lightrft.strategy

• lightrft.strategy.config
  • StrategyConfig
    • StrategyConfig.actor_learning_rate
    • StrategyConfig.adam_betas
    • StrategyConfig.adam_offload
    • StrategyConfig.advantage_clip
    • StrategyConfig.advantage_estimator
    • StrategyConfig.advantages_norm
    • StrategyConfig.bf16
    • StrategyConfig.critic_learning_rate
    • StrategyConfig.critic_pretrain
    • StrategyConfig.dynamic_sampling
    • StrategyConfig.enable_engine_sleep
    • StrategyConfig.engine_tp_size
    • StrategyConfig.engine_type
    • StrategyConfig.extra_args
    • StrategyConfig.from_args()
    • StrategyConfig.fsdp
    • StrategyConfig.fsdp_cpu_offload
    • StrategyConfig.fused_linear_logprob
    • StrategyConfig.grad_accum_dtype
    • StrategyConfig.kl_estimator
    • StrategyConfig.l2
    • StrategyConfig.local_rank
    • StrategyConfig.lr_warmup_ratio
    • StrategyConfig.max_norm
    • StrategyConfig.micro_rollout_batch_size
    • StrategyConfig.micro_train_batch_size
    • StrategyConfig.mixed_mm_data
    • StrategyConfig.n_samples_per_prompt
    • StrategyConfig.opd_kl_coef
    • StrategyConfig.overlap_comm
    • StrategyConfig.overlong_buffer
    • StrategyConfig.overlong_buffer_len
    • StrategyConfig.overlong_buffer_penalty_factor
    • StrategyConfig.plot_every
    • StrategyConfig.pretrain_data
    • StrategyConfig.print_config_summary()
    • StrategyConfig.remote_rm_url
    • StrategyConfig.reward_clip
    • StrategyConfig.reward_running_norm
    • StrategyConfig.reward_running_norm_minus_mean
    • StrategyConfig.seed
    • StrategyConfig.sp_size
    • StrategyConfig.teacher_model_url
    • StrategyConfig.train_batch_size
    • StrategyConfig.use_kl_loss
    • StrategyConfig.use_mp_opt
    • StrategyConfig.use_task_reward
    • StrategyConfig.use_tensorboard
    • StrategyConfig.zero_stage
    • StrategyConfig.zpg
• lightrft.strategy.deepspeed
  • lightrft.strategy.deepspeed.deepspeed
    • ModelOptimPair
    • ModelOrModelOptimPair
    • DeepspeedStrategy
  • lightrft.strategy.deepspeed.deepspeed_utils
    • get_train_ds_config
    • get_eval_ds_config
    • offload_deepspeed_states
    • reload_deepspeed_states
• lightrft.strategy.fake_strategy
  • FakeStrategy
    • FakeStrategy.all_gather()
    • FakeStrategy.all_reduce()
    • FakeStrategy.backward()
    • FakeStrategy.create_optimizer()
    • FakeStrategy.engine_generate_local()
    • FakeStrategy.gather_and_generate()
    • FakeStrategy.get_rank()
    • FakeStrategy.init_model_context()
    • FakeStrategy.is_rank_0()
    • FakeStrategy.load_ckpt()
    • FakeStrategy.maybe_load_optimizer()
    • FakeStrategy.maybe_offload_optimizer()
    • FakeStrategy.maybe_sleep_inference_engine()
    • FakeStrategy.optimizer_step()
    • FakeStrategy.prepare()
    • FakeStrategy.save_ckpt()
    • FakeStrategy.setup_distributed()
    • FakeStrategy.setup_inference_engine()
    • FakeStrategy.update_engine_weights()
    • FakeStrategy.wakeup_inference_engine()
  • get_fake_strategy()
• lightrft.strategy.fsdp
  • lightrft.strategy.fsdp.fsdp_optimizer
    • DTENSOR_SUPPORTED
    • FSDPadaptOptimizer
    • offload_fsdp_optimizer
    • load_fsdp_optimizer
  • lightrft.strategy.fsdp.fsdp_utils
    • is_meta_initialized
    • multi_tensor_l2norm_torch
    • calc_l2_norm
    • calc_lp
    • get_norm
    • reduce_grads
    • get_tensor_norm
    • compute_norm
    • BaseGradScaler
    • DynamicGradScaler
    • BaseOptimizer
  • lightrft.strategy.fsdp.fsdpv2
    • FSDPV2Strategy
    • is_mp_optimizer
    • ModelOptimPair
    • ModelOrModelOptimPair
    • manual_transformer_cls_names_to_wrap
• lightrft.strategy.sglang_utils
  • lightrft.strategy.sglang_utils.sglang_engine
    • RLGenerationEngine
  • lightrft.strategy.sglang_utils.sgl_model_saver
    • release_memory_occupation()
    • resume_memory_occupation()
  • get_sglang_engine
    • get_sglang_engine()
• lightrft.strategy.strategy
  • get_strategy
    • get_strategy()
• lightrft.strategy.strategy_base
  • EngineStatus
    • EngineStatus
  • StrategyBase
    • StrategyBase
  • is_actor
    • is_actor()
• lightrft.strategy.utils
  • lightrft.strategy.utils.broadcast_utils
    • BroadcastManager
  • lightrft.strategy.utils.ckpt_utils
    • find_latest_checkpoint_dir
  • lightrft.strategy.utils.data_utils
    • DistributedSampler
  • lightrft.strategy.utils.distributed_util
    • all_gather_all_prompt_token_ids()
    • create_sub_group()
    • gather_inputs_object_for_inference()
    • init_process_group()
  • lightrft.strategy.utils.optimizer_utils
    • _DEFAULT_NO_DECAY_NAME_LIST
    • get_optimizer_grouped_parameters
  • lightrft.strategy.utils.parallel_utils
    • set_sequence_parallel_group
    • get_sequence_parallel_group
    • get_sequence_parallel_world_size
    • get_sequence_parallel_rank
    • SPDataProcessor
    • sp_slice_and_pad_input
    • gather_forward_split_backward_and_unpad
  • lightrft.strategy.utils.statistic
    • GenLenAnalyser
    • analyse_output_lengths
    • collect_local_output_lengths
    • gather_all_lengths
    • analyze_output_lengths
• lightrft.strategy.vllm_utils
  • lightrft.strategy.vllm_utils.vllm_worker_wrap_no_ray
    • WorkerWrap
  • get_vllm_engine
    • get_vllm_engine()