# torchtitan Config.toml
# NOTE: this toml config is a preset for 64 A100 GPUs.

[job]
dump_folder = "./llama3_1b_output"
description = "Llama 3 1B training"
enable_wandb = true

[profiling]
enable_profiling = true
save_traces_folder = "profile_trace"
profile_freq = 100

[metrics]
log_freq = 10
enable_tensorboard = true
save_tb_folder = "tb"

[model]
name = "llama3"
flavor = "1B"
tokenizer_path = "./assets/tokenizer/Llama-3.1-8B"
# converters = ["float8"]

[optimizer]
name = "AdamW"
lr = 3e-4
eps = 1e-8

[lr_scheduler]
warmup_steps = 2000  # lr scheduler warm up

[training]
local_batch_size = 4
global_batch_size = 128
seq_len = 8192
max_norm = 1.0  # grad norm clipping
steps = 20000
compile = false
dataset = "fineweb_edu_100bt"

[parallelism]
data_parallel_replicate_degree = 1
data_parallel_shard_degree = -1
tensor_parallel_degree = 1
pipeline_parallel_degree = 1
context_parallel_degree = 1

[checkpoint]
enable_checkpoint = true
folder = "./llama3_1b_output/checkpoint"
interval = 500
last_save_model_only = true
export_dtype = "float32"
async_mode = "disabled" # ["disabled", "async", "async_with_pinned_mem"]

[activation_checkpoint]
mode = "selective"  # ["none", "selective", "full"]
selective_ac_option = "op"  # "int" = ac every positive int layer or 'op', ac based on ops policy

[float8]
enable_fsdp_float8_all_gather = false
precompute_float8_dynamic_scale_for_fsdp = false
filter_fqns = ["output"]