin_size": 32,
  "species": "mouse",
  "input_ids_shape": [
    1,
    524288
  ],
  "dtype": "int64"
}

真实推理结果

{
  "logits_shape": [
    1,
    6144,
    2608
  ],
  "coverage_shape": [
    1,
    6144,
    2608
  ],
  "dtype": "float32",
  "device": "npu:0",
  "logits_stats": {
    "min": -9.012091,
    "max": 4.025692,
    "mean": -2.252657,
    "std": 2.167655
  },
  "coverage_stats": {
    "min": 0.000122,
    "max": 4.043386,
    "mean": 0.351522
  },
  "finite_ratio": 1.0,
  "output_act": "softplus",
  "center_bin": 3072,
  "center_track0_logit": -4.303623,
  "center_track0_coverage": 0.013429
}

CPU-NPU 一致性

{
  "cpu_shape": [
    1,
    6144,
    2608
  ],
  "npu_shape": [
    1,
    6144,
    2608
  ],
  "max_abs_diff": 0.0040531158447265625,
  "mean_abs_diff": 0.0005342856898301472,
  "atol": 0.005,
  "rtol": 0.001,
  "passed": true,
  "correlation_sample": 0.999999985,
  "elements": 16023552,
  "dtype": "float32"
}

性能

{
  "warmup_iters": 3,
  "stable_iters": 10,
  "avg_per_window_s": 0.2308,
  "windows_per_s": 4.333,
  "bases_per_s": 2271601.8,
  "elapsed_total_s": 2.308,
  "input_length": 524288,
  "output_shape": [
    1,
    6144,
    2608
  ],
  "dtype": "float32",
  "peak_memory_MB": 3737.7,
  "device": "npu:0",
  "first_compile_included": true
}

自验证截图

Agent workflow

NPU device call

Model result

三张图片均由 xterm.js 根据本次真实日志生成。

已知限制

  • 仅对本次记录的模型 revision、输入契约、dtype 与 Ascend 环境完成验证。
  • 长上下文 524288 bp 窗口已在单卡 Ascend910 (64GB) 验证通过;更小显存设备需单独评估。
  • 2608 mouse tracks 全量输出已验证;下游任务需按 track 语义筛选。

标签

#NPU #Ascend #Ascend910