liucu/optimus5prime
模型介绍
文件和版本
Pull Requests
讨论
分析

multimolecule/optimus5prime 在 Ascend NPU 上

模型与任务

  • 模型:multimolecule/optimus5prime
  • 修订版本:main
  • 适配路线:utr-translation
  • 设备:npu:0

环境依赖

{
  "torch": "2.9.0+cpu",
  "torch_npu_available": true,
  "npu_device_name": "Ascend910_9362",
  "device_count": 2,
  "hf_endpoint": "https://hf-mirror.com",
  "python": "3.11",
  "multimolecule": "0.2.1"
}

安装:pip install -r requirements.txt

NPU 推理

默认执行:python inference.py

真实输入

{
  "sequence_length": 50,
  "vocab_size": 5,
  "alphabet": "RNA A/C/G/U/N pad_id 0 replace_T_with_U true",
  "strand": "5'->3' forward",
  "primary_sequence_name": "generic_50",
  "primary_sha256": "217b42655c86b30cd6b0835e3af600cca7664f3c77268b372779e7c5692e0e0f",
  "primary_seq": "GGGACAUAGCUAGCUAGCUAGCUAGCUAGCUAGCUAGCUAGCUAGCUAGC",
  "batch": 1,
  "dtype": "float32",
  "task": "mean ribosome load (MRL) regression scalar",
  "variant_pair": {
    "ref_sha256": "217b42655c86b30cd6b0835e3af600cca7664f3c77268b372779e7c5692e0e0f",
    "alt_sha256": "d0fe94bdccc4f3e9ee3997aa3ae20acc31282c6d3bd3a3f35507035906924d7c",
    "delta": 0.008986473083496094
  }
}

真实推理结果

{
  "logits_shape": [
    1,
    1
  ],
  "logits_value": [
    [
      1.2541437149047852
    ]
  ],
  "logits_scalar": 1.2541437149047852,
  "variant_delta_alt_minus_ref": 0.008986473083496094,
  "batch_5_logits_cpu": [
    1.2541533708572388,
    -1.3085063695907593,
    0.5420016050338745,
    -0.7953201532363892,
    0.15475201606750488
  ],
  "batch_5_logits_npu": [
    1.2541437149047852,
    -1.3085277080535889,
    0.5420478582382202,
    -0.7953556776046753,
    0.1546841412782669
  ]
}

CPU-NPU 一致性

{
  "max_abs_diff": 9.5367431640625e-06,
  "mean_abs_diff": 9.5367431640625e-06,
  "atol": 0.0001,
  "rtol": 0.0001,
  "passed": true,
  "batch_max_abs_diff": 6.787478923797607e-05
}

性能

{
  "avg_ms_per_iter": 0.7978352019563317,
  "p50_ms": 0.8023250265978277,
  "throughput_seq_per_s": 1253.3916748069653,
  "throughput_bases_per_s": 62669.583740348266,
  "warmup": 3,
  "stable_iters": 10,
  "batch": 1,
  "sequence_length": 50,
  "dtype": "float32"
}

自验证截图

Agent 工作流

NPU 设备调用

模型结果

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。

标签

#NPU #Ascend #Ascend910