pymllm.orchestrator.parallel_state
==================================

.. py:module:: pymllm.orchestrator.parallel_state

.. autoapi-nested-parse::

   Minimal parallel state for single-GPU serving.

   pymllm targets single-GPU, high-concurrency inference. This module keeps
   the TP / DP / PP scaffolding so the rest of the codebase can query ranks
   and groups uniformly, but the default (and expected) case is world_size=1.



Attributes
----------

.. autoapisummary::

   pymllm.orchestrator.parallel_state.logger


Functions
---------

.. autoapisummary::

   pymllm.orchestrator.parallel_state.initialize_model_parallel
   pymllm.orchestrator.parallel_state.get_tp_group
   pymllm.orchestrator.parallel_state.get_dp_group
   pymllm.orchestrator.parallel_state.get_pp_group
   pymllm.orchestrator.parallel_state.get_tensor_model_parallel_rank
   pymllm.orchestrator.parallel_state.get_tensor_model_parallel_world_size
   pymllm.orchestrator.parallel_state.get_data_parallel_rank
   pymllm.orchestrator.parallel_state.get_data_parallel_world_size
   pymllm.orchestrator.parallel_state.get_pipeline_model_parallel_rank
   pymllm.orchestrator.parallel_state.get_pipeline_model_parallel_world_size
   pymllm.orchestrator.parallel_state.model_parallel_is_initialized
   pymllm.orchestrator.parallel_state.tensor_model_parallel_all_reduce
   pymllm.orchestrator.parallel_state.tensor_model_parallel_all_gather
   pymllm.orchestrator.parallel_state.data_parallel_all_reduce


Module Contents
---------------

.. py:data:: logger

.. py:function:: initialize_model_parallel(tensor_model_parallel_size = 1, data_parallel_size = 1, pipeline_model_parallel_size = 1, backend = 'nccl')

.. py:function:: get_tp_group()

.. py:function:: get_dp_group()

.. py:function:: get_pp_group()

.. py:function:: get_tensor_model_parallel_rank()

.. py:function:: get_tensor_model_parallel_world_size()

.. py:function:: get_data_parallel_rank()

.. py:function:: get_data_parallel_world_size()

.. py:function:: get_pipeline_model_parallel_rank()

.. py:function:: get_pipeline_model_parallel_world_size()

.. py:function:: model_parallel_is_initialized()

.. py:function:: tensor_model_parallel_all_reduce(tensor)

.. py:function:: tensor_model_parallel_all_gather(tensor, dim = 0)

.. py:function:: data_parallel_all_reduce(tensor)