apiche

Apiche's group workspace

Group: debug_gspo12

1-3

of 3

Tags

Notes

Author

apiche

State

Crashed

Start time

September 27th, 2025 11:52:56 PM

Runtime

2h 48m 34s

Tracked hours

45m 48s

Run path

apiche/pipeline-rl/debug_gspo12_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_gspo12/finetune" 9e31adf0b44882c5c296b7a083de7c1e82353394

Command

pipelinerl/entrypoints/run_finetune.py --config-dir results/debug_gspo12/conf --config-name exp_config output_dir=results/debug_gspo12 hydra.run.dir=results/debug_gspo12/finetune +me.weight_update_group_init_method=tcp://localhost:9000 +me.weight_update_group_world_size=5 +me.llm_urls=http://localhost:8080+http://localhost:8081+http://localhost:8082+http://localhost:8083

System Hardware

CPU count	112
Logical CPU count	224
GPU count	8
GPU type	NVIDIA H100 80GB HBM3

W&B CLI Version

0.19.11

Group

debug_gspo12

Config parameters are your model's inputs. Learn more

▶
Config parameters:{} 186 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_prompt:
  ""
- 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 0x7ffc246da510>, 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=None, options=None, disable=False, use_regional_compilation=False)"
- environment._target_:
  "pipelinerl.domains.math.MathEnvironment"
- eval_every_n_versions:
  0
- finetune.also_save_steps:[] 0 items
- finetune.attempts:
  1
- finetune.attn_implementation:
  "flash_attention_2"
- finetune.auto_device_map:
  false
- finetune.config_name:
  "Qwen/Qwen2.5-0.5B"
- 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:
  16
- finetune.gradient_checkpointing:
  true
- finetune.gradient_clipping_threshold:
  0.3
- finetune.input:
  "training_data"
- finetune.interrupt_train_steps:
  -1
- finetune.keep_intermediate_checkpoints:
  true
- world.environment_start_port:
  7,777
- world.finetune_fraction:
  1
- world.preprocessor_fraction:
  0
- world.replicas:
  1

Summary metrics are your model's outputs. Learn more

▶
Summary metrics:{} 65 keys
- rl/advantage:
  -12.359375
- rl/clamp_log_ratio_ref_new_indicator:
  0
- rl/entropy:
  0.01837158203125
- rl/entropy_bonus_coef:
  0
- rl/ess:
  0.9999998455639107
- rl/input_size:
  99.25
- rl/kl:
  0.00000061094760894775
- rl/kl_coef:
  0
- rl/kl_new_old:
  0.00000063329935073853
- rl/loss:
  -189.3769073486328
- rl/max_advantage:
  -13.125
- rl/max_kl:
  0.00000095367431640625
- rl/max_loss:
  -25.255136489868164
- rl/max_reward:
  0
- rl/max_token_weight:
  0.0625
- rl/min_advantage:
  -13.1875
- rl/min_kl:
  0.00000023841857910156
- rl/min_loss:
  -75.77116394042969
- rl/min_reward:
  0
- rl/min_token_weight:
  0.0625
- rl/new_logprobs:
  -0.0015120506286621094
- rl/num_output_tokens_sum:
  15
- rl/old_logprobs:
  -0.0004391679249238223
- rl/ratio_new_old:
  0.9364277124404908
- rl/ratio_new_old_squared_sum:
  14.96570873260498
- rl/ratio_new_old_sum:
  14.982843399047852
- rl/ratio_ref_new:
  0.9385735392570496
- rl/ratio_ref_old:
  0.9375
- rl/ref_logprobs:
  -0.0004391679249238223
- rl/reward:
  0
- rl/token_weight:
  0.05859375
- rl/value_loss:
  81.468994140625
- rl/value_max:
  13.1875
- rl/value_mean:
  12.359375
- rl/value_min:
  13.125
- rl/value_mse:
  162.93798828125
- stats/completed_steps:
  1,141
- stats/epoch:
  0
- stats/grad_norm:
  53.05396074183922
- stats/lag:
  304
- stats/lr:
  0.000001
- stats/max_actor_version:
  17,936
- stats/min_actor_version:
  17,936
- stats/passes:
  1,145
- stats/queue/batches:
  1
- stats/samples:
  18,256
- throughput/tokens_per_micro_batch:
  397
- throughput/tokens_per_sec:
  2,659.8437979086425
- throughput/tokens_per_step:
  1,588
- throughput/tokens_perGPU_per_sec:
  664.9609494771606

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

wandb-history

run-debug_gspo12_finetune-history:v0