• G
    Clean fluid APIs in distributed and fleet files (#48851) · 911d6bb1
    Ghost Screaming 提交于
    * Fix bug of reduce_sum op. When input.numel() > INT32_MAX, its result
    is wrong.
    
    * Remove climits.
    
    * Clean fluid API in paddle/distributed and paddle/fleetx folders.
    Include following files:
    python/paddle/distributed/__init__.py
    python/paddle/distributed/collective.py
    python/paddle/distributed/fleet/utils/fs.py
    python/paddle/distributed/fleet/utils/hybrid_parallel_inference.py
    python/paddle/distributed/fleet/utils/hybrid_parallel_util.py
    python/paddle/distributed/fleet/utils/internal_storage.py
    python/paddle/distributed/launch/context/device.py
    python/paddle/distributed/parallel.py
    python/paddle/distributed/parallel_with_gloo.py
    python/paddle/distributed/spawn.py
    python/paddle/framework/__init__.py
    To be mentioned, 'paddle.fluid.dygraph.parallel.ParallelEnv'
     and 'fluid.framework.core' keeps unchanged in those files.
    ParallelEnv is used by paddle.fluid.dygraph.parallel.DataParallel.
    However, APIs in paddle.fluid.dygraph.parallel can't be
    migrated to paddle.distributed, as there exists cyclic import
    dependencies in modules like paddle.static, paddle.tensor. And
    'fluid.framework.core' will be changed to import framework.core
    after fluid.core is transmitted.
    
    * Change TODO authors.
    911d6bb1
hybrid_parallel_util.py 8.8 KB