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

multimolecule/hal on Ascend NPU

模型与任务

  • 模型:multimolecule/hal
  • Revision:8d9a34ef9a5df3a44a832786d069020d269e9208
  • 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": "ACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGU",
  "sequence_alt": "ACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGCCCGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGUACGU",
  "sequence_telomerase_160": "GGGUUGCGGAGGGUGGGCCUGGGAGGGGUGGUGGCCAUUUUUUGUCUAACCCUAACUGAGAAGGGCGUAGGCGCCGUGCUUUUGCUCCCCGCGCGCUGUUUUUCUCGCUGACUUUCAGCGGGCGGAAAAGCCUCGGCCUGCCGCCUUCCACCGUUCAUUCUAG",
  "sequence_ref_sha256": "ea3085513a5844442a654c669e7c6db8c2ac35e96bc2616d84912ae0aef8a1bf",
  "sequence_alt_sha256": "318754da6d29dcaa6a3fc6d03632e9c89454faf4cac048b22a3e53b72afe84a9",
  "sequence_telomerase_sha256": "ad6cde49ba3dfd8c32b27622e440363b3068a0d3ded231a2fa654bf68fc62444",
  "length": 160,
  "length_alt": 160,
  "length_telo": 160,
  "alphabet": "ACGU",
  "variant": "pos 80 (1-indexed) ref U -> alt C single SNV in 160 nt donor window",
  "tokenizer": "RnaTokenizer",
  "vocab_size": 5,
  "replace_T_with_U": true,
  "source": "synthetic ACGU*40 donor window (160 nt balanced) + single SNV variant + telomerase RNA 160 nt realistic sequence from model card trimmed to region_length",
  "attention_mask": "implicit all ones, shape [1,160] for all sequences",
  "region_length": 160,
  "kmer_size": 6,
  "hexamer_features": 4096
}

真实推理结果

{
  "logits_shape": [
    1,
    1
  ],
  "output_type": "scalar splicing score per 160 nt donor window (regression); hexamer linear model with 4096 hexamer weights averaged over 8 columns",
  "logits_ref_cpu": 0.183743,
  "logits_alt_cpu": 0.181706,
  "logits_telo_cpu": 0.010065,
  "logits_ref_npu": 0.183743,
  "logits_alt_npu": 0.181706,
  "logits_telo_npu": 0.010065,
  "delta_cpu": -0.002037,
  "delta_npu": -0.002037,
  "sigmoid_delta_cpu": 0.499491,
  "sigmoid_delta_npu": 0.499491,
  "dtype": "float32",
  "hexamer_weights": 4096,
  "num_params_M": 0.004,
  "flops": 8192
}

CPU-NPU 一致性

{
  "logits_ref_max_abs_diff": 0.0,
  "logits_ref_mean_abs_diff": 0.0,
  "logits_alt_max_abs_diff": 0.0,
  "logits_alt_mean_abs_diff": 0.0,
  "logits_telo_max_abs_diff": 0.0,
  "delta_max_abs_diff": 0.0,
  "delta_mean_abs_diff": 0.0,
  "atol": 0.0001,
  "rtol": 0.001,
  "passed": true,
  "note": "HAL linear hexamer model: CPU vs NPU scalar outputs identical within 1e-6; explicit npu:0 placement with torch.npu.synchronize timing; variant effect delta verified"
}

性能

{
  "sequence_length": 160,
  "total_bases_per_run": 160,
  "output_shape": [
    1,
    1
  ],
  "first_latency_s": 0.229,
  "stable_avg_latency_s": 0.0276,
  "sequence_per_s": 36.21,
  "bases_per_s": 5793.54,
  "warmup_iters": 3,
  "stable_iters": 10,
  "dtype": "float32",
  "device": "npu:0",
  "model_params_M": 0.004,
  "flops": 8192
}

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 synthetic ACGU*40 160 nt window with single SNV and telomerase RNA 160 nt window; HAL scoring is linear over normalized hexamer frequencies across fixed 160 nt donor window.

标签

#NPU #Ascend #Ascend910