{
  "dataset.debug": false,
  "dataset.git_diff": "",
  "dataset.git_sha1": "unknown",
  "dataset.manual_sample_ids": [],
  "dataset.max_read_items": null,
  "dataset.output_dir": "output",
  "dataset.path": "/mnt/hg_cache/temp_llama_dataset/datasets/ds_Llama-2-7b-chat-hf",
  "dataset.read_eagle_format": false,
  "dataset.run_name": "temp_run",
  "dataset.seed": 42,
  "dataset_generation.batch_size": 1,
  "dataset_generation.debug": false,
  "dataset_generation.debug_target": null,
  "dataset_generation.ds_prefix": "ds_",
  "dataset_generation.git_diff": "",
  "dataset_generation.git_sha1": "unknown",
  "dataset_generation.max_length": 4096,
  "dataset_generation.output_dir": "output",
  "dataset_generation.run_name": "temp_run",
  "dataset_generation.save_every": 1000,
  "dataset_generation.seed": 42,
  "inference.debug": false,
  "inference.detail_time_stats": false,
  "inference.draft_tree_shape": "mc_sim_7b_64",
  "inference.git_diff": "",
  "inference.git_sha1": "unknown",
  "inference.interactive": false,
  "inference.max_new_tokens": 512,
  "inference.mode": "speculative",
  "inference.output_dir": "output",
  "inference.run_name": "temp_run",
  "inference.seed": 42,
  "modeling.add_noise": true,
  "modeling.attention_offset": "random.randrange(0, 3)",
  "modeling.attention_wind": "5",
  "modeling.ckpt_path": null,
  "modeling.debug": false,
  "modeling.decoder_key_remap": {},
  "modeling.draft_growing": false,
  "modeling.dtype": "torch.float32",
  "modeling.frozen_targets": [],
  "modeling.git_diff": "",
  "modeling.git_sha1": "unknown",
  "modeling.layer_path": "model.layers",
  "modeling.lmhead_path": "lm_head",
  "modeling.load_config_from_model_path": false,
  "modeling.model_path": "beagle/models/llama2/Llama-2-7b-chat-hf/",
  "modeling.norm_path": "model.norm",
  "modeling.output_dir": "output",
  "modeling.reuse_layer": null,
  "modeling.rotary_path": "model.rotary_emb",
  "modeling.run_name": "temp_run",
  "modeling.save_loading": true,
  "modeling.seed": 42,
  "modeling.strictly_follow_eagle_decoder": false,
  "modeling.tokenizer_path": "meta-llama/Llama-2-7b-chat-hf",
  "modeling.use_fc_eagle": false,
  "modeling.use_lower_layers": 0,
  "modeling.use_state_distill": false,
  "training.adam_beta1": 0.9,
  "training.adam_beta2": 0.95,
  "training.bf16": true,
  "training.ddp_find_unused_parameters": false,
  "training.debug": false,
  "training.eval_steps": 100,
  "training.eval_strategy": "steps",
  "training.filter_out_shorts": false,
  "training.git_diff": "diff --git a/README.md b/README.md\nindex 906761d..da3f70e 100644\n--- a/README.md\n+++ b/README.md\n@@ -21,13 +21,14 @@ Training:\n ```sh\n wandb login\n # multi-GPU training:\n-torchrun --standalone --nnodes=1 --nproc-per-node=2 -m beagle.train \\\n+CUDA_VISIBLE_DEVICES=0,1,2,3 \\\n+    torchrun --standalone --nnodes=1 --nproc-per-node=4 -m beagle.train \\\n     --@llama2_7b_chat --@rtx4070tis_bs4_ctx4096 --modeling.use_lower_layers 0 \\\n     --dataset.path /mnt/hg_cache/temp_llama_dataset/datasets/ds_Llama-2-7b-chat-hf \\\n     --modeling.save_loading --training.report_to wandb\n \n # customized training:\n-CUDA_VISIBLE_DEVICES=1,2 \\\n+CUDA_VISIBLE_DEVICES=0,1 \\\n     torchrun --standalone --nnodes=1 --nproc-per-node=2 -m beagle.train \\\n     --@exaone3.5_2.4B_instr --@rtx4070tis_dev_bs8 \\\n     --dataset.path /mnt/truenas_sync/beagle_train_data/datasets/ds_EXAONE-3.5-2.4B-Instruct \\\ndiff --git a/beagle/train.py b/beagle/train.py\nindex 74795e1..c25096e 100644\n--- a/beagle/train.py\n+++ b/beagle/train.py\n@@ -199,8 +199,6 @@ def train(configs, run_name, tokenizer, model,\n     if not configs.overwrite_output_dir:\n         assert not os.path.exists(configs.output_dir), configs.output_dir\n \n-    print('[rank]', rank, '/', world_size)\n-\n     random.seed(configs.seed)\n     torch.manual_seed(configs.seed)\n \n@@ -326,6 +324,9 @@ def main(config_file='beagle/configs.ini', **injects):\n     configs.set_obj('training.git_diff', git_diff)\n     master_print('[git]', git_sha1, '\\n', git_diff)\n \n+    print('[rank]', rank, '/', world_size)\n+    configs.set_obj('training.world_size', world_size)\n+\n     if rank == 0 and configs.training.report_to == 'wandb':\n         import wandb\n         wandb.init(",
  "training.git_sha1": "453d65d9cbee2ac0dedb03c3c49d99b1344a81ed",
  "training.gradient_accumulation_steps": 8,
  "training.learning_rate": 3e-05,
  "training.logging_steps": 5,
  "training.lr_scheduler_type": "constant_with_warmup",
  "training.max_grad_norm": 0.5,
  "training.max_length": 4096,
  "training.max_steps": -1,
  "training.model_init_ckpt": null,
  "training.num_train_epochs": 10,
  "training.optim": "adamw_torch_fused",
  "training.output_dir": "output",
  "training.overwrite_output_dir": true,
  "training.per_device_eval_batch_size": 1,
  "training.per_device_train_batch_size": 2,
  "training.project": "beagle",
  "training.report_to": "wandb",
  "training.resume_from_checkpoint": false,
  "training.resume_wandb_runid": null,
  "training.run_name": "temp_run",
  "training.save_steps": 500,
  "training.save_strategy": "steps",
  "training.save_total_limit": 2,
  "training.save_vram": true,
  "training.seed": 42,
  "training.tf32": false,
  "training.use_eagle_pipeline": false,
  "training.warmup_steps": 2000,
  "training.world_size": 1
}