Save model at global step 40

config.yaml

@@ -0,0 +1,177 @@
+data:
+  tokenizer: null
+  train_files: dapo_ds_train_sample.parquet
+  val_files: matheval.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: 768
+  max_response_length: 4096
+  train_batch_size: 128
+  val_batch_size: 640
+  return_raw_input_ids: false
+  return_raw_chat: false
+  shuffle: true
+  filter_overlong_prompts: false
+  filter_overlong_prompts_workers: 1
+  truncation: left
+  image_key: images
+  custom_cls:
+    path: null
+    name: null
+actor_rollout_ref:
+  hybrid_engine: true
+  model:
+    path: meta-llama/Llama-3.2-3B-Instruct
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    use_remove_padding: false
+    use_liger: false
+    save_hf_repo_id: RyanYr/brm-dapo-llama3.2-3B-it_lr1.2e-6_rspL4096-beta0.002
+    tokenizer_chat_template: null
+  actor:
+    brm:
+      norm_factor: value
+      value_constant: null
+    buffer:
+      buffer_type: null
+      offline_dataset_buffer:
+        train_files: null
+        response_key: response
+        response_truncation: right
+        shuffle: true
+        update_size: 128
+    strategy: fsdp
+    ppo_mini_batch_size: 128
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: 1
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    grad_clip: 1.0
+    use_torch_compile: true
+    ppo_epochs: 1
+    shuffle: false
+    ulysses_sequence_parallel_size: 1
+    checkpoint:
+      contents:
+      - model
+      - optimizer
+      - extra
+    optim:
+      lr: 1.2e-06
+      lr_warmup_steps: -1
+      lr_warmup_steps_ratio: 0
+      min_lr_ratio: null
+      warmup_style: constant
+      total_training_steps: 400
+      weight_decay: 0.01
+    fsdp_config:
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      fsdp_size: -1
+    report_entropy: true
+  ref:
+    ref_model_path: meta-llama/Llama-3.2-3B-Instruct
+    strategy: fsdp
+    fsdp_config:
+      param_offload: false
+      wrap_policy:
+        min_num_params: 0
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: 8
+    log_prob_use_dynamic_bsz: false
+    log_prob_max_token_len_per_gpu: 16384
+    ulysses_sequence_parallel_size: 1
+  rollout:
+    name: vllm
+    temperature: 1.0
+    top_k: -1
+    top_p: 1
+    use_fire_sampling: false
+    prompt_length: 768
+    response_length: 4096
+    dtype: bfloat16
+    gpu_memory_utilization: 0.6
+    ignore_eos: false
+    enforce_eager: false
+    free_cache_engine: false
+    load_format: dummy_dtensor
+    tensor_model_parallel_size: 4
+    max_num_batched_tokens: 5864
+    max_model_len: null
+    max_num_seqs: 1024
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: 8
+    log_prob_use_dynamic_bsz: false
+    log_prob_max_token_len_per_gpu: 16384
+    disable_log_stats: true
+    enable_chunked_prefill: true
+    do_sample: true
+    'n': 1
+    engine_kwargs:
+      swap_space: null
+    val_kwargs:
+      top_k: -1
+      top_p: 1.0
+      temperature: 0.6
+      'n': 32
+      do_sample: true
+reward_model:
+  enable: false
+  strategy: fsdp
+  model:
+    input_tokenizer: meta-llama/Llama-3.2-3B-Instruct
+    path: ~/models/FsfairX-LLaMA3-RM-v0.1
+    external_lib: null
+    use_remove_padding: false
+    fsdp_config:
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      fsdp_size: -1
+  micro_batch_size: null
+  micro_batch_size_per_gpu: null
+  max_length: null
+  ulysses_sequence_parallel_size: 1
+  use_dynamic_bsz: false
+  forward_max_token_len_per_gpu: 16384
+  reward_manager: prime
+  reward_kwargs:
+    format_reward: 0.0
+    format_type: null
+custom_reward_function:
+  path: null
+  name: compute_score
+trainer:
+  balance_batch: true
+  total_epochs: 100
+  total_training_steps: 400
+  project_name: value-LLM
+  experiment_name: brm-dapo-llama3.2-3B-it_lr1.2e-6_rspL4096-beta0.002
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: 40
+  resume_mode: auto
+  resume_from_path: null
+  val_before_train: false
+  test_freq: -1
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: ./BRM
+  max_actor_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  hf_token: null
+  resume_from_hf:
+    enable: false
+    hf_repo_id: null
+    hf_token: null
+    revision: main
+algorithm:
+  kl_ctrl:
+    kl_coef: 0.002

data.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:abb75c3ac8d43d5d7e8492e245f16591382108d8cd44cc4c3ceefaced2d979bc
+size 1492

extra_state_world_size_4_rank_0.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0d097a47fb7461da5d27460785c7433b2a97e1215dde8f79f9905108150c14b2
+size 14632

extra_state_world_size_4_rank_1.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d4781f9e12c4faa98139099a21d9101e55271dad50e9dc4a1978c324d946bd4a
+size 14632

extra_state_world_size_4_rank_2.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c6fd74afcb4ec9c9da0c26634975d3ce792a3b5c6de14ff8a94b26a0b498c8ab
+size 14632

extra_state_world_size_4_rank_3.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f4e050958b4f2a0a92d3b95efe6615acd3d54029cc29d368dc5261ebab3027c0
+size 14632

model_world_size_4_rank_0.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f92ead5ead02701ffe64f6a6e827d21c3c63fb644b9a4aaf9ea63ede113d32ee
+size 3606904378

model_world_size_4_rank_1.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:72284fd5d26ac03da4d3d5af26fa94ec439718d91fc60938692b4cd89721f0b1
+size 3606904378

model_world_size_4_rank_2.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2c5f06b9aa96c62657c7c2d1fe729ad5c22784468932d9bee2c5c32bd6b30b69
+size 3606904378

model_world_size_4_rank_3.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b30ea09f06f23f1cb685b39ea0f94fd66f7174884b901fa60fe6869090af3f13
+size 3606904378

optim_world_size_4_rank_0.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24fce421015ca1afbb9a28c4a26b78224cc0e08656545b1848a1409ce90548e0
+size 6425526231

optim_world_size_4_rank_1.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2990ad8b971e390fd46d19517b6b75bbb18a744f50f81cf5affece74211de969
+size 6425526231

optim_world_size_4_rank_2.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1ae74814c5aacce3c8c1b5dda75e301610e4f79907b9f2f6176f05fc3a7a6254
+size 6425526231

optim_world_size_4_rank_3.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:94e744a990ef773a63e6353d2131f99e53078528a19d5a881f210f6eff0dae50
+size 6425526231