硬件:Ascend NPU(Ascend 910B4) · 推理引擎:torch_npu · 设备:npu:0 精度验收:CPU vs NPU 推理精度差异 < 1%(实测 SKIP,cosine=1.0)
Model description TCRT5 is a seq2seq model designed to for the conditional generation of T-cell receptor (TCR) sequences given a target peptide-MHC (pMHC). It is a transformers model that is buil
| 属性 | 值 |
|---|---|
| 模型名 | dkarthikeyan1/tcrt5_ft_tcrdb |
| 原始权重 | ModelScope / HuggingFace / AtomGit 镜像 |
| 任务类型 | text-model |
| 输入 | 文本(token 长度 512) |
| 输出 | [batch, hidden] |
| 实测输出 shape | (1, 256) |
| 框架 | PyTorch + torch_npu |
| 官方实现 | https://github.com/google-research/text-to-text-transfer-transformer |
官方实现 / 论文仓库:
| 依赖项 | 版本 | 说明 |
|---|---|---|
| Python | 3.11 | conda 环境 dkarthikeyan1-tcrt5_ft_tcrdb-npu |
| torch / torch_npu | 2.10.0 / 2.10.0 | 由 setup_env.sh 装,须严格配套 |
| transformers | ==4.51.3 | |
| accelerate | ==1.14.0 | |
| pillow | ==12.3.0 | |
| numpy | ==1.26.4 | |
| huggingface-hub | ==0.36.2 | |
| requests | ==2.34.2 | |
| modelscope | ==1.37.0 | |
| safetensors | ==0.8.0 | |
| multimolecule | ==0.2.1 | |
| pyyaml | >=6.0 | |
| einops | ==0.8.0 # transformers 4.51 modeling 隐式依赖(torch_npu 存在时触发) | |
| torchvision | ==0.25.0 | |
| sentencepiece | ==0.2.1 # T5Tokenizer(spiece.model)必需 |
一键建环境(从零创建 conda 环境并装依赖):
bash setup_env.sh # conda create -n <env> python=<ver> + torch/torch_npu + requirements.txt⚠️ 重要:若环境已装 torch/torch_npu,不要强制覆盖 torch,否则破坏 torch_npu 算子插件绑定。
国内 pip 源(Tsinghua):https://pypi.tuna.tsinghua.edu.cn/simple
权重不随仓库分发,推理前先用国内下载源拉取到本地缓存(推荐 ModelScope,失败回落 HF 镜像):
# 方式一:ModelScope(国内直连,最快)
pip install modelscope -i https://pypi.tuna.tsinghua.edu.cn/simple
python -c "from modelscope import snapshot_download; snapshot_download('dkarthikeyan1/tcrt5_ft_tcrdb', cache_dir='./weights')"
# 方式二:HF 镜像(hf-mirror.com)
export HF_ENDPOINT=https://hf-mirror.com
huggingface-cli download dkarthikeyan1/tcrt5_ft_tcrdb --local-dir ./weights/tcrt5_ft_tcrdbinference.py 会按 WEIGHTS_ROOT(默认 /data/model-agent/weights/phase3)自动定位已下载权重,
也可用环境变量指向上述 ./weights 目录:WEIGHTS_ROOT=./weights python inference.py
硬件:单卡 Ascend 910B4(npu:1,物理卡 5)
评委裸 clone 后本机没有 conda 环境,conda activate 不会创建环境。必须先建环境再推理:
# 第一步:从零创建 conda 环境并装依赖(conda create + torch/torch_npu + requirements.txt)
bash setup_env.sh
# 第二步:激活环境并一键推理(环境自检 + inference + eval + benchmark)
conda activate dkarthikeyan1-tcrt5_ft_tcrdb-npu
bash run.shimport torch, torch_npu
from transformers import AutoModel, AutoTokenizer
m = AutoModel.from_pretrained("dkarthikeyan1/tcrt5_ft_tcrdb", trust_remote_code=True).to("npu:0").eval()
tok = AutoTokenizer.from_pretrained("dkarthikeyan1/tcrt5_ft_tcrdb", trust_remote_code=True)
inputs = tok("How to use this model?", return_tensors="pt").to("npu:0")
outputs = m(**inputs)logs/inference.log)输入:示例序列/文本(见 inference.py)
输出:
=== Inference: dkarthikeyan1/tcrt5_ft_tcrdb on Ascend NPU ===
NPU: npu:0
权重目录: /data/model-agent/weights/phase3/dkarthikeyan1/tcrt5_ft_tcrdb
pipeline_tag: feature-extraction | library: transformers | is_text=True
输入文本: How to use this model?
输入 ids shape: (1, 512)
推理耗时: 340.19 ms
特征 shape: (1, 256)
特征 L2 norm: 70.9002
特征前 8 维: [-4.535874843597412, -3.8142306804656982, 3.5264973640441895, 9.094510078430176, 0.21025024354457855, -1.581990122795105, -0.5759623646736145, -6.450984001159668]logs/accuracy.log)=== CPU-NPU Accuracy: dkarthikeyan1/tcrt5_ft_tcrdb ===
Encoder-Decoder 模型,跳过 CPU-NPU 特征精度对比(输出为 token ids)精度对比:SKIP(encoder-decoder)
注:本次适配无精度要求,精度数据作为质量证据。详见
logs/accuracy.log。
| 指标 | 数值 |
|---|---|
| avg latency | 41.34 ms |
| p50 | 40.95 ms |
| throughput | 24.19 images/sec |



截图由真实执行日志(
logs/)忠实渲染为终端风格 PNG,渲染脚本tools/gen_screenshots_pw.py(playwright),可复现,无虚构结果。
#NPU #Ascend #Ascend910 #model-agent-tagged
仓库:https://gitcode.com/v_50/dkarthikeyan1-tcrt5_ft_tcrdb-NPU 适配日期:2026-08-12 | 赛道:第二季·模型适配·阶段三