hfai
=====

幻方萤火深度学习套件 ``hfai`` 提供了 **训练加速** 、 **环境管理** 与 **萤火调度** 三类核心接口与命令行工具，以此加速用户本地的训练，打通用户个人主机与萤火二号集群。


.. toctree::
   :maxdepth: 2
   :caption: Contents:



hfai.nn
----------------------
.. py:currentmodule:: hfai.nn
.. autosummary::
   :nosignatures:

   to_hfai
   to_torch
   bench
   compare
   LSTM
   LSTM_fullc
   GRU
   LayerNorm
   MultiheadAttention
   Dropout
   Hardtanh
   LogSoftmax
   Softmax
   Softmax2d
   Softmin
   Softplus
   ReLU
   ReLU6
   Threshold
   LeakyReLU
   RReLU
   Hardsigmoid
   Hardshrink
   Softshrink


hfai.nn.functional
------------------------
.. py:currentmodule:: hfai.nn.functional
.. autosummary::
   :nosignatures:

   set_replace_torch
   dropout
   hardtanh
   hardtanh_
   log_softmax
   softmax
   softmin
   softplus
   sync
   relu
   relu_
   relu6
   threshold
   threshold_
   leaky_relu
   leaky_relu_
   rrelu
   rrelu_
   hardsigmoid
   hardshrink
   softshrink
   abs
   abs_
   minimum
   maximum
   min
   max
   clip
   clip_
   clamp
   clamp_
   clamp_max
   clamp_max_
   clamp_min
   clamp_min_
   where
   masked_fill
   masked_fill_
   masked_select
   masked_scatter
   masked_scatter_
   scan
   associative_scan

hfai.datasets
--------------------------
.. py:currentmodule:: hfai.datasets
.. autosummary::
   :nosignatures:

   BaseDataset
   set_data_dir
   get_data_dir
   download
   ADE20k
   AlphafoldData
   CIFAR10
   CIFAR100
   CLUEForCLS
   CLUEForMLM
   CocoCaption
   CocoDetection
   CocoKeypoint
   CocoPanoptic
   ERA5
   GoogleConceptualCaption
   ImageNet
   ImageNet21K
   ImageNet1K
   ISC2021
   KITTIObject2D
   LJSpeech
   LTSF
   NuScenes
   OGB
   WeatherBench


hfai.distributed
-------------------
.. py:currentmodule:: hfai.distributed
.. autosummary::
   :nosignatures:

   init_process_group
   barrier
   gather
   broadcast
   all_gather
   reduce_scatter
   set_nccl_opt_level
   pairwise_apply


hfai.multiprocessing
------------------------
.. py:currentmodule:: hfai.multiprocessing
.. autosummary::
   :nosignatures:

   Process
   spawn
   fork


hfai.checkpoint
------------------
.. py:currentmodule:: hfai.checkpoint
.. autosummary::
   :nosignatures:

   init
   load
   save


hfai.cuda
------------
.. py:currentmodule:: hfai.cuda
.. autosummary::
   :nosignatures:

   pin_memory
   unpin_memory


hfai.utils
-------------
.. py:currentmodule:: hfai.utils
.. autosummary::
   :nosignatures:

   which_numa
   num_gpus
   profile_memory
   bind_numa
   get_current_numa
   timeout


hfai.utils.{mm, detr2}
-----------------------------
.. py:currentmodule:: hfai.utils
.. autosummary::
   :nosignatures:

   mm.FFRecordClient
   detr2.register_ffrecord_handler


hfai.autotune
-------------
.. py:currentmodule:: hfai.autotune
.. autosummary::
   :nosignatures:

   run
   run_grid_search
   get_args
   report
   show_result
   set_debug_mode

hfai.pl
-------------
.. py:currentmodule:: hfai.pl
.. autosummary::
   :nosignatures:

   ModelCheckpointHF
   HFAIEnvironment

   DDPStrategyBindNuma
   DDPSpawnStrategyBindNuma
   HFReduceStrategyBindNuma
   HFReduceSpawnStrategyBindNuma

   nn_to_hfai


hfai.file
------------
.. py:currentmodule:: hfai.file
.. autosummary::
   :nosignatures:

   read_file
   read_file_as_bytesio


hfai.client
----------------
.. py:currentmodule:: hfai.client
.. autosummary::
   :nosignatures:

   set_watchdog_time
   set_whole_life_state
   get_whole_life_state
   receive_suspend_command
   go_suspend

   EXP_PRIORITY
   set_priority

   create_experiment_v2
   bind_hf_except_hook
   get_experiment
   get_experiments
   self_health_check

.. py:currentmodule:: hfai.client.api
.. autosummary::
   :nosignatures:

   Experiment


hfai.client.remote
--------------------
.. py:currentmodule:: hfai.client.remote
.. autosummary::
   :nosignatures:

   GlobalSession
   SessionConfig