Save model at global step 80

config.yaml

@@ -0,0 +1,223 @@
+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: null
+  return_raw_input_ids: false
+  return_raw_chat: false
+  shuffle: true
+  filter_overlong_prompts: true
+  filter_overlong_prompts_workers: 1
+  truncation: error
+  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: true
+    use_liger: false
+    save_hf_repo_id: RyanYr/ppo-dapo-llama3.2-3B-it-lr-mbs64_actor
+  actor:
+    strategy: fsdp
+    ppo_mini_batch_size: 64
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: 2
+    use_dynamic_bsz: false
+    ppo_max_token_len_per_gpu: 16384
+    grad_clip: 1.0
+    clip_ratio: 0.2
+    clip_ratio_low: 0.2
+    clip_ratio_high: 0.2
+    clip_ratio_c: 3.0
+    loss_agg_mode: token-mean
+    entropy_coeff: 0.001
+    use_kl_loss: false
+    use_torch_compile: true
+    kl_loss_coef: 0.001
+    kl_loss_type: low_var_kl
+    ppo_epochs: 1
+    shuffle: false
+    ulysses_sequence_parallel_size: 1
+    checkpoint:
+      contents:
+      - model
+      - optimizer
+      - extra
+    optim:
+      lr: 1.0e-06
+      lr_warmup_steps: -1
+      lr_warmup_steps_ratio: 0.0
+      min_lr_ratio: null
+      warmup_style: constant
+      total_training_steps: 1620
+      weight_decay: 0.01
+    fsdp_config:
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      fsdp_size: -1
+  ref:
+    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: 0.75
+    ignore_eos: false
+    enforce_eager: false
+    free_cache_engine: false
+    load_format: dummy_dtensor
+    tensor_model_parallel_size: 4
+    max_num_batched_tokens: 4864
+    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
+      'n': 1
+      do_sample: false
+critic:
+  rollout_n: 1
+  strategy: fsdp
+  optim:
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    min_lr_ratio: null
+    warmup_style: constant
+    total_training_steps: 1620
+    weight_decay: 0.01
+  model:
+    path: meta-llama/Llama-3.2-3B-Instruct
+    tokenizer_path: meta-llama/Llama-3.2-3B-Instruct
+    override_config: {}
+    external_lib: null
+    enable_gradient_checkpointing: false
+    use_remove_padding: true
+    fsdp_config:
+      param_offload: false
+      optimizer_offload: false
+      wrap_policy:
+        min_num_params: 0
+      fsdp_size: -1
+    save_hf_repo_id: RyanYr/ppo-dapo-llama3.2-3B-it-lr-mbs64_critic
+  ppo_mini_batch_size: 64
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: 2
+  forward_micro_batch_size: null
+  forward_micro_batch_size_per_gpu: 2
+  use_dynamic_bsz: false
+  ppo_max_token_len_per_gpu: 32768
+  forward_max_token_len_per_gpu: 32768
+  ulysses_sequence_parallel_size: 1
+  ppo_epochs: 1
+  shuffle: false
+  grad_clip: 1.0
+  cliprange_value: 0.5
+  checkpoint:
+    contents:
+    - model
+    - optimizer
+    - extra
+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: 32768
+  reward_manager: prime
+custom_reward_function:
+  path: null
+  name: compute_score
+algorithm:
+  gamma: 1.0
+  lam: 1.0
+  adv_estimator: gae
+  use_kl_in_reward: true
+  kl_penalty: kl
+  kl_ctrl:
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+trainer:
+  balance_batch: true
+  total_epochs: 12
+  total_training_steps: null
+  project_name: value-LLM
+  experiment_name: ppo-dapo-llama3.2-3B-it_lr-mbs64
+  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
+  critic_warmup: 0
+  default_hdfs_dir: null
+  del_local_ckpt_after_load: false
+  default_local_dir: checkpoints/value-LLM/ppo-dapo-llama3.2-3B-it_lr-mbs64
+  max_actor_ckpt_to_keep: 1
+  max_critic_ckpt_to_keep: 1
+  ray_wait_register_center_timeout: 300
+  hf_token: null
+  resume_from_hf:
+    enable: false
+    actor_hf_repo_id: null
+    actor_revision: main
+    critic_hf_repo_id: null
+    critic_revision: main
+    hf_token: null

extra_state_world_size_4_rank_0.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc290cf68578c2e73d6fd0e043d9913f77ab5d6f64a22b0db54130a7dc318c45
+size 14632

extra_state_world_size_4_rank_1.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c727c113079dcd66af227f11cb63bf87ec9554dc475ed3baf5233b98c725834f
+size 14632

extra_state_world_size_4_rank_2.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ecd530762fc74b332d0dc475fb0a6550a58f9316c9c6b760778d182aebd75ee8
+size 14632

extra_state_world_size_4_rank_3.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f26d4ad28de99b3331a02af6f4f9d964244634261e00c7f674c916f095f75c1
+size 14632

model_world_size_4_rank_0.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a382d9098513b634fa46bb9f7b8c1f73f6da0e0ac0d46ed8d9ba85cf0a812b71
+size 3212915290

model_world_size_4_rank_1.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6ad679f10b1528dfc075491086844c997851283b25e56945939361f978428740
+size 3212903002

model_world_size_4_rank_2.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:925449ab8391027d3192e38172dc2d872cc2ecb1d735eadc0a0ddda8118f921e
+size 3212903002

model_world_size_4_rank_3.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:04eef8f99b762bb18bc562b0d7f46bcc3cf401956f0802ab2b5cc021fe81ab35
+size 3212903002

optim_world_size_4_rank_0.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7ad7011cc3769103f0256049b066a28f65e27e5179fd6aafbccf358cee0c3ee0
+size 6425532375

optim_world_size_4_rank_1.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a42bffe39e88ee45b24deafb6fb4b71582afa22980420df586df01dc5ece966d
+size 6425532375

optim_world_size_4_rank_2.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6e1eb3eec2cd23b023aa8a7340c7a0496c412fd675d62aabe25c6a8e4acad89b
+size 6425532375

optim_world_size_4_rank_3.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4fbb5a9e45a8ca5f5a719c16899e480c446358dbebd805d301eba065f670f85c
+size 6425532375