fml / state /step_4000 /policy /cosmos_config
Jiaxincc's picture
Add step_4000 checkpoint
f36b364 verified
{
"custom": {},
"train": {
"train_policy": {
"type": "sft",
"dataset": {
"name": "LNTANOooo/sharegpt52k",
"subset": "",
"revision": "",
"split": [
"train"
],
"test_size": 0.1
},
"mini_batch": 8,
"dataloader_shuffle": true,
"enable_dataset_cache": false,
"dataloader_num_workers": 0,
"dataloader_prefetch_factor": null,
"conversation_column_name": "conversation",
"system_prompt": ""
},
"optm_name": "AdamW",
"optm_lr": 0.0002,
"optm_impl": "fused",
"optm_weight_decay": 0.01,
"optm_betas": [
0.9,
0.999
],
"optm_warmup_steps": 20,
"optm_decay_ratio": null,
"optm_decay_type": null,
"optm_min_lr_factor": 0.0,
"optm_grad_norm_clip": 1.0,
"master_dtype": "float32",
"param_dtype": "bfloat16",
"transfer_dtype": "bfloat16",
"fsdp_reduce_dtype": "float32",
"fsdp_offload": false,
"fsdp_reshard_after_forward": "default",
"train_batch_per_replica": 32,
"enable_validation": false,
"validation_step": 20,
"validation_batch_per_replica": 24,
"fp8": {
"enable_fp8": false,
"fp8_recipe": "dynamic_scaling",
"quant_recipe": "rowwise"
},
"ckpt": {
"enable_checkpoint": true,
"save_freq": 2000,
"save_mode": "async",
"max_keep": 5,
"export_safetensors": true,
"upload_hf": false,
"hf_repo_name": "Comos-Reason1",
"upload_s3": false,
"s3_bucket": null,
"s3_prefix": "outputs"
},
"resume": true,
"epoch": 1000,
"output_dir": "./outputs/qwen2-5-3b-tp1-sft/20251001112805",
"timestamp": "20251001112805",
"epsilon": 1e-06,
"async_tp_enabled": false,
"compile": false,
"sync_weight_interval": 1,
"deterministic": false,
"seed": null,
"max_num_steps": null,
"sequence_packing": false
},
"rollout": {
"parallelism": {
"n_init_replicas": 1,
"tp_size": 2,
"cp_size": 1,
"ep_size": 1,
"dp_shard_size": -1,
"pp_size": 1,
"pp_dynamic_shape": false,
"pp_micro_batch_size": 1,
"dp_replicate_size": 1
},
"enforce_eager": true,
"include_stop_str_in_output": false,
"gpu_memory_utilization": 0.8,
"enable_chunked_prefill": false,
"max_response_length": 2048,
"n_generation": 16,
"batch_size": 1,
"val_batch_size": null,
"quantization": "none",
"seed": null,
"sampling_config": {
"temperature": 1.0,
"top_p": 1.0,
"top_k": -1,
"repetition_penalty": 1.0,
"use_flashinfer": false
},
"vllm_use_flashinfer": false,
"backend": "vllm"
},
"policy": {
"parallelism": {
"n_init_replicas": 1,
"tp_size": 1,
"cp_size": 1,
"ep_size": 1,
"dp_shard_size": 8,
"pp_size": 1,
"pp_dynamic_shape": false,
"pp_micro_batch_size": 1,
"dp_replicate_size": 8
},
"model_name_or_path": "Qwen/Qwen3-4B-Instruct-2507",
"model_revision": null,
"model_max_length": 4096,
"model_gradient_checkpointing": true,
"lora": null,
"trainable_map": null,
"enable_liger_kernel": false
},
"logging": {
"logger": [
"console",
"wandb"
],
"project_name": "cosmos_rl",
"experiment_name": "None",
"report_mfu": false
},
"profiler": {
"enable_profiler": false,
"enable_nsys": false,
"sub_profiler_config": {
"do_profile": false,
"active_steps": 1,
"warmup_steps": 1,
"wait_steps": 1,
"rank_filter": [],
"record_shape": false,
"profile_memory": false,
"with_stack": false,
"with_modules": false
}
},
"validation": {
"dataset": {
"name": "",
"subset": "",
"revision": "",
"split": [
""
],
"test_size": null
},
"temperature": 0.0,
"top_p": null,
"top_k": 1,
"repetition_penalty": 1.0,
"n_generation": 1,
"max_response_length": null,
"reward_function": {
"single_choice": 1.0
}
},
"redis": "12800",
"eth_ips": "10.0.0.171;172.17.0.1;100.64.0.5;10.50.83.240;10.50.154.240;169.254.66.66;172.20.0.10"
}