apiche

Apiche's group workspace

Group: debug_dataset_buffer_size_again

1-3

of 3

Tags

Notes

Author

apiche

State

Crashed

Start time

July 17th, 2025 9:06:46 PM

Runtime

21h 49m 29s

Tracked hours

3h 12m 14s

Run path

apiche/pipeline-rl/debug_dataset_buffer_size_again_finetune

Linux-5.15.0-1067-nvidia-x86_64-with-glibc2.39

Python version

CPython 3.11.11

Git repository

git clone git@github.com:ServiceNow/pipelinerl.git

Git state

git checkout -b "debug_dataset_buffer_size_again/finetune" a9030a7e0fcd6ae9de0539c73b91e6b45f731ef3

Command

pipelinerl/entrypoints/run_finetune.py --config-dir results/debug_dataset_buffer_size_again/conf --config-name exp_config output_dir=results/debug_dataset_buffer_size_again hydra.run.dir=results/debug_dataset_buffer_size_again/finetune +me.weight_update_group_init_method=tcp://localhost:9000 +me.weight_update_group_world_size=3 +me.llm_urls=http://localhost:8080+http://localhost:8081

System Hardware

CPU count	112
Logical CPU count	224
GPU count	4
GPU type	NVIDIA H100 80GB HBM3

W&B CLI Version

0.19.11

Group

debug_dataset_buffer_size_again

Config parameters are your model's inputs. Learn more

▶
Config parameters:{} 182 keys
- _cpu:
  "False"
- _mixed_precision:
  "no"
- accelerate_config:
  null
- actor.discount_factor:
  1
- actor.llm_max_rollouts:
  64
- actor.log_each_n_secs:
  0
- actor.problem_queue_size:
  64
- actor.result_queue_size:
  64
- actor.rollout_policy:
  "pipelinerl.domains.math.generate_math_rollout"
- actor.rollout_workers:
  1
- actor.shared_memory_entry_size:
  10,000,000
- actor.system_prompt:
  "Please reason step by step, and put your final answer within \boxed{}."
- actor.task_template:
  "{task}"
- actor.throughput_window_size:
  50
- attempts:
  1
- backend:
  "nccl"
- dataset_loader:
  "pipelinerl.domains.math.load_datasets"
- debug:
  "False"
- debug.mode:
  ""
- debug.place_inference_workers:
  true
- debug.streams_from:
  null
- debug.use_existing_llms:
  false
- deepspeed_config:
  "deepspeed_stage3_bf16"
- deepspeed_plugins:
  "DeepSpeedPlugin(hf_ds_config=<accelerate.utils.deepspeed.HfDeepSpeedConfig object at 0x7ffc77091650>, gradient_accumulation_steps=1, gradient_clipping='auto', zero_stage=3, is_train_batch_min=True, offload_optimizer_device='none', offload_param_device='none', offload_optimizer_nvme_path='none', offload_param_nvme_path='none', zero3_init_flag=True, zero3_save_16bit_model=True, transformer_moe_cls_names=None, enable_msamp=False, msamp_opt_level='O1')"
- device:
  "cuda:0"
- distributed_type:
  "DistributedType.DEEPSPEED"
- dynamo_plugin:
  "TorchDynamoPlugin(backend=<DynamoBackend.NO: 'NO'>, mode='default', fullgraph=False, dynamic=False, options=None, disable=False)"
- environment._target_:
  "pipelinerl.domains.math.MathEnvironment"
- eval_every_n_versions:
  78,000
- finetune.also_save_steps:[] 0 items
- finetune.attn_implementation:
  "flash_attention_2"
- finetune.auto_device_map:
  false
- finetune.config_name:
  "Qwen/Qwen2.5-1.5B-Instruct"
- finetune.cuda_empty_cache:
  true
- finetune.data:
  null
- finetune.eval_callback._target_:
  "tapeagents.finetune.eval.dummy_eval_callback"
- finetune.eval_callback.config_name:
  ""
- finetune.force_restart:
  true
- finetune.gradient_accumulation_passes:
  256
- finetune.gradient_checkpointing:
  true
- finetune.gradient_clipping_threshold:
  0.3
- finetune.input:
  "training_data"
- finetune.interrupt_train_steps:
  -1
- finetune.keep_intermediate_checkpoints:
  true
- finetune.learning_rate:
  0
- finetune.load_as_bf16:
  true
- world.environment_start_port:
  7,777
- world.finetune_fraction:
  4
- world.preprocessor_fraction:
  0
- world.replicas:
  1

Summary metrics are your model's outputs. Learn more

▶
Summary metrics:{} 69 keys
- rl/advantage:
  0.1459902971982956
- rl/clamp_log_ratio_new_old_indicator:
  0
- rl/clamp_log_ratio_ref_new_indicator:
  0
- rl/entropy:
  0.4675585627555847
- rl/entropy_bonus_coef:
  0
- rl/ess:
  0.9992681397184988
- rl/input_size:
  918.6015625
- rl/kl:
  0.0004211392952129245
- rl/kl_coef:
  0
- rl/loss:
  47.85908508300781
- rl/max_advantage:
  1.5078125
- rl/max_kl:
  0.27544963359832764
- rl/max_loss:
  6.1685791015625
- rl/max_reward:
  1
- rl/max_token_weight:
  0.00390625
- rl/min_advantage:
  -0.5546875
- rl/min_kl:
  -0.00000011920928955078
- rl/min_loss:
  0.01759498193860054
- rl/min_reward:
  0
- rl/min_token_weight:
  0.00390625
- rl/new_logprobs:
  -0.4707274436950683
- rl/num_nans:
  0
- rl/num_output_tokens_sum:
  208,609
- rl/old_logprobs:
  -0.470413476228714
- rl/policy_loss:
  -0.07589671015739441
- rl/ratio_new_old:
  1.0001059770584106
- rl/ratio_new_old_squared_sum:
  208,816.546875
- rl/ratio_new_old_sum:
  208,636.359375
- rl/ratio_ref_new:
  1.0007351636886597
- rl/ratio_ref_old:
  1
- rl/ref_logprobs:
  -0.470413476228714
- rl/reward:
  0.15625
- rl/surr1:
  0
- rl/surr2:
  0
- rl/token_weight:
  0.00390625
- rl/value_loss:
  0.2074708491563797
- rl/value_max:
  0.5546875
- rl/value_mean:
  0.01025969348847866
- rl/value_min:
  -0.5078125
- rl/value_mse:
  0.17431235313415527
- stats/completed_steps:
  14
- stats/epoch:
  0
- stats/grad_norm:
  3.8672189818135014
- stats/lag:
  0
- stats/lr:
  0
- stats/max_actor_version:
  3,072
- throughput/tokens_per_micro_batch:
  10,660.4
- throughput/tokens_per_sec:
  22,123.64113206965
- throughput/tokens_per_step:
  213,208
- throughput/tokens_perGPU_per_sec:
  11,061.820566034825

This run produced these artifacts as outputs. Total: 1. Learn more

wandb-history

run-debug_dataset_buffer_size_again_finetune-history:v0