TRIBE v2 Ascend NPU 适配

docker pull swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:8.5.1-a3-ubuntu22.04-py3.11

docker run -d \
  --privileged \
  --device /dev/davinci_manager \
  --device /dev/devmm_svm \
  --device /dev/hisi_hdc \
  --device /dev/davinci0 \
  --device /dev/davinci1 \
  -v /data:/data \
  -v /usr/local/Ascend:/usr/local/Ascend \
  --name tribev2_npu \
  swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:8.5.1-a3-ubuntu22.04-py3.11

# 进入容器
docker exec -it tribev2_npu bash

# 设置 CANN 环境变量
source /usr/local/Ascend/ascend-toolkit/set_env.sh

# 克隆 TRIBE v2 仓库
cd /data
git clone https://github.com/facebookresearch/tribev2.git

# 创建 exca 补丁文件（解决 exca 0.5.17 与 neuralset 的兼容性问题）
cat > /data/tribev2/exca_patch.py << 'EOF'
#!/usr/bin/env python3
"""
exca 0.5.17 Compatibility Patch for Python 3.10

Problem: neuralset 0.0.2 requires exca >= 0.5.20, but exca 0.5.20+
requires Python >= 3.11. This patch adds missing methods to exca 0.5.17.

Methods added:
1. _get_discriminated_subclasses() - Class dispatcher for DiscriminatedModel
2. __new__() - Subclass instantiation method
"""
from exca import helpers

@classmethod
def _get_discriminated_subclasses(cls) -> dict:
    """Get all subclasses with the same discriminator key."""
    key = cls._exca_discriminator_key
    val_classes = {}
    for s in helpers._get_subclasses(cls) + [cls]:
        if s._exca_discriminator_key != key:
            continue
        val = s.__name__
        past = val_classes.get(val, None)
        if past is not None and past.__module__ != s.__module__:
            raise RuntimeError(f"2 subclasses from different modules are named {val!r}")
        val_classes[val] = s
    return val_classes

def __new__(cls, /, **kwargs):
    """Instantiate the correct subclass based on discriminator."""
    key = cls._exca_discriminator_key
    type_name = kwargs.get(key)
    if type_name is not None and type_name != cls.__name__:
        target = cls._get_discriminated_subclasses().get(type_name)
        if target is not None:
            return super(helpers.DiscriminatedModel, target).__new__(target)
    return super(helpers.DiscriminatedModel, cls).__new__(cls)

# Apply patch
helpers.DiscriminatedModel._get_discriminated_subclasses = _get_discriminated_subclasses
helpers.DiscriminatedModel.__new__ = __new__

print("exca 0.5.17 patch applied successfully")
EOF

cd /data/tribev2

# 安装基础依赖
pip3 install pydantic==2.10.6

# 安装 exca 0.5.17（neuralset 需要 exca>=0.5.20，但 0.5.20+ 需要 Python 3.11）
pip3 install exca==0.5.17

# 安装 neuralset/neuraltrain（跳过依赖检查，因为 exca 已通过补丁兼容）
pip3 install neuralset==0.0.2 neuraltrain==0.0.2 --no-deps

# 安装 transformers 及兼容版本（最新版支持 V-JEPA2）
pip3 install transformers --upgrade
pip3 install huggingface_hub==0.29.0 tokenizers==0.21.2

# 安装 CANN TBE 必需依赖
pip3 install decorator scipy attrs psutil sympy cffi pyyaml

# 应用 exca 补丁（解决 exca 0.5.17 与 neuralset 的兼容性问题）
python3 exca_patch.py

# 安装 ModelScope
pip install modelscope

# 下载 LLaMA 3.2-3B
modelscope download --model LLM-Research/Llama-3.2-3B-Instruct --local_dir /data/cache/modelscope/LLM-Research/Llama-3___2-3B-Instruct

# 创建 HuggingFace 缓存映射
mkdir -p ~/.cache/huggingface/hub/models--meta-llama--Llama-3.2-3B/snapshots
ln -s /data/cache/modelscope/LLM-Research/Llama-3___2-3B-Instruct ~/.cache/huggingface/hub/models--meta-llama--Llama-3.2-3B/snapshots/main

# 使用中国镜像加速
export HF_ENDPOINT=https://hf-mirror.com

# Wav2Vec-BERT 2.0
huggingface-cli download facebook/w2v-bert-2.0 --local-dir /data/cache/models--facebook--w2v-bert-2.0

# V-JEPA2
huggingface-cli download facebook/vjepa2-vitg-fpc64-256 --local-dir /data/cache/models--facebook--vjepa2-vitg-fpc64-256

cd /data/tribev2
python run_inference_npu.py

TRIBE v2 NPU Inference Test
============================
Device: npu:0 (Ascend 910C)
Model loaded successfully!

Text Encoder (LLaMA 3.2-3B):
  Output shape: [1, 10, 3072]
  Latency: 24.88 ms

Audio Encoder (Wav2Vec-BERT 2.0):
  Output shape: [1, 100, 1024]
  Latency: 75.93 ms (seq=1000)

Video Encoder (V-JEPA2 ViT-G):
  Output shape: [1, 1024, 1408]
  Latency: 420.25 ms (frames=32)

All encoders verified on NPU!

# exca_patch.py 为 helpers.DiscriminatedModel 添加：
# 1. _get_discriminated_subclasses() - 类分发器方法
# 2. __new__() - 子类实例化方法

import sys
sys.path.insert(0, "/data/tribev2")
import exca_patch  # 必须在导入 neuralset 前执行

docker run --cap-add=SYS_PTRACE --cap-add=IPC_LOCK --cap-add=SYS_ADMIN ...