| { | |
| "dataset.debug": false, | |
| "dataset.eval_path": "./output/datasets/self_ds_Llama-2-7b-chat-hf", | |
| "dataset.git_diff": "", | |
| "dataset.git_sha1": "unknown", | |
| "dataset.manual_sample_ids": [], | |
| "dataset.max_read_items": null, | |
| "dataset.output_dir": "output", | |
| "dataset.path": "output/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, | |
| "dataset_generation.sharegpt_path": "Aeala/ShareGPT_Vicuna_unfiltered", | |
| "device_names": [ | |
| "NVIDIA A10G" | |
| ], | |
| "inference.alpha_stats": false, | |
| "inference.debug": false, | |
| "inference.draft_growing": false, | |
| "inference.draft_tree_shape": "mc_sim_7b_64", | |
| "inference.dynamic_draft": true, | |
| "inference.dynamic_draft_all_top_k": 59, | |
| "inference.dynamic_draft_max_depth": 5, | |
| "inference.dynamic_draft_top_k": 10, | |
| "inference.git_diff": "", | |
| "inference.git_sha1": "unknown", | |
| "inference.interactive": false, | |
| "inference.max_draft_growing_depth": 100, | |
| "inference.max_new_tokens": 512, | |
| "inference.mode": "speculative", | |
| "inference.output_dir": "output", | |
| "inference.run_name": "temp_run", | |
| "inference.seed": 42, | |
| "inference.timer": false, | |
| "modeling.add_noise": true, | |
| "modeling.attention_wind": "5", | |
| "modeling.ckpt_path": null, | |
| "modeling.debug": false, | |
| "modeling.decoder_key_remap": {}, | |
| "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.mask_token": "\u2581blank", | |
| "modeling.mask_token_rate": 0, | |
| "modeling.model_path": "beagle/models/llama2/Llama-2-7b-chat-hf/", | |
| "modeling.norm_path": "model.norm", | |
| "modeling.only_first_state_distill": false, | |
| "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_dyt": false, | |
| "modeling.use_fc_eagle": false, | |
| "modeling.use_lower_layers": 0, | |
| "modeling.use_moe": true, | |
| "modeling.use_state_distill": true, | |
| "training.adam_beta1": 0.9, | |
| "training.adam_beta2": 0.95, | |
| "training.average_tokens_across_devices": false, | |
| "training.bf16": true, | |
| "training.ddp_find_unused_parameters": false, | |
| "training.debug": false, | |
| "training.disable_sampled_print": false, | |
| "training.eval_max_tti_wind": null, | |
| "training.eval_steps": 100, | |
| "training.eval_strategy": "steps", | |
| "training.fast_dry_run": false, | |
| "training.filter_out_shorts": false, | |
| "training.force_model_parallel": false, | |
| "training.git_diff": "diff --git a/beagle/models/llama2/modeling_speculative_llama.py b/beagle/models/llama2/modeling_speculative_llama.py\nindex febae2d..b0c256d 100644\n--- a/beagle/models/llama2/modeling_speculative_llama.py\n+++ b/beagle/models/llama2/modeling_speculative_llama.py\n@@ -79,7 +79,7 @@ class LlamaBeagleDecoderLayer(LlamaDecoderLayer):\n \"beagle/models/deepseek-moe\", trust_remote_code=True)\n moe_config.hidden_size = config.hidden_size\n moe_config.moe_intermediate_size = config.intermediate_size\n- moe_config.n_routed_experts = 3\n+ moe_config.n_routed_experts = 5\n moe_config.n_shared_experts = None\n moe_config.num_experts_per_tok = 1\n self.mlp = DeepseekMoE(moe_config, init_exp)", | |
| "training.git_sha1": "928ca280f55b12f05d0109a87e5b54703590e21e", | |
| "training.gradient_accumulation_steps": 8, | |
| "training.learning_rate": 3e-05, | |
| "training.logging_first_step": true, | |
| "training.logging_steps": 20, | |
| "training.loss_pivot_only": false, | |
| "training.loss_reweighting": false, | |
| "training.loss_reweighting_use_surrogate": false, | |
| "training.lr_scheduler_type": "constant_with_warmup", | |
| "training.max_grad_norm": 0.5, | |
| "training.max_length": 2048, | |
| "training.max_steps": -1, | |
| "training.max_tti_wind": 1, | |
| "training.model_init_ckpt": null, | |
| "training.num_train_epochs": 10, | |
| "training.number_sampled_print": "index", | |
| "training.optim": "adamw_torch_fused", | |
| "training.output_dir": "output/fresh-darkness-542", | |
| "training.overwrite_output_dir": true, | |
| "training.per_device_eval_batch_size": 1, | |
| "training.per_device_train_batch_size": 1, | |
| "training.project": "beagle", | |
| "training.ranking_distill_topk": null, | |
| "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.slow_dry_run": false, | |
| "training.tf32": false, | |
| "training.topk_w": 1.0, | |
| "training.use_eagle_pipeline": false, | |
| "training.warmup_steps": 0, | |
| "training.world_size": 1 | |
| } |