gcw_6a3w2yXz/multimolecule-mtsplice-NPU
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

multimolecule/mtsplice on Ascend NPU

模型与任务

  • 模型:multimolecule/mtsplice
  • Revision:a3826976df62ae133f72d72a8c13d9b18777841b
  • 适配路线:splicing
  • 设备:npu:0

环境依赖

{
  "torch_version": "2.9.0+cpu",
  "torch_npu_version": "2.9.0.post1+gitee7ba04",
  "device_name": "Ascend910_9362",
  "device_count": 2,
  "npu_available": true,
  "python_version": "3.11.14"
}

安装:pip install -r requirements.txt

NPU 推理

默认执行:python inference.py

真实输入

{
  "sequence_ref": "AUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGC",
  "sequence_alt": "AUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGGAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGCAUGCUAGCUAGCUAGCUAGC",
  "sequence_ref_sha256": "68c0e574d015c6c80ccfa65641f48670f1e4f27888cdae383c9220b6cdf54f66",
  "sequence_alt_sha256": "b6fb8aff51e45518906c1990f255748da27562d3552ab3a3464e7701997e94c3",
  "length": 800,
  "length_alt": 800,
  "alphabet": "ACGU",
  "variant": "pos 400 ref C -> alt G at acceptor/donor boundary (single SNV, cassette exon window)",
  "tokenizer": "RnaTokenizer",
  "vocab_size": 5,
  "replace_T_with_U": true,
  "source": "synthetic 800 nt cassette-exon-with-flanks window (AUGCUAGC*40), single SNV variant; diverse A/C/G/U for splicing task",
  "attention_mask": "all ones, shape [1,800] for both ref and alt",
  "acceptor_length": 400,
  "donor_length": 400
}

真实推理结果

{
  "logits_shape": [
    1,
    56
  ],
  "delta_logits_shape": [
    1,
    56
  ],
  "alternative_logits_shape": [
    1,
    56
  ],
  "output_type": "tissue-specific delta-logit-PSI vector across 56 GTEx tissues (regression); ref logits + delta = alt",
  "ref_logits_mean_cpu": 0.101872,
  "ref_logits_max_cpu": 3.802699,
  "ref_logits_min_cpu": -0.318739,
  "ref_logits_mean_npu": 0.101876,
  "ref_logits_max_npu": 3.802851,
  "delta_mean_cpu": 6.6e-05,
  "delta_max_cpu": 0.005666,
  "delta_min_cpu": -0.004431,
  "delta_abs_max_cpu": 0.005666,
  "delta_mean_npu": 6.6e-05,
  "delta_max_npu": 0.005662,
  "ref_logits_sample_cpu": [
    0.708108127117157,
    -0.08909831941127777,
    0.1294664740562439,
    -0.03364101052284241,
    -0.3187394440174103
  ],
  "delta_sample_cpu": [
    -0.00304412841796875,
    0.0008879676461219788,
    0.0014516264200210571,
    0.0013770461082458496,
    0.0007616281509399414
  ],
  "top5_abs_delta_indices_cpu": [
    51,
    13,
    12,
    15,
    14
  ],
  "top5_abs_delta_values_cpu": [
    0.005666017532348633,
    0.00443112850189209,
    0.00425410270690918,
    0.003945166245102882,
    0.0036490261554718018
  ],
  "dtype": "float32",
  "tissues": 56,
  "example_tissues": "GTEx v8 56 tissues (adipose, muscle, brain, liver, etc., index 0-55)"
}

CPU-NPU 一致性

{
  "ref_logits_max_abs_diff": 0.000152,
  "ref_logits_mean_abs_diff": 3.1e-05,
  "delta_max_abs_diff": 4e-06,
  "delta_mean_abs_diff": 3e-07,
  "atol_logits": 0.001,
  "atol_delta": 0.0001,
  "rtol": 0.001,
  "passed": true,
  "note": "All 56 tissue outputs compared; explicit npu:0 placement with torch.npu.synchronize timing; splice variant effect delta verified"
}

性能

{
  "sequence_length": 800,
  "sequence_length_alt": 800,
  "total_bases_per_run": 1600,
  "output_shape": [
    1,
    56
  ],
  "first_latency_s": 0.231,
  "stable_avg_latency_s": 0.0048,
  "sequence_per_s": 209.87,
  "bases_per_s": 335792.81,
  "warmup_iters": 3,
  "stable_iters": 10,
  "dtype": "float32",
  "device": "npu:0",
  "model_params_M": 0.211,
  "flops_M": 164.36
}

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

  • Only validated on 800 nt window with single SNV; GTEx tissue list is index-labeled (LABEL_0..55) without explicit tissue name mapping in checkpoint.

标签

#NPU #Ascend #Ascend910