导出Llama 3.2 1B SpinQuant模型到Executorch时遭遇ImportError：无法从torch.export导入default_decompositions

我现在在M1 MacBook Air（16G内存，系统是macOS Sequoia 15.1）的Anaconda环境里，尝试用Executorch在iPhone上部署Llama 3.2 1B SpinQuant模型，但遇到了这个错误：

ImportError: cannot import name 'default_decompositions' from 'torch.export' (/Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages/torch/export/__init__.py)

操作背景与步骤 #

我是按照Executorch的XNNPACK部署指南操作的，前面的环境搭建步骤都顺利完成：

1. 创建并激活Conda环境：

conda create -n et_xnnpack python=3.10.0
conda activate et_xnnpack

2. 克隆Executorch仓库并同步子模块：

git clone https://github.com/pytorch/executorch.git
cd executorch
git submodule sync
git submodule update --init

3. 安装依赖：

./install_requirements.sh

问题出在执行模型导出命令时，我用的是Meta发布的预量化INT4 SpinQuant Llama 3.2模型，导出命令如下：

python -m examples.models.llama.export_llama \
  --model "llama3_2" \
  --checkpoint "/Users/questionone/.llama/checkpoints/Llama3.2-1B-Instruct-int4-spinquant-eo8/consolidated.00.pth" \
  --params "/Users/questionone/.llama/checkpoints/Llama3.2-1B-Instruct-int4-spinquant-eo8/params.json" \
  -kv \
  --use_sdpa_with_kv_cache \
  -X \
  -d fp32 \
  --xnnpack-extended-ops \
  --preq_mode 8da4w_output_8da8w \
  --preq_group_size 32 \
  --max_seq_length 2048 \
  --preq_embedding_quantize 8,0 \
  --use_spin_quant native \
  --metadata '{"get_bos_id":128000, "get_eos_ids":[128009, 128001]}' \
  --output_name "llama3_2_spinquant.pte"

完整报错栈：

(et_xnnpack) questionone@curious-MacBookAir executorch % python -m examples.models.llama.export_llama \
  --model "llama3_2" \
  --checkpoint "/Users/questionone/.llama/checkpoints/Llama3.2-1B-Instruct-int4-spinquant-eo8/consolidated.00.pth" \
  --params "/Users/questionone/.llama/checkpoints/Llama3.2-1B-Instruct-int4-spinquant-eo8/params.json" \
  -kv \
  --use_sdpa_with_kv_cache \
  -X \
  -d fp32 \
  --xnnpack-extended-ops \
  --preq_mode 8da4w_output_8da8w \
  --preq_group_size 32 \
  --max_seq_length 2048 \
  --preq_embedding_quantize 8,0 \
  --use_spin_quant native \
  --metadata '{"get_bos_id":128000, "get_eos_ids":[128009, 128001]}' \
  --output_name "llama3_2_spinquant.pte"
Traceback (most recent call last):
  File "<frozen runpy>", line 198, in _run_module_as_main
  File "<frozen runpy>", line 88, in _run_code
  File "/Users/questionone/executorch/examples/models/llama/export_llama.py", line 14, in <module>
    from .export_llama_lib import build_args_parser, export_llama
  File "/Users/questionone/executorch/examples/models/llama/export_llama_lib.py", line 25, in <module>
    from executorch.devtools.etrecord import generate_etrecord
  File "/Users/questionone/executorch/devtools/__init__.py", line 7, in <module>
    import executorch.devtools.inspector as inspector
  File "/Users/questionone/executorch/devtools/inspector/__init__.py", line 9, in <module>
    from executorch.devtools.inspector._inspector import (
  File "/Users/questionone/executorch/devtools/inspector/_inspector.py", line 31, in <module>
    import executorch.devtools.etdump.schema_flatcc as flatcc
  File "/Users/questionone/executorch/devtools/etdump/schema_flatcc.py", line 18, in <module>
    from executorch.exir.scalar_type import ScalarType
  File "/Users/questionone/executorch/exir/__init__.py", line 9, in <module>
    from executorch.exir.capture import (
  File "/Users/questionone/executorch/exir/capture/__init__.py", line 9, in <module>
    from executorch.exir.capture._capture import (
  File "/Users/questionone/executorch/exir/capture/_capture.py", line 15, in <module>
    from executorch.exir.capture._config import CaptureConfig
  File "/Users/questionone/executorch/exir/capture/_config.py", line 15, in <module>
    from executorch.exir.passes import MemoryPlanningPass, ToOutVarPass
  File "/Users/questionone/executorch/exir/passes/__init__.py", line 19, in <module>
    from executorch.exir import control_flow, memory, memory_planning
  File "/Users/questionone/executorch/exir/control_flow.py", line 58, in <module>
    from executorch.exir.tracer import (
  File "/Users/questionone/executorch/exir/tracer.py", line 50, in <module>
    from torch.export import default_decompositions
ImportError: cannot import name 'default_decompositions' from 'torch.export' (/Library/Frameworks/Python.framework/Versions/3.11/lib/python3.11/site-packages/torch/export/__init__.py)

我已经确认et_xnnpack环境里安装了Executorch和PyTorch，甚至用export PYTHONPATH=/Users/questionone/executorch重置了路径，但错误还是存在。以下是我的环境包列表：

(et_xnnpack) questionone@curious-MacBookAir executorch % conda list -n et_xnnpack
# packages in environment at /opt/anaconda3/envs/et_xnnpack:
#
# Name                    Version                   Build  Channel
annotated-types           0.7.0                    pypi_0    pypi
anyio                     4.7.0                    pypi_0    pypi
attrs                     24.3.0                   pypi_0    pypi
blas                      1.0                    openblas  
blobfile                  3.0.0                    pypi_0    pypi
bzip2                     1.0.8                h80987f9_6  
ca-certificates           2024.11.26           hca03da5_0  
certifi                   2024.12.14               pypi_0    pypi
charset-normalizer        3.4.0                    pypi_0    pypi
click                     8.1.8                    pypi_0    pypi
cmake                     3.31.2                   pypi_0    pypi
distro                    1.9.0                    pypi_0    pypi
exceptiongroup            1.2.2                    pypi_0    pypi
execnet                   2.1.1                    pypi_0    pypi
executorch                0.5.0a0+82763a9          pypi_0    pypi
expecttest                0.3.0                    pypi_0    pypi
filelock                  3.16.1                   pypi_0    pypi
fire                      0.7.0                    pypi_0    pypi
flatbuffers               24.3.25                  pypi_0    pypi
fsspec                    2024.12.0                pypi_0    pypi
gmp                       6.2.1                hc377ac9_3  
gmpy2                     2.1.2           py310h8c48613_0  
h11                       0.14.0                   pypi_0    pypi
httpcore                  1.0.7                    pypi_0    pypi
httpx                     0.28.1                   pypi_0    pypi
huggingface-hub           0.27.0                   pypi_0    pypi
hypothesis                6.123.0                  pypi_0    pypi
idna                      3.10                     pypi_0    pypi
iniconfig                 2.0.0                    pypi_0    pypi
jinja2                    3.1.5                    pypi_0    pypi
libabseil                 20240116.2      cxx17_h313beb8_0  
libcxx                    14.0.6               h848a8c0_0  
libffi                    3.4.4                hca03da5_1  
libgfortran               5.0.0           11_3_0_hca03da5_28  
libgfortran5              11.3.0              h009349e_28  
libopenblas               0.3.21               h269037a_0  
libprotobuf               4.25.3               h514c7bf_0  
llama-models              0.0.63                   pypi_0    pypi
llama-stack               0.0.63                   pypi_0    pypi
llama-stack-client        0.0.63                   pypi_0    pypi
llvm-openmp               14.0.6               hc6e5704_0  
lxml                      5.3.0                    pypi_0    pypi
markdown-it-py            3.0.0                    pypi_0    pypi
markupsafe                3.0.2                    pypi_0    pypi
mdurl                     0.1.2                    pypi_0    pypi
mpc                       1.1.0                h8c48613_1  
mpfr                      4.0.2                h695f6f0_1  
mpmath                    1.3.0           py310hca03da5_0  
ncurses                   6.4                  h313beb8_0  
networkx                  3.4.2                    pypi_0    pypi
numpy                     1.21.3                   pypi_0    pypi
numpy-base                1.26.4          py310ha9811e2_0  
openssl                   3.0.15               h80987f9_0  
packaging                 24.2                     pypi_0    pypi
pandas                    2.0.3                    pypi_0    pypi
parameterized             0.9.0                    pypi_0    pypi
pillow                    11.0.0                   pypi_0    pypi
pip                       24.2            py310hca03da5_0  
pluggy                    1.5.0                    pypi_0    pypi
prompt-toolkit            3.0.48                   pypi_0    pypi
pyaml                     24.12.1                  pypi_0    pypi
pycryptodomex             3.21.0                   pypi_0    pypi

问题分析与解决方法 #

看起来这个问题的核心是PyTorch版本与Executorch版本不兼容，加上Conda环境隔离失效导致的：

1. 环境隔离问题 #

你创建的是Python 3.10的Conda环境，但报错里的Torch路径指向了系统全局的Python 3.11目录——这说明你的Conda环境没有被正确激活，系统全局的Python被优先调用了，导致环境内的Torch没有被使用。

2. API版本兼容问题 #

torch.export.default_decompositions这个API的位置在PyTorch版本中是有变化的：

• PyTorch 2.1及更早版本：不存在这个API，或者放在其他模块；
• PyTorch 2.2~2.4版本：这个API存在于torch.export模块下；
• PyTorch 2.5+版本：该API可能被移除或重命名。

而你环境里的Executorch是0.5.0a0版本，需要搭配PyTorch 2.3~2.4版本才能正常兼容。

具体解决步骤 #

步骤1：彻底重置Conda环境

# 退出当前环境
conda deactivate
# 删除旧环境
conda remove -n et_xnnpack --all -y
# 重新创建干净的Python 3.10环境
conda create -n et_xnnpack python=3.10.0 -y
conda activate et_xnnpack
# 验证Python路径是否为环境内的（输出应为/opt/anaconda3/envs/et_xnnpack/bin/python）
which python

步骤2：安装匹配版本的PyTorch和Executorch

# 安装PyTorch 2.3.1（和Executorch 0.5.0兼容的稳定版本）
pip install torch==2.3.1 torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
# 进入executorch目录，以开发模式安装本地包
cd executorch
pip install -e .
# 重新安装依赖
./install_requirements.sh

步骤3：验证API可用性

激活环境后，进入Python交互模式测试：

from torch.export import default_decompositions
print(default_decompositions)

如果没有报错，说明环境已经正常，此时再重新运行模型导出命令即可。

额外注意事项 #

• 不要手动设置PYTHONPATH为全局的executorch路径，pip install -e .已经将本地包链接到环境中，手动设置会导致路径混乱；
• 所有命令必须在激活et_xnnpack环境后执行，避免调用系统全局的Python。

备注：内容来源于stack exchange，提问作者swimcode