Save model at global step 40

config.yaml

@@ -0,0 +1,224 @@
+data:
+  tokenizer: null
+  train_files: dapo-r1_ds_train_sample.parquet
+  val_files: matheval.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: 768
+  max_response_length: 3328
+  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: Qwen/Qwen2.5-Math-1.5B
+    external_lib: null
+    override_config: {}
+    enable_gradient_checkpointing: true
+    use_remove_padding: true
+    use_liger: false
+    save_hf_repo_id: RyanYr/grpo-respL-dapo-r1-qwen2.5math-1.5B-base-mbs64-n4_actor
+    tokenizer_chat_template: deepseek-r1-plain
+  actor:
+    strategy: fsdp
+    ppo_mini_batch_size: 64
+    ppo_micro_batch_size: null
+    ppo_micro_batch_size_per_gpu: 16
+    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
+    use_kl_loss: true
+    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: 400
+      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: 64
+    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: 3328
+    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: 4096
+    max_model_len: null
+    max_num_seqs: 1024
+    log_prob_micro_batch_size: null
+    log_prob_micro_batch_size_per_gpu: 64
+    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': 4
+    engine_kwargs:
+      swap_space: null
+    val_kwargs:
+      top_k: -1
+      top_p: 1.0
+      temperature: 0
+      'n': 1
+      do_sample: false
+critic:
+  rollout_n: 4
+  strategy: fsdp
+  optim:
+    lr: 1.0e-05
+    lr_warmup_steps_ratio: 0.0
+    min_lr_ratio: null
+    warmup_style: constant
+    total_training_steps: 400
+    weight_decay: 0.01
+  model:
+    path: ~/models/deepseek-llm-7b-chat
+    tokenizer_path: Qwen/Qwen2.5-Math-1.5B
+    override_config: {}
+    external_lib: null
+    enable_gradient_checkpointing: true
+    use_remove_padding: false
+    fsdp_config:
+      param_offload: false
+      optimizer_offload: false
+      wrap_policy:
+        min_num_params: 0
+      fsdp_size: -1
+    save_hf_repo_id: null
+  ppo_mini_batch_size: 64
+  ppo_micro_batch_size: null
+  ppo_micro_batch_size_per_gpu: null
+  forward_micro_batch_size: null
+  forward_micro_batch_size_per_gpu: null
+  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: Qwen/Qwen2.5-Math-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: grpo
+  use_kl_in_reward: false
+  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: 400
+  project_name: value-LLM
+  experiment_name: grpo-respL-dapo-r1-qwen2.5math-1.5B-base-mbs64-n4
+  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/grpo-respL-dapo-r1-qwen2.5math-1.5B-base-mbs64-n4
+  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 sha256:461914b89a787ce8f050b5de9bb0dc5b8848b41b8065137b4ac2b4f007f41d9b
+size 1492

extra_state_world_size_4_rank_0.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1ab08b34f7000fc99b714608686487f4b0ef5f419a1c23610e1c44c7aaad8137
+size 14632

extra_state_world_size_4_rank_1.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af9fc86f6e469cb0834476b71b525358133f74ff5ee77ed4bce6eac57aa550ca
+size 14632

extra_state_world_size_4_rank_2.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d321ba942354b0228acd80bb829660d58c1006ff7e77e69386c788b91137c6b8
+size 14632

extra_state_world_size_4_rank_3.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:720ecba3014916b6c126dc326620e82c0dc2b8c061000f556b103b296edc5656
+size 14632

model_world_size_4_rank_0.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a7070b5ae42aa1ca2492ac6c0689cd6c9f8a47b4ee24475e995f7133b83796a1
+size 1777276538

model_world_size_4_rank_1.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:07dbc7518d9133d25f43d0d52bdfe00c5fd6a22d6ca2686065bf4cbb473bbdb7
+size 1777276538

model_world_size_4_rank_2.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e6ecb12d7a13cb77265f8c091d066ba51a400759b02f2f01483e505c87a64cb2
+size 1777276538

model_world_size_4_rank_3.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:57760d35f7aac726cdab7468354dc82203e988a6c8e085e3a6130bdf579b3edd
+size 1777276538

optim_world_size_4_rank_0.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e8cb8dd3ead82d0970dcd634d3bd467215bb197714f1a04a7041ef39cf0f5807
+size 3087454775

optim_world_size_4_rank_1.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:179f27f4a81c958b9a0d3c577e2bd98e7c318c4d175ea0d09c4a2b0957c3fd40
+size 3087454775

optim_world_size_4_rank_2.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:74375c68e623834a56ece1745747d213d724c77f11e40669cd282a3d67711030
+size 3087454775

optim_world_size_4_rank_3.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3256c37225fd7d18609da6b72446ea844c89b1923b3be8870e3584c4f0a79717
+size 3087454775