m
mengsihan/deepsea-npu
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

multimolecule/deepsea on Ascend NPU

模型与任务

  • 模型:multimolecule/deepsea
  • Revision:7025225200b133f2c21f0e2ed26efe54bbcd3262
  • 适配路线:regulatory-sequence
  • 设备:npu:0

环境依赖

{
  "python": "3.11.14",
  "torch": "2.9.0",
  "torch_npu": "2.9.0.post1",
  "transformers": "5.9.0",
  "huggingface_hub": "1.28.0",
  "tokenizers": "0.22.0",
  "multimolecule": "0.2.1",
  "safetensors": "0.8.0",
  "numpy": "2.4.6",
  "npu_name": "Ascend910_9362",
  "npu_count": 2,
  "cann": "8.5.1",
  "os": "linux-aarch64"
}

安装:pip install -r requirements.txt

NPU 推理

默认执行:python inference.py

真实输入

{
  "sequence_source": "hg38 chr7:140453130-140454130 (GRCh38, UCSC Genome Browser)",
  "sequence_length": 1000,
  "sequence_sha256": "989a29126896b2f0fc171f0ecf2e3c904a903b37a34947d2c643584fd985782c",
  "sequence_head": "ATTTGTTTTTTCTTAAAGTGAAAGAATGGGTGGGATCCAAGGAATCAAAG...TCCCATGTCTGGGGCAGAGGGCTATTGGCCAGCTTAGGTCCCTTCACCCA",
  "alphabet": "ACGTN",
  "batch_size": 1,
  "tokenizer": "DnaTokenizer",
  "model_dir": "./model_dir"
}

真实推理结果

{
  "logits_shape": [
    1,
    919
  ],
  "logits_dtype": "float32",
  "probs_shape": [
    1,
    919
  ],
  "top5_chromatin_feature_indices": [
    917,
    894,
    872,
    875,
    861
  ],
  "top5_chromatin_feature_probs": [
    0.583474,
    0.532285,
    0.523805,
    0.519651,
    0.516672
  ],
  "active_features_p_gt_0_5": 6,
  "num_labels": 919,
  "output_device": "npu:0 -> cpu postprocess"
}

CPU-NPU 一致性

{
  "passed": true,
  "logits_max_abs_diff": 0.000949859619140625,
  "logits_mean_abs_diff": 0.00021286497259425391,
  "probs_max_abs_diff": 0.00012725591659545898,
  "probs_mean_abs_diff": 3.924125849152915e-06,
  "atol": 0.001,
  "rtol": 0.001,
  "cpu_logits_sha256": "b7ab6c364a156580789ae55ed9d7d1131f4f9c38842848197880f6f3cb4dcd65",
  "npu_logits_sha256": "5d185ca3f8e6eb276d3bb7537dc2f26d2cc91df3439df6162ef16d260c4b2e85"
}

性能

{
  "warmup_runs": 3,
  "stable_runs": 10,
  "mean_latency_ms": 2.4798,
  "sequences_per_second": 403.26,
  "bases_per_second": 403263.46,
  "peak_npu_memory_allocated_mb": 289.94,
  "peak_npu_memory_reserved_mb": 331.35,
  "device": "npu:0",
  "dtype": "float32",
  "batch_size": 1,
  "sequence_length": 1000
}

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。

标签

#NPU #Ascend #Ascend910