Save model at global step 5

config.yaml

@@ -0,0 +1,224 @@
+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: 13312
+  train_batch_size: 1024
+  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: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    use_remove_padding: true
+    use_liger: false
+    save_hf_repo_id: RyanYr/ppo-dapo-r1qwen1.5B-base-lr-mbs256_actor
+    tokenizer_chat_template: null
+  actor:
+    strategy: fsdp
+    ppo_mini_batch_size: 256
+    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
+    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: 100
+      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: 1
+    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: 13312
+    dtype: bfloat16
+    gpu_memory_utilization: 0.7
+    ignore_eos: false
+    enforce_eager: false
+    free_cache_engine: false
+    load_format: dummy_dtensor
+    tensor_model_parallel_size: 4
+    max_num_batched_tokens: 14080
+    max_model_len: null
+    max_num_seqs: 1024
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: 1
+    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: 100
+    weight_decay: 0.01
+  model:
+    path: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
+    tokenizer_path: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
+    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-r1qwen1.5B-base-lr-mbs256_critic
+  ppo_mini_batch_size: 256
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: 1
+  forward_micro_batch_size: null
+  forward_micro_batch_size_per_gpu: 1
+  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: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
+    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: 100
+  total_training_steps: 100
+  project_name: value-LLM
+  experiment_name: ppo-dapo-r1qwen1.5B-base_lr-mbs256
+  logger:
+  - console
+  - wandb
+  log_val_generations: 0
+  nnodes: 1
+  n_gpus_per_node: 4
+  save_freq: 5
+  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-r1qwen1.5B-base_lr-mbs256
+  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

data.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:09588a074d0c922390123da1d3852004cb21d12d5f82876411cc90ca54e0eda0
+size 1492

extra_state_world_size_4_rank_0.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9643feae11f4f0e283422cef05239947ebd23ec4f3843fe7fcc04142509148dc
+size 14632

extra_state_world_size_4_rank_1.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c5fcce58b4da9fde5c7bb2601be01ac994234e59af6cddc57fa5ec320abfb5c2
+size 14632

extra_state_world_size_4_rank_2.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:510a69fb2b013c3a8d5d7ce19e81741d496a5a06d7bcab6c3a147b7853f8e9ac
+size 14632

extra_state_world_size_4_rank_3.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f337c45b712efaaab96794f367a02895f264bb168328bf922f713c35d0bc27c8
+size 14632

model_world_size_4_rank_0.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:77cfbb5fd1cc893640b17ee31f19dcc15850a21a3c0129a01f2ce992c1e7a291
+size 1777276538

model_world_size_4_rank_1.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2e151605b92aefc51b01502c210b10f0343b41aab09b4ea3b106d12d3fea036a
+size 1777276538

model_world_size_4_rank_2.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:965994447a6722be4274387a1e482a6e181af2b205b04184aa85ebcc215a6b61
+size 1777276538

model_world_size_4_rank_3.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ae02cbe7f949dbcc7f5079299f672c4e122bce24a596ffe8f6eb78e93f24e638
+size 1777276538

optim_world_size_4_rank_0.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad3f56f87d518fe22c6cd7a86d7ddba76b04c10a5e85821b8b39da094ac1eed9
+size 3554202167

optim_world_size_4_rank_1.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8593ae4f399369ecd7a61ac744d1cfdafbf5d599f601ee6913001284ad1bca0d
+size 3554202167

optim_world_size_4_rank_2.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e52f706d8d05980bfc18050f6e819e03ffbd63c9b1fb85471381cd536921ad31
+size 3554202167

optim_world_size_4_rank_3.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f560680942fe2e7345a29e4a881b6b9a3c03236c445af52c2faa856bfe221855
+size 3554202167