WIP: Mixture-of-Experts

config_files/training/config_lorem_ipsum_long_moe_ep_fsdp2.yaml

@@ -0,0 +1,384 @@
+settings:
+  experiment_id: ${modalities_env:experiment_id}
+  config_file_path: ${modalities_env:config_file_path}
+  referencing_keys:
+    sample_key: input_ids
+    target_key: target_ids
+    prediction_key: logits
+  cuda_env:
+    local_rank: ${cuda_env:LOCAL_RANK}
+    global_rank: ${cuda_env:RANK}
+    world_size: ${cuda_env:WORLD_SIZE}
+  paths:
+    checkpoint_saving_path: data/checkpoints
+    train_dataset_path: ./data/lorem_ipsum_long.pbin
+    test_dataset_path: ./data/lorem_ipsum.pbin
+    experiments_root_path: ${modalities_env:experiments_root_path}
+  intervals:
+    training_log_interval_in_steps: 1
+    checkpointing_interval_in_steps: 32
+    evaluation_interval_in_steps: 32
+  consistency_enforcement:
+    enforce_tokens_per_step_consistency: false
+    enforce_last_step_logged: false
+    enforce_last_step_evaluated: false
+    enforce_last_step_checkpointed: false
+  step_profile:
+    gradient_accumulation_steps: 1
+    local_train_micro_batch_size: 1
+    sequence_length: 256
+    dp_degree:
+      instance_key: dp_degree
+      pass_type: BY_REFERENCE
+  training_target:
+    num_target_tokens:
+      component_key: number_conversion
+      variant_key: num_tokens_from_packed_mem_map_dataset_continuous
+      config:
+        dataset_path: ${settings.paths.train_dataset_path}
+        sequence_length: ${settings.step_profile.sequence_length}
+        dp_degree:
+          instance_key: dp_degree
+          pass_type: BY_REFERENCE
+        local_micro_batch_size: ${settings.step_profile.local_train_micro_batch_size}
+        gradient_accumulation_steps: ${settings.step_profile.gradient_accumulation_steps}
+    num_target_steps:
+      component_key: number_conversion
+      variant_key: num_steps_from_num_tokens
+      config:
+        dp_degree:
+          instance_key: dp_degree
+          pass_type: BY_REFERENCE
+        local_micro_batch_size: ${settings.step_profile.local_train_micro_batch_size}
+        global_num_tokens: ${settings.training_target.num_target_tokens}
+        sequence_length: ${settings.step_profile.sequence_length}
+        gradient_accumulation_steps: ${settings.step_profile.gradient_accumulation_steps}
+  training_progress:
+    global_num_seen_tokens: 0
+    num_seen_steps: 0
+    num_seen_samples: 0
+    last_step: -1
+
+collate_fn:
+  component_key: collate_fn
+  variant_key: gpt_2_llm_collator
+  config:
+    sample_key: ${settings.referencing_keys.sample_key}
+    target_key: ${settings.referencing_keys.target_key}
+
+train_dataset:
+  component_key: dataset
+  variant_key: packed_mem_map_dataset_continuous
+  config:
+    raw_data_path: ${settings.paths.train_dataset_path}
+    sequence_length: ${settings.step_profile.sequence_length}
+    sample_key: ${settings.referencing_keys.sample_key}
+
+train_dataloader:
+  component_key: data_loader
+  variant_key: default
+  config:
+    num_workers: 2
+    pin_memory: true
+    dataloader_tag: train
+    dataset:
+      instance_key: train_dataset
+      pass_type: BY_REFERENCE
+    batch_sampler:
+      component_key: batch_sampler
+      variant_key: default
+      config:
+        batch_size: ${settings.step_profile.local_train_micro_batch_size}
+        drop_last: true
+        sampler:
+          component_key: sampler
+          variant_key: resumable_distributed_sampler
+          config:
+            dataset:
+              instance_key: train_dataset
+              pass_type: BY_REFERENCE
+            rank: ${settings.cuda_env.global_rank}
+            num_replicas: ${settings.cuda_env.world_size}
+            shuffle: true
+            seed: 42
+            drop_last: true
+            skip_num_global_samples: ${settings.training_progress.num_seen_samples}
+    collate_fn:
+      instance_key: collate_fn
+      pass_type: BY_REFERENCE
+
+test_dataset:
+  component_key: dataset
+  variant_key: packed_mem_map_dataset_continuous
+  config:
+    raw_data_path: ${settings.paths.test_dataset_path}
+    sequence_length: ${settings.step_profile.sequence_length}
+    sample_key: ${settings.referencing_keys.sample_key}
+
+test_dataloader:
+  component_key: data_loader
+  variant_key: default
+  config:
+    num_workers: 2
+    pin_memory: true
+    dataloader_tag: test
+    dataset:
+      instance_key: test_dataset
+      pass_type: BY_REFERENCE
+    batch_sampler:
+      component_key: batch_sampler
+      variant_key: default
+      config:
+        batch_size: ${settings.step_profile.local_train_micro_batch_size}
+        drop_last: true
+        sampler:
+          component_key: sampler
+          variant_key: distributed_sampler
+          config:
+            rank: ${settings.cuda_env.global_rank}
+            num_replicas: ${settings.cuda_env.world_size}
+            shuffle: false
+            drop_last: true
+            dataset:
+              instance_key: test_dataset
+              pass_type: BY_REFERENCE
+    collate_fn:
+      instance_key: collate_fn
+      pass_type: BY_REFERENCE
+
+eval_dataloaders:
+  - instance_key: test_dataloader
+    pass_type: BY_REFERENCE
+
+checkpoint_saving:
+  component_key: checkpoint_saving
+  variant_key: default
+  config:
+    checkpoint_saving_strategy:
+      component_key: checkpoint_saving_strategy
+      variant_key: save_k_most_recent_checkpoints_strategy
+      config:
+        k: -1
+    checkpoint_saving_execution:
+      component_key: checkpoint_saving_execution
+      variant_key: dcp
+      config:
+        checkpoint_path: ${settings.paths.checkpoint_saving_path}
+        global_rank: ${settings.cuda_env.global_rank}
+        experiment_id: ${settings.experiment_id}
+
+loss_fn:
+  component_key: loss
+  variant_key: moe_cross_entropy
+  config:
+    target_key: ${settings.referencing_keys.target_key}
+    prediction_key: ${settings.referencing_keys.prediction_key}
+    model:
+      instance_key: model_raw
+      pass_type: BY_REFERENCE
+
+device_mesh:
+  component_key: device_mesh
+  variant_key: default
+  config:
+    device_type: cuda
+    data_parallel_replicate_degree: 1
+    data_parallel_shard_degree: -1
+    expert_parallel_degree: 4
+    world_size: ${settings.cuda_env.world_size}
+
+dp_degree:
+  component_key: number_conversion
+  variant_key: parallel_degree
+  config:
+    device_mesh:
+      instance_key: device_mesh
+      pass_type: BY_REFERENCE
+    parallelism_methods: [dp_shard, dp_replicate]
+
+app_state:
+  component_key: app_state
+  variant_key: raw
+  config:
+    model:
+      instance_key: initialized_model
+      pass_type: BY_REFERENCE
+    optimizer:
+      instance_key: optimizer
+      pass_type: BY_REFERENCE
+    lr_scheduler:
+      instance_key: lr_scheduler
+      pass_type: BY_REFERENCE
+
+initialized_model:
+  component_key: model
+  variant_key: model_initialized
+  config:
+    model:
+      instance_key: fsdp_model
+      pass_type: BY_REFERENCE
+    model_initializer:
+      component_key: model_initialization
+      variant_key: composed
+      config:
+        model_type: gpt2
+        weight_init_type: scaled
+        mean: 0.0
+        std: 0.02
+        num_layers: ${model_raw.config.num_layers}
+        multi_device_generator_policy: error
+
+ep_model:
+  component_key: model
+  variant_key: ep_wrapped
+  config:
+    model:
+      instance_key: model_raw
+      pass_type: BY_REFERENCE
+    device_mesh:
+      instance_key: device_mesh
+      pass_type: BY_REFERENCE
+    mixed_precision_settings:
+      param_dtype: BF_16
+      reduce_dtype: BF_16
+    block_names: [TransformerBlock]
+
+ac_model:
+  component_key: model
+  variant_key: activation_checkpointed
+  config:
+    model:
+      instance_key: ep_model
+      pass_type: BY_REFERENCE
+    ac_variant: full_activation_checkpointing
+    layers_fqn: layers
+    ac_fun_params:
+      ac_freq: 1
+
+fsdp_model:
+  component_key: model
+  variant_key: fsdp2_wrapped
+  config:
+    model:
+      instance_key: ac_model
+      pass_type: BY_REFERENCE
+    device_mesh:
+      instance_key: device_mesh
+      pass_type: BY_REFERENCE
+    mixed_precision_settings:
+      param_dtype: BF_16
+      reduce_dtype: BF_16
+    reshard_after_forward: true
+    block_names: [TransformerBlock]
+
+model_raw:
+  component_key: model
+  variant_key: moe
+  config:
+    sample_key: ${settings.referencing_keys.sample_key}
+    prediction_key: ${loss_fn.config.prediction_key}
+    vocab_size: 50304
+    max_seq_len: ${settings.step_profile.sequence_length}
+    d_model: 128
+    n_heads: 8
+    n_kv_heads: 4
+    num_layers: 2
+    d_ff: 128
+    attn_dropout: 0.0
+    ffn_dropout: 0.0
+    tie_embeddings: false
+    norm_eps: 1e-6
+    rope_base: 1000000.0
+    moe_num_experts: 8
+    moe_top_k: 2
+    moe_d_ff: 128
+    moe_capacity_factor: 1.25
+    moe_min_capacity: 4
+    moe_overflow_policy: residual
+    moe_router_noise_std: 0.0
+    moe_router_temperature: 1.0
+    moe_router_dropout: 0.0
+    moe_aux_loss_coef: 0.001
+    moe_z_loss_coef: 0.0
+
+lr_scheduler:
+  component_key: scheduler
+  variant_key: onecycle_lr
+  config:
+    optimizer:
+      instance_key: optimizer
+      pass_type: BY_REFERENCE
+    max_lr: 6e-4
+    div_factor: 10
+    final_div_factor: 1
+    total_steps: ${settings.training_target.num_target_steps}
+    pct_start: 0.01
+    anneal_strategy: cos
+    last_epoch: ${settings.training_progress.last_step}
+
+optimizer:
+  component_key: optimizer
+  variant_key: ep_adam_w
+  config:
+    lr: 0.0001
+    betas: [0.9, 0.95]
+    eps: 1e-8
+    weight_decay: 1e-1
+    weight_decay_groups_excluded: [embedding, layernorm]
+    wrapped_model:
+      instance_key: initialized_model
+      pass_type: BY_REFERENCE
+    device_mesh:
+      instance_key: device_mesh
+      pass_type: BY_REFERENCE
+
+gradient_clipper:
+  component_key: gradient_clipper
+  variant_key: ep
+  config:
+    wrapped_model:
+      instance_key: initialized_model
+      pass_type: BY_REFERENCE
+    norm_type: P2_NORM
+    max_norm: 1.0
+    device_mesh:
+      instance_key: device_mesh
+      pass_type: BY_REFERENCE
+
+progress_subscriber:
+  component_key: progress_subscriber
+  variant_key: rich
+  config:
+    global_rank: ${settings.cuda_env.global_rank}
+    num_seen_steps: ${settings.training_progress.num_seen_steps}
+    num_target_steps: ${settings.training_target.num_target_steps}
+    train_dataloader_tag: ${train_dataloader.config.dataloader_tag}
+    eval_dataloaders:
+      instance_key: eval_dataloaders
+      pass_type: BY_REFERENCE
+
+evaluation_subscriber:
+  component_key: results_subscriber
+  variant_key: wandb
+  config:
+    global_rank: ${settings.cuda_env.global_rank}
+    project: modalities_dcp_tests
+    mode: OFFLINE
+    experiment_id: ${settings.experiment_id}
+    directory: wandb_storage
+    config_file_path: ${settings.config_file_path}
+
+mfu_calculator:
+  component_key: mfu_calculator
+  variant_key: gpt2
+  config:
+    n_layer: ${model_raw.config.num_layers}
+    sequence_length: ${settings.step_profile.sequence_length}
+    n_embd: ${model_raw.config.d_model}
+    world_size: ${settings.cuda_env.world_size}
+    wrapped_model:
+      instance_key: initialized_model
+      pass_type: BY_REFERENCE
+    device_mesh:
+      instance_key: device_mesh
+      pass_type: BY_REFERENCE