Save model at global step 1235

config.yaml

@@ -0,0 +1,227 @@
+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_neg0.01-aime24-qwen2.5math-1.5B-base-mbs128-n4-ref1230-975b46d_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: 2000
+      weight_decay: 0.01
+    fsdp_config:
+      wrap_policy:
+        min_num_params: 0
+      param_offload: false
+      optimizer_offload: false
+      fsdp_size: -1
+  ref:
+    ref_model_path: RyanYr/grpo-aime24-qwen2.5math-1.5B-base-mbs128-n4_actor_1230-975b46d
+    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: 2000
+    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_neg
+  use_kl_in_reward: false
+  grpo_neg:
+    mean_penalty: 0.01
+  kl_penalty: kl
+  kl_ctrl:
+    type: fixed
+    kl_coef: 0.001
+    horizon: 10000
+    target_kl: 0.1
+trainer:
+  balance_batch: true
+  total_epochs: 1000000000000
+  total_training_steps: 2000
+  project_name: value-LLM
+  experiment_name: grpo_neg0.01-aime24-qwen2.5math-1.5B-base-mbs128-n4-ref1230-975b46d
+  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_neg0.01-aime24-qwen2.5math-1.5B-base-mbs128-n4-ref1230-975b46d
+  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: true
+    actor_hf_repo_id: RyanYr/grpo-aime24-qwen2.5math-1.5B-base-mbs128-n4-ref895-82bb89a_actor
+    actor_revision: 975b46d1ee3ee658c46b85220a34a95c384f4078
+    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 sha256:9bc1bd198dfb33b20ea3ec9c6472f90016c8013b8bd571a556148caa4054af8a
+size 1492

extra_state_world_size_4_rank_0.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:128166d590b7b4c0ece3cac613ae83a5b32b8a808475bdf746d97e9cb75e7604
+size 14632

extra_state_world_size_4_rank_1.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:57f4f8e9a23b316750904c6ea4eaf1dc2ee49bbc7ae854e41de331e8261930dd
+size 14632

extra_state_world_size_4_rank_2.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac50bfad296a68ebe814cae144038535c2525f2de2342aba397a420a7460d806
+size 14632

extra_state_world_size_4_rank_3.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eafa30e154fae36c6e9384b02baab8b66a392e6e7edaf354d5f0eb0722512f3c
+size 14632

model_world_size_4_rank_0.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e76e166aedfba42b626aa7f805b702ddc8da8b0a16301f00dbd52ff0ba79de30
+size 1777276538

model_world_size_4_rank_1.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2bd98b5f6c4f68a6763498b6fe673fef4a9227bba8ff23ffac56e55de7d7f040
+size 1777276538

model_world_size_4_rank_2.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de07ff493005f7cfabb2b455d0874c859a05cbcc84a4d0dfe8819103bf64f1a0
+size 1777276538

model_world_size_4_rank_3.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cea24478fe54f2d20d8f05053e58ff9ddb4dad0dd5a8484dc6343735a5f2c467
+size 1777276538

optim_world_size_4_rank_0.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3681059ac0968807dd1023bf2df98d51db34b6b2d7aafd4b95ac8367d7ebe861
+size 3087454775

optim_world_size_4_rank_1.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5990306ed23970b15fb6729d5e44118008fcf5e42189556ed85877d7a060ffbd
+size 3087454775

optim_world_size_4_rank_2.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:950aff4abb79f05a1c676918bf5775f0ab26afb7f8603b9cfa239bbdcc88a3e6
+size 3087454775

optim_world_size_4_rank_3.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3cba4b3950d3de6548265c62252c922ef5b90f7bd71dc369ae4a4b9e9fa7bd01
+size 3087454775