Transformer-340M-0428 / config.json
windsornguyen's picture
add: tokenizer/config files
b7c583f verified
{
"model_type": "Transformer",
"_name_or_path": "Transformer-340M-0408",
"architectures": ["TransformerForCausalLM"],
"dim": 1024,
"num_heads": 4,
"num_layers": 11,
"seq_len": 4096,
"vocab_size": 200064,
"inter_dim": 4096,
"mlp_scale": 12,
"bias": false,
"weight_tying": true,
"rope_theta": 10000.0,
"num_epochs": 1,
"global_bsz": 524288,
"bsz": 8,
"warmup_steps": 1907,
"eval_period": 50,
"save_period": 1000,
"max_lr": 4.0e-4,
"min_lr": 4.0e-5,
"max_norm": 1.0,
"fsdp": true,
"ddp": false,
"reshard_after_forward_policy": "default",
"mixed_precision": true,
"torch_dtype": "bfloat16",
"cpu_offload": false,
"sharding_strategy": "full_shard",
"state_dict_type": "full",
"auto_wrap_policy": "partial",
"backward_prefetch": "backward_pre",
"forward_prefetch": false,
"sync_module_states": true,
"use_orig_params": true,
"device_id": null,
"precision": {
"param": "bfloat16",
"reduce": "bfloat16",
"buffer": "bfloat16"
},
"fsdp_modules": [
"AttentionLayer"
],
"num_workers": 0,
"snapshot_every_n_steps": 50,
"use_activation_checkpointing": true,
"torch_compile": true,
"torch_compile_kwargs": {
"mode": "default",
"fullgraph": true
},
"enable_compiled_autograd": false
}