m
mengsihan/deepcpgdna-hou2016-mesc
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

multimolecule/deepcpgdna-hou2016-mesc on Ascend NPU

模型与任务

  • 模型:multimolecule/deepcpgdna-hou2016-mesc
  • Revision:a21c2965ee47880c49d923766bbfdc657a707c64
  • 适配路线:regulatory-sequence
  • 设备:npu:0

环境依赖

{
  "CANN": "8.5.1",
  "torch": "2.5.0",
  "torch_npu": "available",
  "transformers": "4.40+",
  "multimolecule": "0.4.3",
  "npu_name": "Ascend910_9362",
  "npu_count": 2,
  "platform": "Atlas 800 A2 (Ascend910)"
}

安装:pip install -r requirements.txt

NPU 推理

默认执行:python inference.py

真实输入

{
  "alphabet": "A/C/G/T/N (N=all-zero, pad)",
  "sequence_length": 1001,
  "strand": "5'->3'",
  "padding": "right-pad N to 1001, no truncation beyond 1001",
  "tokenizer": "DnaTokenizer 1-mer, vocab_size=5, pad_token=N",
  "sequences": [
    {
      "name": "p53_fragment",
      "length": 1001,
      "sha256_prefix": "72558a15e155d8dd",
      "source": "model card tumor protein p53 extended to 1001"
    },
    {
      "name": "BRCA1_fragment",
      "length": 1001,
      "sha256_prefix": "83c71b4c3f0ad74a",
      "source": "model card BRCA1"
    },
    {
      "name": "HBB_fragment",
      "length": 1001,
      "sha256_prefix": "3b3517a325b4a251",
      "source": "model card hemoglobin subunit beta"
    },
    {
      "name": "synthetic_ACGT",
      "length": 1001,
      "sha256_prefix": "f8f15f3762978b8a",
      "source": "ACGT*250+A"
    },
    {
      "name": "synthetic_polyC",
      "length": 1001,
      "sha256_prefix": "cf8aabbc1bd3a50f",
      "source": "C*1001"
    }
  ],
  "batch": 1,
  "dtype": "float32",
  "revision_source": "local snapshot a21c2965ee47880c49d923766bbfdc657a707c64"
}

真实推理结果

{
  "task": "per-cell single-cell methylation logits (binary) 6 cells: mESC1..mESC6",
  "output_shape": "[1,6]",
  "output_dtype": "float32",
  "example_logits_cpu": {
    "p53_fragment": [
      -0.1924,
      0.0799,
      -0.1036,
      -0.1836,
      0.0024,
      -0.142
    ],
    "BRCA1_fragment": [
      2.0551,
      2.0831,
      2.2459,
      2.0745,
      2.432,
      2.0259
    ],
    "HBB_fragment": [
      2.7851,
      2.7452,
      2.9012,
      2.7535,
      3.4083,
      2.792
    ],
    "synthetic_ACGT": [
      -2.8783,
      -3.3375,
      -3.1813,
      -2.6691,
      -3.2541,
      -3.0509
    ],
    "synthetic_polyC": [
      0.1845,
      0.1248,
      0.4128,
      -0.3846,
      -0.0184,
      -1.0491
    ]
  },
  "example_probs_sigmoid": {
    "p53_fragment": [
      0.452,
      0.52,
      0.4741,
      0.4542,
      0.5006,
      0.4646
    ],
    "synthetic_ACGT": [
      0.0532,
      0.0343,
      0.0399,
      0.0648,
      0.0372,
      0.0452
    ]
  },
  "postprocess": "sigmoid -> per-cell methylation probability, >0.5 methylated",
  "device_proof": "logits device npu:0 verified, first param on npu:0"
}

CPU-NPU 一致性

{
  "comparison": "CPU vs NPU per-cell logits, same weights, same sequences, float32",
  "metric": "max_abs_diff and mean_abs_diff",
  "atol": 0.005,
  "rtol": 0.0,
  "per_sequence": {
    "p53_fragment": {
      "max_abs_diff": 0.000137,
      "mean_abs_diff": 0.0001,
      "pass": true
    },
    "BRCA1_fragment": {
      "max_abs_diff": 7.9e-05,
      "mean_abs_diff": 7.2e-05,
      "pass": true
    },
    "HBB_fragment": {
      "max_abs_diff": 3.3e-05,
      "mean_abs_diff": 2e-05,
      "pass": true
    },
    "synthetic_ACGT": {
      "max_abs_diff": 0.002193,
      "mean_abs_diff": 0.001699,
      "pass": true
    },
    "synthetic_polyC": {
      "max_abs_diff": 0.000283,
      "mean_abs_diff": 0.000245,
      "pass": true
    }
  },
  "worst_max_abs_diff": 0.002193,
  "overall_pass": true,
  "sign_agreement": true
}

性能

{
  "warmup_iters": 3,
  "measured_iters": 10,
  "batch": 1,
  "sequence_length": 1001,
  "avg_latency_ms": 1.02,
  "throughput_seq_per_s": 980.0,
  "bases_per_s": 981072,
  "dtype": "float32",
  "device": "npu:0",
  "synchronize": "torch.npu.synchronize before and after timing",
  "peak_memory_note": "model params 4.1M (~16MB), activation small, HBM < 3110 MB baseline"
}

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。
  • 固定 1001bp 窗口,不支持可变长 padding 截断冒充。
  • NPU 浮点误差在监管序列模型预期范围内,atol=5e-3。

标签

#NPU #Ascend #Ascend910