Alit/mamba tmp

examples/nlp/language_modeling/conf/megatron_mamba_config.yaml

@@ -0,0 +1,191 @@
+name: megatron_mamba
+restore_from_path: null # used when starting from a .nemo file
+
+trainer:
+  devices: 1
+  num_nodes: 1
+  accelerator: gpu
+  precision: bf16
+  logger: False # logger provided by exp_manager
+  enable_checkpointing: False
+  use_distributed_sampler: False
+  max_epochs: -1 # PTL default. In practice we don't usually train for more than 1 epoch.
+  max_steps: 100000 # consumed_samples = global_step * micro_batch_size * data_parallel_size * accumulate_grad_batches
+  log_every_n_steps: 10
+  val_check_interval: 100
+  limit_val_batches: 50
+  limit_test_batches: 500
+  accumulate_grad_batches: 1
+  gradient_clip_val: 1.0
+  benchmark: False
+
+exp_manager:
+  explicit_log_dir: null
+  exp_dir: null
+  name: megatron_mamba
+  create_wandb_logger: False
+  wandb_logger_kwargs:
+    project: null
+    name: null
+  resume_if_exists: True
+  resume_ignore_no_checkpoint: True
+  create_checkpoint_callback: True
+  checkpoint_callback_params:
+    monitor: val_loss
+    save_top_k: 10
+    mode: min
+    always_save_nemo: False # saves nemo file during validation, not implemented for model parallel
+    filename: 'megatron_mamba--{val_loss:.2f}-{step}-{consumed_samples}'
+    model_parallel_size: ${multiply:${model.tensor_model_parallel_size}, ${model.pipeline_model_parallel_size}}
+
+
+model:
+  restore_from_path: null
+  # model parallelism 
+  mcore_gpt: True
+  micro_batch_size: 1
+  global_batch_size: 8
+  tensor_model_parallel_size: 1
+  pipeline_model_parallel_size: 1
+  virtual_pipeline_model_parallel_size: null
+  expert_model_parallel_size: 1 # expert model parallelism
+  hybrid_override_pattern: null
+  vocab_size: 256000
+  # model architecture
+  encoder_seq_length: 4096
+  max_position_embeddings: ${.encoder_seq_length}
+  position_embedding_type: 'none' # Position embedding type. Options ['learned_absolute', 'rope', 'alibi', 'kerple' , 'xpos', 'sandwich'] xpos and sandwich are experimental.
+  num_layers: 56
+  gated_linear_unit: False
+  add_bias_linear: False
+  num_query_groups: 8
+  mamba_ssm_ngroups: 8
+  attention_dropout: 0.0
+  hidden_dropout: 0.0
+  hidden_size: 4096
+  ffn_hidden_size: 14336 # Transformer FFN hidden size. Usually 4 * hidden_size.
+  num_attention_heads: 32
+  transformer_block_type: pre_ln
+  init_method_std: 0.02 # Standard deviation of the zero mean normal distribution used for weight initialization.')
+  kv_channels: null # Projection weights dimension in multi-head attention. Set to hidden_size // num_attention_heads if null
+  apply_query_key_layer_scaling: True # scale Q * K^T by 1 / layer-number.
+  normalization: RMSNorm
+  layernorm_epsilon: 1e-5
+  num_moe_experts: 16
+  moe_router_topk: 2
+  moe_aux_loss_coeff: 0.001
+  make_vocab_size_divisible_by: 128 # Pad the vocab size to be divisible by this value for computation efficiency.
+  pre_process: True # add embedding
+  post_process: True # add pooler
+  megatron_legacy: False
+  persist_layer_norm: True
+
+  tokenizer:
+    library: 'huggingface'
+    type: 'EleutherAI/gpt-neox-20b' 
+    model: null 
+    vocab_file: null
+    merge_file: null 
+    sentencepiece_legacy: False
+    use_fast: True
+
+  # Distributed checkpoint setup
+  dist_ckpt_format: 'zarr' # Set to 'torch_dist' to use PyTorch distributed checkpoint format.
+  dist_ckpt_load_on_device: True # whether to load checkpoint weights directly on GPU or to CPU
+  dist_ckpt_parallel_save: False # if true, each worker will write its own part of the dist checkpoint
+
+  # precision
+  native_amp_init_scale: 4294967296 # 2 ** 32
+  native_amp_growth_interval: 1000
+  fp32_residual_connection: False # Move residual connections to fp32
+  fp16_lm_cross_entropy: False # Move the cross entropy unreduced loss calculation for lm head to fp16
+
+  # Megatron O2-style half-precision
+  megatron_amp_O2: False # Enable O2-level automatic mixed precision using main parameters
+  grad_allreduce_chunk_size_mb: 125
+
+
+  # Fusion
+  grad_div_ar_fusion: True # Fuse grad division into torch.distributed.all_reduce. Only used with O2 and no pipeline parallelism..
+  gradient_accumulation_fusion: False # Fuse weight gradient accumulation to GEMMs. Only used with pipeline parallelism and O2.
+  bias_activation_fusion: False # Use a kernel that fuses the bias addition from weight matrices with the subsequent activation function.
+  bias_dropout_add_fusion: True # Use a kernel that fuses the bias addition, dropout and residual connection addition.
+  masked_softmax_fusion: True # Use a kernel that fuses the attention softmax with it's mask.
+  get_attention_mask_from_fusion: True # When using fused softmax it will create the attention mask so we won't copy it to the pipeline stages.
+  apply_rope_fusion: True # Use a kernel to add rotary positional embeddings. Only used if position_embedding_type=rope
+
+
+  # miscellaneous
+  seed: 1234
+  use_cpu_initialization: False # Init weights on the CPU (slow for large models)
+  onnx_safe: False # Use work-arounds for known problems with Torch ONNX exporter.
+  gradient_as_bucket_view: True # PyTorch DDP argument. Allocate gradients in a contiguous bucket to save memory (less fragmentation and buffer memory)
+
+  ## Activation Checkpointing
+  # NeMo Megatron supports 'selective' activation checkpointing where only the memory intensive part of attention is checkpointed.
+  # These memory intensive activations are also less compute intensive which makes activation checkpointing more efficient for LLMs (20B+).
+  # See Reducing Activation Recomputation in Large Transformer Models: https://arxiv.org/abs/2205.05198 for more details.
+  # 'full' will checkpoint the entire transformer layer.
+  activations_checkpoint_granularity: null # 'selective' or 'full' 
+  activations_checkpoint_recurrent: False # If set to True, the checkpointing is only done for rglru and conv1d and not for attention and mlp layers
+  activations_checkpoint_method: null # 'uniform', 'block'
+  # 'uniform' divides the total number of transformer layers and checkpoints the input activation
+  # of each chunk at the specified granularity. When used with 'selective', 'uniform' checkpoints all attention blocks in the model.
+  # 'block' checkpoints the specified number of layers per pipeline stage at the specified granularity
+  activations_checkpoint_num_layers: null
+  # when using 'uniform' this creates groups of transformer layers to checkpoint. Usually set to 1. Increase to save more memory.
+  # when using 'block' this this will checkpoint the first activations_checkpoint_num_layers per pipeline stage.
+  num_micro_batches_with_partial_activation_checkpoints: null
+  # This feature is valid only when used with pipeline-model-parallelism.
+  # When an integer value is provided, it sets the number of micro-batches where only a partial number of Transformer layers get checkpointed
+  # and recomputed within a window of micro-batches. The rest of micro-batches in the window checkpoint all Transformer layers. The size of window is
+  # set by the maximum outstanding micro-batch backpropagations, which varies at different pipeline stages. The number of partial layers to checkpoint
+  # per micro-batch is set by 'activations_checkpoint_num_layers' with 'activations_checkpoint_method' of 'block'.
+  # This feature enables using activation checkpoint at a fraction of micro-batches up to the point of full GPU memory usage.
+  activations_checkpoint_layers_per_pipeline: null
+  # This feature is valid only when used with pipeline-model-parallelism.
+  # When an integer value (rounded down when float is given) is provided, it sets the number of Transformer layers to skip checkpointing at later
+  # pipeline stages. For example, 'activations_checkpoint_layers_per_pipeline' of 3 makes pipeline stage 1 to checkpoint 3 layers less than
+  # stage 0 and stage 2 to checkpoint 6 layers less stage 0, and so on. This is possible because later pipeline stage
+  # uses less GPU memory with fewer outstanding micro-batch backpropagations. Used with 'num_micro_batches_with_partial_activation_checkpoints',
+  # this feature removes most of activation checkpoints at the last pipeline stage, which is the critical execution path.
+  sequence_parallel: False
+
+  data:
+    # Path to data must be specified by the user.
+    # can override from the CLI: "model.data.data_prefix=[.5,/raid/data/pile/my-gpt3_00_text_document,.5,/raid/data/pile/my-gpt3_01_text_document]",
+    # Or see example below: 
+    # data_prefix: 
+    #   - .5
+    #   - /raid/data/pile/my-gpt3_00_text_document
+    #   - .5
+    #   - /raid/data/pile/my-gpt3_01_text_document
+    data_prefix: [1.0, /path/to/data]
+    index_mapping_dir: null # path to save index mapping .npy files, by default will save in the same location as data_prefix
+    data_impl: mmap
+    splits_string: 900,50,50
+    seq_length: ${model.encoder_seq_length}
+    skip_warmup: True
+    num_workers: 0
+    dataloader_type: single  # cyclic, LDDL
+    reset_position_ids: False # Reset position ids after end-of-document token
+    reset_attention_mask: False # Reset attention mask after end-of-document token
+    eod_mask_loss: False # Mask loss for the end of document tokens
+    masked_lm_prob: 0.15 # Probability of replacing a token with mask.
+    short_seq_prob: 0.1 # Probability of producing a short sequence.
+    ceil_to_power_2: True
+    get_attention_mask_from_fusion: True
+    pad_to_max_length: True
+
+  optim:
+    name: distributed_fused_adam
+    lr: 2e-4
+    weight_decay: 0.01 
+    betas: 
+    - 0.9
+    - 0.98
+    sched:
+      name: CosineAnnealing
+      warmup_steps: 500
+      constant_steps: 50000
+      min_lr: 2e-5