pretrain core 3

README.md

@@ -292,6 +292,31 @@ Epoch 1 | iter 2048 step 8 | loss train: 2.919, val: 3.078 | iter time: 464.37 m
 Epoch 1 | iter 2304 step 9 | loss train: 2.922, val: 3.078 | iter time: 464.24 ms (step) remaining time: 6:51:45
 Epoch 1 | iter 2560 step 10 | loss train: 2.924, val: 3.078 | iter time: 464.71 ms (step) remaining time: 6:48:39
 # ...
+Epoch 1 | iter 53760 step 210 | loss train: 2.904, val: 3.013 | iter time: 468.66 ms (step) remaining time: 0:23:26
+Epoch 1 | iter 54016 step 211 | loss train: 2.903, val: 3.013 | iter time: 468.81 ms (step) remaining time: 0:21:32
+Epoch 1 | iter 54272 step 212 | loss train: 2.951, val: 3.013 | iter time: 463.52 ms (step) remaining time: 0:19:39
+Epoch 1 | iter 54528 step 213 | loss train: 2.941, val: 3.013 | iter time: 466.12 ms (step) remaining time: 0:17:45
+Epoch 1 | iter 54784 step 214 | loss train: 2.950, val: 3.013 | iter time: 468.24 ms (step) remaining time: 0:15:52
+Epoch 1 | iter 55040 step 215 | loss train: 2.943, val: 3.013 | iter time: 466.65 ms (step) remaining time: 0:13:59
+Epoch 1 | iter 55296 step 216 | loss train: 2.903, val: 3.013 | iter time: 464.44 ms (step) remaining time: 0:12:05
+Epoch 1 | iter 55552 step 217 | loss train: 2.954, val: 3.013 | iter time: 465.12 ms (step) remaining time: 0:10:12
+Epoch 1 | iter 55808 step 218 | loss train: 2.907, val: 3.013 | iter time: 464.96 ms (step) remaining time: 0:08:19
+Epoch 1 | iter 56064 step 219 | loss train: 2.909, val: 3.013 | iter time: 467.12 ms (step) remaining time: 0:06:25
+Epoch 1 | iter 56320 step 220 | loss train: 2.908, val: 3.013 | iter time: 466.43 ms (step) remaining time: 0:04:32
+Epoch 1 | iter 56576 step 221 | loss train: 2.894, val: 3.013 | iter time: 469.70 ms (step) remaining time: 0:02:38
+Epoch 1 | iter 56832 step 222 | loss train: 2.809, val: 3.013 | iter time: 463.88 ms (step) remaining time: 0:00:45
+Validating ...
+Final evaluation | val loss: 3.011 | val ppl: 20.306
+Saving checkpoint to '../out/pretrain-core-1/final/lit_model.pth'
+----------------------------------------
+| Performance
+| - Total tokens  : 466,411,520
+| - Training Time : 25263.31 s
+| - Tok/sec       : 371.33 tok/s
+| ----------------------------------------
+| Memory Usage
+| - Memory Used   : 22.33 GB
+----------------------------------------
 ```
 
 ```bash

config-3.json

@@ -0,0 +1,29 @@
+{
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 0,
+  "eos_token_id": 1,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 512,
+  "initializer_range": 0.02,
+  "intermediate_size": 2048,
+  "max_position_embeddings": 131072,
+  "mlp_bias": false,
+  "model_type": "llama",
+  "num_attention_heads": 8,
+  "num_hidden_layers": 32,
+  "num_key_value_heads": 8,
+  "pretraining_tp": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "rope_theta": 84000.0,
+  "tie_word_embeddings": true,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.45.0.dev0",
+  "use_cache": true,
+  "vocab_size": 131072
+}

scripts/pretrain_core_model_2.yaml

@@ -25,7 +25,7 @@ model_config:
 
 # Directory in which to save checkpoints and logs. If running in a Lightning Studio Job, look for it in
 # /teamspace/jobs/<job-name>/share. (type: <class 'Path'>, default: out/pretrain)
-out_dir: "../out/pretrain-core-1/"
+out_dir: "../out/pretrain-core-2/"
 
 # The precision to use for pretraining. Possible choices: "bf16-true", "bf16-mixed", "32-true". (type: Optional[str], default: null)
 # precision: bf16-mixed

scripts/pretrain_core_model_3.yaml

@@ -0,0 +1,149 @@
+# The name of the model to pretrain. Choose from names in ``litgpt.config``. Mutually exclusive with
+# ``model_config``. (type: Optional[str], default: null)
+model_name: 'tangled-alpha-0.9-core'
+
+# A ``litgpt.Config`` object to define the model architecture. Mutually exclusive with
+# ``model_config``. (type: Optional[Config], default: null)
+model_config:
+  name: 'tangled-alpha-0.9-core'
+  block_size: 131072
+  vocab_size: 131072
+  padded_vocab_size: 131072
+  n_layer: 32
+  n_head: 8
+  n_embd: 512
+  n_query_groups: 8
+  rotary_percentage: 1.0
+  parallel_residual: False
+  bias: False
+  norm_class_name: "RMSNorm"
+  mlp_class_name: "LLaMAMLP"
+  intermediate_size: 2048 # n_embd * 4
+  norm_eps: 1e-5
+  rope_base: 84000 # https://arxiv.org/pdf/2405.14591
+  head_size: 128 # n_embd / n_head
+
+# Directory in which to save checkpoints and logs. If running in a Lightning Studio Job, look for it in
+# /teamspace/jobs/<job-name>/share. (type: <class 'Path'>, default: out/pretrain)
+out_dir: "../out/pretrain-core-3/"
+
+# The precision to use for pretraining. Possible choices: "bf16-true", "bf16-mixed", "32-true". (type: Optional[str], default: null)
+# precision: bf16-mixed
+precision: bf16-true
+
+# Optional path to a checkpoint directory to initialize the model from.
+# Useful for continued pretraining. Mutually exclusive with ``resume``. (type: Optional[Path], default: null)
+initial_checkpoint_dir: "../out/pretrain-core-2/checkpoint"
+
+# Path to a checkpoint directory to resume from in case training was interrupted, or ``True`` to resume
+# from the latest checkpoint in ``out_dir``. An error will be raised if no checkpoint is found. Passing
+# ``'auto'`` will resume from the latest checkpoint but not error if no checkpoint exists.
+# (type: Union[bool, Literal["auto"], Path], default: False)
+resume:
+
+# Data-related arguments. If not provided, the default is ``litgpt.data.TinyLlama``.
+data:
+  class_path: LitData
+
+  init_args:
+    data_path: "../core-data-3-4097-8193-8193-2000/"
+    num_workers: 32
+
+# Training-related arguments. See ``litgpt.args.TrainArgs`` for details
+train:
+  # Number of optimizer steps between saving checkpoints (type: Optional[int], default: 1000)
+  save_interval: 10
+
+  # Number of iterations between logging calls (type: int, default: 1)
+  log_interval: 1
+
+  # Number of samples between optimizer steps across data-parallel ranks (type: int, default: 512)
+  global_batch_size: 512
+
+  # Number of samples per data-parallel rank (type: int, default: 4)
+  micro_batch_size: 1
+
+  # Number of iterations with learning rate warmup active (type: int, default: 2000)
+  lr_warmup_steps: 0
+
+  # Number of epochs to train on (type: Optional[int], default: null)
+  epochs:
+
+  # Total number of tokens to train on (type: Optional[int], default: 3000000000000)
+  max_tokens: 464649609
+
+  # Limits the number of optimizer steps to run. (type: Optional[int], default: null)
+  max_steps:
+
+  # Limits the length of samples. Off by default (type: Optional[int], default: null)
+  max_seq_length: 8192
+
+  # Whether to tie the embedding weights with the language modeling head weights. (type: Optional[bool], default: False)
+  tie_embeddings: true
+
+  #   (type: Optional[float], default: 1.0)
+  max_norm: 1.0
+
+  #   (type: float, default: 4e-05)
+  min_lr: 5e-5
+
+# Evaluation-related arguments. See ``litgpt.args.EvalArgs`` for details
+eval:
+  # Number of optimizer steps between evaluation calls (type: int, default: 1000)
+  interval: 10
+
+  # Number of tokens to generate (type: Optional[int], default: null)
+  max_new_tokens:
+
+  # Number of iterations (type: int, default: 100)
+  max_iters: 100
+
+  # Whether to evaluate on the validation set at the beginning of the training
+  initial_validation: true
+
+  # Whether to evaluate on the validation set at the end the training
+  final_validation: true
+
+# Optimizer-related arguments
+
+# optimizer:
+#   class_path: torch.optim.AdamW
+#   # class_path: torchao.prototype.low_bit_optim.AdamW8bit
+#   # class_path: torchao.prototype.low_bit_optim.AdamW4bit
+#   # class_path: bitsandbytes.optim.AdamW8bit
+#   # class_path: bitsandbytes.optim.PagedAdamW8bit
+#   init_args:
+#     # (type: float, default: 0.001)
+#     lr: 3e-4
+#     # (type: float, default: 0.01)
+#     weight_decay: 0.01
+#     # (type: tuple, default: (0.9,0.999))
+#     betas:
+#       - 0.9
+#       - 0.999
+
+optimizer:
+  class_path: sophia_opt.SophiaG
+  init_args:
+    lr: 1e-4
+    betas:
+      - 0.9
+      - 0.95
+    rho: 0.05
+    weight_decay: 0.1
+
+# How many devices/GPUs to use. Uses all GPUs by default. (type: Union[int, str], default: auto)
+devices: auto
+
+# How many nodes to use. (type: int, default: 1)
+num_nodes: 1
+
+# Optional path to the tokenizer dir that was used for preprocessing the dataset. Only some data
+# module require this. (type: Optional[Path], default: null)
+tokenizer_dir: "../tokenizer"
+
+# The name of the logger to send metrics to. (type: Literal['wandb', 'tensorboard', 'csv'], default: tensorboard)
+logger_name: "wandb"
+
+# The random seed to use for reproducibility. (type: int, default: 42)
+seed: 23