gcw_6a3w2yXz/tcrt5_ft_tcrdb
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

dkarthikeyan1/tcrt5_ft_tcrdb on Ascend NPU

模型与任务

  • 模型:dkarthikeyan1/tcrt5_ft_tcrdb
  • Revision:f559f8f7f1862d23f863b0d0d3b2f858cdab9705
  • 适配路线:protein-conditional-generation
  • 设备:npu:0 (Ascend910_9362)

环境依赖

{
  "torch": "2.9.0+cpu",
  "torch_npu": "2.9.0.post1+gitee7ba04",
  "transformers": "4.57.6",
  "device_name": "Ascend910_9362",
  "weights_sha256": "8d702775dfdffa16f77fedef3301672260b9fa2793ce6d6f679d3f09ac61b26b",
  "dtype": "float32"
}

安装:pip install -r requirements.txt

NPU 推理

默认执行:python inference.py

真实输入

{
  "pmhc_example": "[PMHC]KLGGALQAK[SEP]YFAMYQENVAQTDVDTLYIIYRDYTWAELAYTWY[EOS]",
  "peptide": "KLGGALQAK",
  "mhc_pseudosequence_hla": "HLA-A*03:01",
  "source": "official model card usage example",
  "target_tcr_teacher_forced": "[TCR]CASSLGYNEQFF[EOS]",
  "encoder_shape": [
    1,
    46
  ]
}

真实推理结果

{
  "logits_shape": [
    1,
    13,
    128
  ],
  "greedy_cdr3b": [
    "CASSLGGGGTDTQYF"
  ],
  "beam_cdr3b_top": [
    "CASSLGTGGTDTQYF",
    "CASSPGTGGTDTQYF",
    "CASSLGQGGTEAFF",
    "CASSVGTGGTDTQYF",
    "CASSLGTGGSYEQYF",
    "CASSPGQGGTEAFF",
    "CASSSGTGGTDTQYF",
    "CASSLGGGGTDTQYF",
    "CASSLGGGSYEQYF",
    "CASSLGTGGNQPQHF"
  ],
  "validity_rate": 1.0
}

CPU-NPU 一致性

{
  "metric": "teacher-forced logits + NLL + greedy decode match",
  "max_abs_diff": 1.9073486328125e-05,
  "mean_abs_diff": 3.4129381954419214e-06,
  "atol": 0.001,
  "rtol": 0.01,
  "logits_ok": true,
  "cpu_nll": 0.7684825658798218,
  "npu_nll": 0.7684828042984009,
  "nll_abs_diff": 2.384185791015625e-07,
  "npu_perplexity": 2.156491994857788,
  "nll_ok": true,
  "greedy_exact_match_rate": 1.0,
  "cpu_fallback": false
}

性能

{
  "warmup_runs": 3,
  "timed_runs": 10,
  "latency_s_first": 0.3259,
  "latency_s_stable_mean": 0.3228,
  "generated_sequences_per_s": 30.95,
  "peak_npu_memory_mb": 359.1
}

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

  • 仅对固定 revision f559f8f7f1862d23f863b0d0d3b2f858cdab9705、官方模型卡示例输入与 float32 精度完成验证。
  • 一致性阈值 (logits atol=1e-3, rtol=1e-2; NLL atol=5e-3) 在执行前依据 fp32 与算子误差固定,未事后调整。
  • 单样本 smoke consistency + 生物学有效性检查,非完整基准评测。

标签

#NPU #Ascend #Ascend910