Save model at global step 5

config.yaml

@@ -0,0 +1,224 @@
+data:
+  tokenizer: null
+  train_files: aime24_ds_train_sample.parquet
+  val_files: matheval.parquet
+  prompt_key: prompt
+  reward_fn_key: data_source
+  max_prompt_length: 1024
+  max_response_length: 3072
+  train_batch_size: 256
+  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-aime24-qwen2.5math-1.5B-base-mbs128-n4_actor
+    tokenizer_chat_template: null
+  actor:
+    strategy: fsdp
+    ppo_mini_batch_size: 128
+    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: 200
+      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: 1024
+    response_length: 3072
+    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: 200
+    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: 128
+  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: 200
+  project_name: value-LLM
+  experiment_name: grpo-aime24-qwen2.5math-1.5B-base-mbs128-n4
+  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/grpo-aime24-qwen2.5math-1.5B-base-mbs128-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:a4d7a5086567ce89f4490ecd05565998e017d7b1f9b0275bb174c98d9441607b
+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 sha256:186b43acaf161c319d98e0350f446280663e9ca2ac147cd812631eef8c2f3cba
+size 1777276538

model_world_size_4_rank_1.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:86cc0336d80c31df11103c4f174b751930f8b442666f45db29acffc6e343d31e
+size 1777276538

model_world_size_4_rank_2.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:10222af218993908ca92a0003aa2f3a116d638e13f5bec596d26de1ec82f8713
+size 1777276538

model_world_size_4_rank_3.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:09facae073e61017dead35d3f92f7894031b5f4636c807ef42879da1a82d3891
+size 1777276538

optim_world_size_4_rank_0.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cc76658f90f9ac30c6ca03f23b4e44549e79deb2ab0b1acb71e96ac4b7cac40f
+size 3087454775

optim_world_size_4_rank_1.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:62487419c231a219a15ad5e68068aaef343f4d795837f4d6373f0348fd5db71f
+size 3087454775

optim_world_size_4_rank_2.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:52c3103518e6f431f0741fb259f13642582f7109dd46d4e78d82e4ac283c6216
+size 3087454775

optim_world_size_4_rank_3.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3b6be44d72ceddd9757b5d9d7c950636e5e538484a6adbe7b599b5bd3b93cdf3
+size 3087454775