The following have been reloaded with a version change: 1) GCCcore/.14.3.0 => GCCcore/14.3.0 Lmod is automatically replacing "GCC/14.3.0" with "nvidia-compilers/25.9-CUDA-13". Deactivating conda environment: /e/scratch/jureap59/feuer1/miniforge3/envs/otagent Activating RL environment: /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl Python executable: /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/bin/python Python path check: /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/bin/python [ray] RAY_TMPDIR=/tmp/ray/ray_653584 [triton_cache] Triton cache: /tmp/triton_cache_feuer1_653584 [triton_cache] TorchInductor cache: /tmp/torchinductor_cache_feuer1_653584 [proxy] ✓ Found proxychains binary at /e/scratch/jureap59/feuer1/proxychains-ng-aarch64/bin/proxychains4 [proxy] Setting up SSH tunnel to jpbl-s01-01 [proxy] SSH key: /e/home/jusers/feuer1/jupiter/.ssh/authorized_keys/id_ed25519_jsc [proxy] Tunnel port: 7003 [proxy] Node IP: 10.128.16.35 (workers will connect here) [proxy] ✓ SSH tunnel started successfully [proxy] ✓ Generated proxychains config at /e/home/jusers/feuer1/jupiter/.proxychains/proxychains_653584.conf [proxy] - Internal traffic (10.x.x.x, 172.x.x.x, 169.254.x.x) → DIRECT [proxy] - External traffic (internet) → PROXY via tunnel [proxy] ✓ Daytona timeout settings configured [proxy] Testing proxy connectivity... [proxychains] config file found: /e/home/jusers/feuer1/jupiter/.proxychains/proxychains_653584.conf [proxychains] preloading /e/scratch/jureap59/feuer1/proxychains-ng-aarch64/lib/libproxychains4.so [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e [proxy] ✓ Proxy connectivity test passed (huggingface.co reachable via wrapped binary) [proxy] ⚠ Tunnel not accessible at 10.128.16.35:7003 (workers may fail) [proxy] ✓ Proxy setup complete (using wrapped binary for Ray workers) [container_runtime] Using cloud backend: daytona (no local container setup) === Universal RL Training Runner === Config: /e/data1/datasets/playground/ot-baf/ablation-pymethods2test-seqmean-arm0-tis/configs/ablation-pymethods2test-seqmean-arm0-tis_rl_config.json Working directory: /e/scratch/jureap59/feuer1/OpenThoughts-Agent Python: /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/bin/python Python version: Python 3.12.12 UV_USE_IO_URING: 0 Proxy: DISABLED (direct internet or not configured) ======================================== === RLJobRunner: ablation-pymethods2test-seqmean-arm0-tis === [wandb_utils] Fixing permissions on: /e/data1/datasets/playground/ot-baf/ablation-pymethods2test-seqmean-arm0-tis/wandb [wandb_utils] WandB directory ready: /e/data1/datasets/playground/ot-baf/ablation-pymethods2test-seqmean-arm0-tis/wandb HF_TOKEN=****pDbg HF_HUB_CACHE=/e/data1/datasets/playground/ot-baf/hf_hub SUPABASE_URL=https://rpzmyuapoqilpghynmza.s... (direct Supabase config) Environment configured: TENSOR_PARALLEL_SIZE=1 NUM_INFERENCE_ENGINES=56 POLICY_NUM_NODES=14 WANDB_DIR=/e/data1/datasets/playground/ot-baf/ablation-pymethods2test-seqmean-arm0-tis/wandb Starting Ray cluster with 14 nodes, 4 GPUs/node Cleaning up existing Ray instances... === Starting Ray Cluster === Nodes: 14 GPUs per node: 4 CPUs per node: 288 Head node: jpbo-001-35 (10.128.16.35) Ray port: 6379 ============================ Starting Ray head on jpbo-001-35 (logging to /e/data1/datasets/playground/ot-baf/experiments/_ray_logs/ray_head_jpbo-001-35.log)... Command: srun --export=ALL,VLLM_HOST_IP=10.128.16.35 --nodes=1 --ntasks=1 --gres=gpu:4 --gpu-bind=none --overlap --cpu-bind=none -w jpbo-001-35 bash -c /e/scratch/jureap59/feuer1/proxychains-ng-aarch64/bin/proxychains4 -f "$PROXYCHAINS_CONF_FILE" ray start --head --node-ip-address=10.128.16.35 --port=6379 --num-gpus=4 --num-cpus=288 --block --object-store-memory=42949672960 Started Ray head on jpbo-001-35 Starting Ray worker on jpbo-001-36 (logging to /e/data1/datasets/playground/ot-baf/experiments/_ray_logs/ray_worker_jpbo-001-36.log)... Command: srun --export=ALL,VLLM_HOST_IP=10.128.16.36 --nodes=1 --ntasks=1 --gres=gpu:4 --gpu-bind=none --overlap --cpu-bind=none -w jpbo-001-36 bash -c /e/scratch/jureap59/feuer1/proxychains-ng-aarch64/bin/proxychains4 -f "$PROXYCHAINS_CONF_FILE" ray start --address=10.128.16.35:6379 --node-ip-address=10.128.16.36 --num-gpus=4 --num-cpus=288 --block --object-store-memory=42949672960 Started Ray worker 1 on jpbo-001-36 Starting Ray worker on jpbo-001-37 (logging to /e/data1/datasets/playground/ot-baf/experiments/_ray_logs/ray_worker_jpbo-001-37.log)... Command: srun --export=ALL,VLLM_HOST_IP=10.128.16.37 --nodes=1 --ntasks=1 --gres=gpu:4 --gpu-bind=none --overlap --cpu-bind=none -w jpbo-001-37 bash -c /e/scratch/jureap59/feuer1/proxychains-ng-aarch64/bin/proxychains4 -f "$PROXYCHAINS_CONF_FILE" ray start --address=10.128.16.35:6379 --node-ip-address=10.128.16.37 --num-gpus=4 --num-cpus=288 --block --object-store-memory=42949672960 Started Ray worker 2 on jpbo-001-37 Starting Ray worker on jpbo-001-38 (logging to /e/data1/datasets/playground/ot-baf/experiments/_ray_logs/ray_worker_jpbo-001-38.log)... Command: srun --export=ALL,VLLM_HOST_IP=10.128.16.38 --nodes=1 --ntasks=1 --gres=gpu:4 --gpu-bind=none --overlap --cpu-bind=none -w jpbo-001-38 bash -c /e/scratch/jureap59/feuer1/proxychains-ng-aarch64/bin/proxychains4 -f "$PROXYCHAINS_CONF_FILE" ray start --address=10.128.16.35:6379 --node-ip-address=10.128.16.38 --num-gpus=4 --num-cpus=288 --block --object-store-memory=42949672960 Started Ray worker 3 on jpbo-001-38 Starting Ray worker on jpbo-001-41 (logging to /e/data1/datasets/playground/ot-baf/experiments/_ray_logs/ray_worker_jpbo-001-41.log)... Command: srun --export=ALL,VLLM_HOST_IP=10.128.16.41 --nodes=1 --ntasks=1 --gres=gpu:4 --gpu-bind=none --overlap --cpu-bind=none -w jpbo-001-41 bash -c /e/scratch/jureap59/feuer1/proxychains-ng-aarch64/bin/proxychains4 -f "$PROXYCHAINS_CONF_FILE" ray start --address=10.128.16.35:6379 --node-ip-address=10.128.16.41 --num-gpus=4 --num-cpus=288 --block --object-store-memory=42949672960 Started Ray worker 4 on jpbo-001-41 Starting Ray worker on jpbo-001-42 (logging to /e/data1/datasets/playground/ot-baf/experiments/_ray_logs/ray_worker_jpbo-001-42.log)... Command: srun --export=ALL,VLLM_HOST_IP=10.128.16.42 --nodes=1 --ntasks=1 --gres=gpu:4 --gpu-bind=none --overlap --cpu-bind=none -w jpbo-001-42 bash -c /e/scratch/jureap59/feuer1/proxychains-ng-aarch64/bin/proxychains4 -f "$PROXYCHAINS_CONF_FILE" ray start --address=10.128.16.35:6379 --node-ip-address=10.128.16.42 --num-gpus=4 --num-cpus=288 --block --object-store-memory=42949672960 Started Ray worker 5 on jpbo-001-42 Starting Ray worker on jpbo-001-43 (logging to /e/data1/datasets/playground/ot-baf/experiments/_ray_logs/ray_worker_jpbo-001-43.log)... Command: srun --export=ALL,VLLM_HOST_IP=10.128.16.43 --nodes=1 --ntasks=1 --gres=gpu:4 --gpu-bind=none --overlap --cpu-bind=none -w jpbo-001-43 bash -c /e/scratch/jureap59/feuer1/proxychains-ng-aarch64/bin/proxychains4 -f "$PROXYCHAINS_CONF_FILE" ray start --address=10.128.16.35:6379 --node-ip-address=10.128.16.43 --num-gpus=4 --num-cpus=288 --block --object-store-memory=42949672960 Started Ray worker 6 on jpbo-001-43 Starting Ray worker on jpbo-001-44 (logging to /e/data1/datasets/playground/ot-baf/experiments/_ray_logs/ray_worker_jpbo-001-44.log)... Command: srun --export=ALL,VLLM_HOST_IP=10.128.16.44 --nodes=1 --ntasks=1 --gres=gpu:4 --gpu-bind=none --overlap --cpu-bind=none -w jpbo-001-44 bash -c /e/scratch/jureap59/feuer1/proxychains-ng-aarch64/bin/proxychains4 -f "$PROXYCHAINS_CONF_FILE" ray start --address=10.128.16.35:6379 --node-ip-address=10.128.16.44 --num-gpus=4 --num-cpus=288 --block --object-store-memory=42949672960 Started Ray worker 7 on jpbo-001-44 Starting Ray worker on jpbo-001-45 (logging to /e/data1/datasets/playground/ot-baf/experiments/_ray_logs/ray_worker_jpbo-001-45.log)... Command: srun --export=ALL,VLLM_HOST_IP=10.128.16.45 --nodes=1 --ntasks=1 --gres=gpu:4 --gpu-bind=none --overlap --cpu-bind=none -w jpbo-001-45 bash -c /e/scratch/jureap59/feuer1/proxychains-ng-aarch64/bin/proxychains4 -f "$PROXYCHAINS_CONF_FILE" ray start --address=10.128.16.35:6379 --node-ip-address=10.128.16.45 --num-gpus=4 --num-cpus=288 --block --object-store-memory=42949672960 Started Ray worker 8 on jpbo-001-45 Starting Ray worker on jpbo-001-46 (logging to /e/data1/datasets/playground/ot-baf/experiments/_ray_logs/ray_worker_jpbo-001-46.log)... Command: srun --export=ALL,VLLM_HOST_IP=10.128.16.46 --nodes=1 --ntasks=1 --gres=gpu:4 --gpu-bind=none --overlap --cpu-bind=none -w jpbo-001-46 bash -c /e/scratch/jureap59/feuer1/proxychains-ng-aarch64/bin/proxychains4 -f "$PROXYCHAINS_CONF_FILE" ray start --address=10.128.16.35:6379 --node-ip-address=10.128.16.46 --num-gpus=4 --num-cpus=288 --block --object-store-memory=42949672960 Started Ray worker 9 on jpbo-001-46 Starting Ray worker on jpbo-001-47 (logging to /e/data1/datasets/playground/ot-baf/experiments/_ray_logs/ray_worker_jpbo-001-47.log)... Command: srun --export=ALL,VLLM_HOST_IP=10.128.16.47 --nodes=1 --ntasks=1 --gres=gpu:4 --gpu-bind=none --overlap --cpu-bind=none -w jpbo-001-47 bash -c /e/scratch/jureap59/feuer1/proxychains-ng-aarch64/bin/proxychains4 -f "$PROXYCHAINS_CONF_FILE" ray start --address=10.128.16.35:6379 --node-ip-address=10.128.16.47 --num-gpus=4 --num-cpus=288 --block --object-store-memory=42949672960 Started Ray worker 10 on jpbo-001-47 Starting Ray worker on jpbo-001-48 (logging to /e/data1/datasets/playground/ot-baf/experiments/_ray_logs/ray_worker_jpbo-001-48.log)... Command: srun --export=ALL,VLLM_HOST_IP=10.128.16.48 --nodes=1 --ntasks=1 --gres=gpu:4 --gpu-bind=none --overlap --cpu-bind=none -w jpbo-001-48 bash -c /e/scratch/jureap59/feuer1/proxychains-ng-aarch64/bin/proxychains4 -f "$PROXYCHAINS_CONF_FILE" ray start --address=10.128.16.35:6379 --node-ip-address=10.128.16.48 --num-gpus=4 --num-cpus=288 --block --object-store-memory=42949672960 Started Ray worker 11 on jpbo-001-48 Starting Ray worker on jpbo-003-37 (logging to /e/data1/datasets/playground/ot-baf/experiments/_ray_logs/ray_worker_jpbo-003-37.log)... Command: srun --export=ALL,VLLM_HOST_IP=10.128.16.133 --nodes=1 --ntasks=1 --gres=gpu:4 --gpu-bind=none --overlap --cpu-bind=none -w jpbo-003-37 bash -c /e/scratch/jureap59/feuer1/proxychains-ng-aarch64/bin/proxychains4 -f "$PROXYCHAINS_CONF_FILE" ray start --address=10.128.16.35:6379 --node-ip-address=10.128.16.133 --num-gpus=4 --num-cpus=288 --block --object-store-memory=42949672960 Started Ray worker 12 on jpbo-003-37 Starting Ray worker on jpbo-003-39 (logging to /e/data1/datasets/playground/ot-baf/experiments/_ray_logs/ray_worker_jpbo-003-39.log)... Command: srun --export=ALL,VLLM_HOST_IP=10.128.16.135 --nodes=1 --ntasks=1 --gres=gpu:4 --gpu-bind=none --overlap --cpu-bind=none -w jpbo-003-39 bash -c /e/scratch/jureap59/feuer1/proxychains-ng-aarch64/bin/proxychains4 -f "$PROXYCHAINS_CONF_FILE" ray start --address=10.128.16.35:6379 --node-ip-address=10.128.16.135 --num-gpus=4 --num-cpus=288 --block --object-store-memory=42949672960 Started Ray worker 13 on jpbo-003-39 Waiting for cluster (56 GPUs, 14 nodes)... Connecting to Ray at 10.128.16.35:6379 (expecting 14 nodes, 56.0 GPUs) Ray connection established, polling for resources... [Ray wait] nodes=14/14 GPUs=56.0/56.0 resources={'GPU': 56.0, 'accelerator_type:GH200': 14.0, 'node:10.128.16.43': 1.0, 'memory': 10289339891712.0, 'CPU': 4032.0, 'object_store_memory': 601295421440.0, 'node:10.128.16.47': 1.0, 'node:10.128.16.45': 1.0, 'node:10.128.16.46': 1.0, 'node:10.128.16.44': 1.0, 'node:10.128.16.133': 1.0, 'node:10.128.16.36': 1.0, 'node:__internal_head__': 1.0, 'node:10.128.16.35': 1.0, 'node:10.128.16.48': 1.0, 'node:10.128.16.135': 1.0, 'node:10.128.16.38': 1.0, 'node:10.128.16.42': 1.0, 'node:10.128.16.37': 1.0, 'node:10.128.16.41': 1.0} ✓ Ray cluster ready === Ray Cluster Ready === Address: 10.128.16.35:6379 Total GPUs: 56 ========================= Ray cluster ready at 10.128.16.35:6379 Total GPUs available: 56 [RLJobRunner] Pinggy check: url=False, token=False, needs_tunnel=False (agent=terminus-2, env=daytona) [RLJobRunner] No Pinggy tunnel needed, using local vLLM Running SkyRL: Python: /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/bin/python Entrypoint: examples.terminal_bench.entrypoints.main_tbench Args: 123 Hydra arguments Working dir: /e/scratch/jureap59/feuer1/OpenThoughts-Agent/SkyRL/skyrl-train Using proxychains binary: /e/scratch/jureap59/feuer1/proxychains-ng-aarch64/bin/proxychains4 Executing command with srun: /e/scratch/jureap59/feuer1/proxychains-ng-aarch64/bin/proxychains4 -f $PROXYCHAINS_CONF_FILE /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/bin/python -m examples.terminal_bench.entrypoints.main_tbench +terminal_bench_config=terminal_bench trainer.strategy=fsdp2 trainer.algorithm.advantage_estimator=rloo_n trainer.algorithm.use_kl_loss=false trainer.algorithm.kl_loss_coef=0.0 trainer.algorithm.eps_clip_low=0.2 trainer.algorithm.eps_clip_high=0.05 trainer.algorithm.loss_reduction=sequence_mean trainer.algorithm.use_tis=true trainer.algorithm.tis_imp_ratio_cap=2.0 trainer.epochs=2 trainer.max_steps=80 trainer.update_epochs_per_batch=1 trainer.train_batch_size=64 trainer.policy_mini_batch_size=64 trainer.eval_batch_size=64 trainer.micro_forward_batch_size_per_gpu=4 trainer.micro_train_batch_size_per_gpu=1 trainer.max_prompt_length=999999 trainer.eval_interval=999999 trainer.eval_before_train=false trainer.ckpt_interval=2 trainer.resume_mode=latest trainer.hf_save_interval=5 ++trainer.hf_hub_repo_id=laion/ablation-pymethods2test-seqmean-arm0-tis ++trainer.hf_hub_private=false ++trainer.hf_hub_revision=main ++trainer.enable_db_registration=false trainer.project_name=OpenThoughts-Agent trainer.log_level=INFO trainer.tracker_commit_each_step=true trainer.logger=console trainer.run_name=ablation-pymethods2test-seqmean-arm0-tis trainer.ckpt_path=/e/data1/datasets/playground/ot-baf/ablation-pymethods2test-seqmean-arm0-tis/ablation-pymethods2test-seqmean-arm0-tis/checkpoints trainer.export_path=/e/data1/datasets/playground/ot-baf/ablation-pymethods2test-seqmean-arm0-tis/ablation-pymethods2test-seqmean-arm0-tis/exports trainer.policy.optimizer_config.lr=8e-6 trainer.policy.optimizer_config.weight_decay=0.0 trainer.policy.optimizer_config.adam_betas=[0.9,0.999] trainer.policy.optimizer_config.max_grad_norm=0.9 trainer.policy.fsdp_config.cpu_offload=false trainer.policy.fsdp_config.reshard_after_forward=true trainer.policy.fsdp_config.fsdp_size=4 trainer.policy.model.path=/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6 trainer.ref.fsdp_config.cpu_offload=false trainer.ref.fsdp_config.reshard_after_forward=true trainer.ref.fsdp_config.fsdp_size=4 trainer.placement.colocate_all=false trainer.placement.policy_num_nodes=2 trainer.placement.ref_num_nodes=2 trainer.placement.policy_num_gpus_per_node=4 trainer.placement.ref_num_gpus_per_node=4 trainer.fully_async.max_staleness_steps=16 trainer.fully_async.num_parallel_generation_workers=338 generator.backend=vllm generator.timeout_multiplier=1.0 generator.model_dtype=bfloat16 generator.inference_engine_tensor_parallel_size=1 generator.num_inference_engines=48 generator.n_samples_per_prompt=8 generator.eval_n_samples_per_prompt=8 generator.gpu_memory_utilization=0.75 generator.max_num_seqs=24 generator.max_num_batched_tokens=65536 generator.enable_prefix_caching=true generator.enable_chunked_prefill=true generator.run_engines_locally=true generator.weight_sync_backend=nccl generator.async_engine=true generator.batched=false generator.enable_http_endpoint=true generator.enable_ray_prometheus_stats=false generator.vllm_stats_interval=1 generator.append_eos_token_after_stop_str_in_multi_turn=true generator.max_turns=999999 generator.sampling_params.max_generate_length=4096 generator.sampling_params.temperature=0.7 generator.sampling_params.top_p=0.95 generator.sampling_params.top_k=20 ++generator.engine_init_kwargs.max_model_len=32768 ++generator.engine_init_kwargs.custom_chat_template_chat_completion_path=chat_templates/qwen3_thinking_acc.jinja2 ++generator.engine_init_kwargs.served_model_name=0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6 data.train_data=["/e/scratch/jureap59/feuer1/tasks/exp_rpt_pymethods2test-large"] data.val_data=[] +terminal_bench_config.trials_dir=/e/data1/datasets/playground/ot-baf/ablation-pymethods2test-seqmean-arm0-tis/ablation-pymethods2test-seqmean-arm0-tis/trace_jobs +terminal_bench_config.harbor.name=terminus-2 +terminal_bench_config.harbor.max_episodes=999999 +terminal_bench_config.harbor.enable_summarize=false +terminal_bench_config.harbor.store_all_messages=true +terminal_bench_config.harbor.trajectory_config.raw_content=true +terminal_bench_config.harbor.enable_episode_logging=false +terminal_bench_config.harbor.record_terminal_session=false +terminal_bench_config.harbor.enable_pane_logging=false +terminal_bench_config.harbor.strict_json_parser=true +terminal_bench_config.harbor.interleaved_thinking=true +terminal_bench_config.harbor.extra_body.chat_template_kwargs.enable_thinking=true +terminal_bench_config.harbor.override_timeout_sec=900 +terminal_bench_config.harbor.override_cpus=1 +terminal_bench_config.harbor.override_memory_mb=2048 +terminal_bench_config.harbor.override_storage_mb=2048 +terminal_bench_config.harbor.auto_snapshot=true +terminal_bench_config.harbor.verifier_override_timeout_sec=120 +terminal_bench_config.harbor.max_retries=3 +terminal_bench_config.harbor.min_wait_sec=60.0 +terminal_bench_config.harbor.max_wait_sec=600.0 +terminal_bench_config.harbor.wait_multiplier=2.0 +terminal_bench_config.harbor.exclude_exceptions=["VerifierTimeoutError","VerifierRuntimeError","RewardFileNotFoundError","RewardFileEmptyError","VerifierOutputParseError"] +terminal_bench_config.harbor.n_concurrent_trials=675 +terminal_bench_config.harbor.log_level=INFO +terminal_bench_config.harbor.enable_reward_shaping=false +terminal_bench_config.harbor.collect_rollout_details=true +terminal_bench_config.harbor.enable_error_classification=true +terminal_bench_config.harbor.mask_exceptions=["DaytonaError","EnvironmentStartTimeoutError","NetworkError","ConnectionError","RewardFileNotFoundError","RewardFileEmptyError","AgentEnvironmentTimeoutError","ContextLengthExceededError"] +terminal_bench_config.harbor.default_error_treatment=zero +terminal_bench_config.harbor.passthrough_exceptions=["AgentTimeoutError"] +terminal_bench_config.harbor.zero_exceptions=[] +terminal_bench_config.model_info.max_input_tokens=32000 +terminal_bench_config.model_info.max_output_tokens=4096 +terminal_bench_config.archiving.enabled=false +terminal_bench_config.trace_upload.enabled=true +terminal_bench_config.trace_upload.repo_org=DCAgent +terminal_bench_config.trace_upload.episodes=last +terminal_bench_config.trace_upload.dataset_type=SFT +terminal_bench_config.trace_upload.cleanup=true [proxychains] config file found: /e/home/jusers/feuer1/jupiter/.proxychains/proxychains_653584.conf [proxychains] preloading /e/scratch/jureap59/feuer1/proxychains-ng-aarch64/lib/libproxychains4.so [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e 2026-06-07 15:37:07.080 | WARNING | skyrl_train.utils.utils:validate_cfg:479 - `generator.sampling_params.logprobs` is `None` but `trainer.algorithm.use_tis` is `True`. Setting `logprobs` to `True`. 2026-06-07 15:37:07.289 | INFO | skyrl_train.utils.utils:prepare_runtime_environment:753 - Exporting wandb api key to ray runtime env 2026-06-07 15:37:07.289 | INFO | skyrl_train.utils.utils:prepare_runtime_environment:772 - Exporting RAY_ADDRESS to ray runtime env 2026-06-07 15:37:07.289 | INFO | skyrl_train.utils.utils:prepare_runtime_environment:797 - Exporting `NCCL_SOCKET_IFNAME` to ray runtime env: ib0 2026-06-07 15:37:07.289 | INFO | skyrl_train.utils.utils:prepare_runtime_environment:797 - Exporting `NCCL_SOCKET_FAMILY` to ray runtime env: AF_INET 2026-06-07 15:37:07.289 | INFO | skyrl_train.utils.utils:prepare_runtime_environment:797 - Exporting `NCCL_DEBUG` to ray runtime env: WARN 2026-06-07 15:37:07,290 INFO worker.py:1680 -- Using address 10.128.16.35:6379 set in the environment variable RAY_ADDRESS 2026-06-07 15:37:07,324 INFO worker.py:1821 -- Connecting to existing Ray cluster at address: 10.128.16.35:6379... 2026-06-07 15:37:07,334 INFO worker.py:2007 -- Connected to Ray cluster. /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/ray/_private/worker.py:2046: FutureWarning: Tip: In future versions of Ray, Ray will no longer override accelerator visible devices env var if num_gpus=0 or num_gpus=None (default). To enable this behavior and turn off this error message, set RAY_ACCEL_ENV_VAR_OVERRIDE_ON_ZERO=0 warnings.warn( (raylet, ip=10.128.16.45) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e 2026-06-07 15:37:09.507 | INFO | skyrl_train.utils.ppo_utils:sync_registries:546 - Synced registries to ray actor [2026-06-07 15:37:37,531 E 1617107 1617929] core_worker_process.cc:842: Failed to establish connection to the metrics exporter agent. Metrics will not be exported. Exporter agent status: RpcError: Running out of retries to initialize the metrics agent. rpc_code: 14 (raylet, ip=10.128.16.45) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e [repeated 5x across cluster] (Ray deduplicates logs by default. Set RAY_DEDUP_LOGS=0 to disable log deduplication, or see https://docs.ray.io/en/master/ray-observability/user-guides/configure-logging.html#log-deduplication for more options.) (skyrl_entrypoint pid=1617955) 2026-06-07 15:37:18.206 | INFO  | skyrl_train.entrypoints.main_base:_configure_log_level:212 - SkyRL log level set to: INFO (skyrl_entrypoint pid=1617955) The tokenizer you are loading from '/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6' with an incorrect regex pattern: https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503/discussions/84#69121093e8b480e709447d5e. This will lead to incorrect tokenization. You should set the `fix_mistral_regex=True` flag when loading this tokenizer to fix this issue. (skyrl_entrypoint pid=1617955) 2026-06-07 15:37:18.546 | INFO  | examples.terminal_bench.dataset:_load_data_files:40 - Loading data from: /e/scratch/jureap59/feuer1/tasks/exp_rpt_pymethods2test-large (skyrl_entrypoint pid=1617955) 2026-06-07 15:37:20.166 | INFO  | examples.terminal_bench.dataset:_load_data_files:50 - Found 5000 valid task directories out of 5000 total directories (skyrl_entrypoint pid=1617955) 2026-06-07 15:37:20.166 | INFO  | examples.terminal_bench.dataset:__init__:27 - TerminalBenchTaskDataset initialized with 5000 task paths (skyrl_entrypoint pid=1617955) 2026-06-07 15:37:20.178 | INFO  | skyrl_train.entrypoints.main_base:_setup_trainer:405 - data: (skyrl_entrypoint pid=1617955) train_data: (skyrl_entrypoint pid=1617955) - /e/scratch/jureap59/feuer1/tasks/exp_rpt_pymethods2test-large (skyrl_entrypoint pid=1617955) val_data: [] (skyrl_entrypoint pid=1617955) trainer: (skyrl_entrypoint pid=1617955) placement: (skyrl_entrypoint pid=1617955) colocate_all: false (skyrl_entrypoint pid=1617955) colocate_policy_ref: true (skyrl_entrypoint pid=1617955) policy_num_nodes: 2 (skyrl_entrypoint pid=1617955) policy_num_gpus_per_node: 4 (skyrl_entrypoint pid=1617955) critic_num_nodes: 1 (skyrl_entrypoint pid=1617955) critic_num_gpus_per_node: 4 (skyrl_entrypoint pid=1617955) ref_num_nodes: 2 (skyrl_entrypoint pid=1617955) ref_num_gpus_per_node: 4 (skyrl_entrypoint pid=1617955) policy_strict_spread_pg: false (skyrl_entrypoint pid=1617955) policy_per_gpu_bundles: false (skyrl_entrypoint pid=1617955) policy_force_cvd_mask: false (skyrl_entrypoint pid=1617955) sequence_parallel_backend: ulysses (skyrl_entrypoint pid=1617955) strategy: fsdp2 (skyrl_entrypoint pid=1617955) policy: (skyrl_entrypoint pid=1617955) model: (skyrl_entrypoint pid=1617955) path: /e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6 (skyrl_entrypoint pid=1617955) lora: (skyrl_entrypoint pid=1617955) rank: 0 (skyrl_entrypoint pid=1617955) alpha: 16 (skyrl_entrypoint pid=1617955) dropout: 0 (skyrl_entrypoint pid=1617955) lora_sync_path: /tmp/skyrl_lora_sync (skyrl_entrypoint pid=1617955) target_modules: all-linear (skyrl_entrypoint pid=1617955) exclude_modules: null (skyrl_entrypoint pid=1617955) deepspeed_config: ${deepspeed_config.train} (skyrl_entrypoint pid=1617955) optimizer_config: (skyrl_entrypoint pid=1617955) optimizer: AdamW (skyrl_entrypoint pid=1617955) lr: 8.0e-06 (skyrl_entrypoint pid=1617955) adam_betas: (skyrl_entrypoint pid=1617955) - 0.9 (skyrl_entrypoint pid=1617955) - 0.999 (skyrl_entrypoint pid=1617955) weight_decay: 0.0 (skyrl_entrypoint pid=1617955) max_grad_norm: 0.9 (skyrl_entrypoint pid=1617955) offload_after_step: true (skyrl_entrypoint pid=1617955) num_warmup_steps: 0 (skyrl_entrypoint pid=1617955) scheduler: constant_with_warmup (skyrl_entrypoint pid=1617955) optimizer_kwargs: {} (skyrl_entrypoint pid=1617955) fsdp_config: (skyrl_entrypoint pid=1617955) cpu_offload: false (skyrl_entrypoint pid=1617955) reshard_after_forward: true (skyrl_entrypoint pid=1617955) fsdp_size: 4 (skyrl_entrypoint pid=1617955) expert_model_parallel_size: 1 (skyrl_entrypoint pid=1617955) expert_tensor_parallel_size: 1 (skyrl_entrypoint pid=1617955) moe_token_dispatcher_type: alltoall (skyrl_entrypoint pid=1617955) moe_router_replay: false (skyrl_entrypoint pid=1617955) moe_grouped_gemm: false (skyrl_entrypoint pid=1617955) ep_comm_backend: torch (skyrl_entrypoint pid=1617955) deepep_num_sms: 20 (skyrl_entrypoint pid=1617955) deepep_token_chunk_size: null (skyrl_entrypoint pid=1617955) sequence_parallel_size: 1 (skyrl_entrypoint pid=1617955) use_torch_compile: false (skyrl_entrypoint pid=1617955) record_memory: false (skyrl_entrypoint pid=1617955) megatron_config: (skyrl_entrypoint pid=1617955) tensor_model_parallel_size: 1 (skyrl_entrypoint pid=1617955) pipeline_model_parallel_size: 1 (skyrl_entrypoint pid=1617955) context_parallel_size: 1 (skyrl_entrypoint pid=1617955) expert_model_parallel_size: 1 (skyrl_entrypoint pid=1617955) expert_tensor_parallel_size: null (skyrl_entrypoint pid=1617955) ddp_config: (skyrl_entrypoint pid=1617955) grad_reduce_in_fp32: true (skyrl_entrypoint pid=1617955) overlap_grad_reduce: false (skyrl_entrypoint pid=1617955) overlap_param_gather: false (skyrl_entrypoint pid=1617955) average_in_collective: true (skyrl_entrypoint pid=1617955) model_config_kwargs: {} (skyrl_entrypoint pid=1617955) torch_profiler_config: (skyrl_entrypoint pid=1617955) enable: false (skyrl_entrypoint pid=1617955) ranks: [] (skyrl_entrypoint pid=1617955) save_path: null (skyrl_entrypoint pid=1617955) optimizer_config_kwargs: (skyrl_entrypoint pid=1617955) overlap_cpu_optimizer_d2h_h2d: false (skyrl_entrypoint pid=1617955) use_precision_aware_optimizer: false (skyrl_entrypoint pid=1617955) optimizer_cpu_offload: false (skyrl_entrypoint pid=1617955) optimizer_offload_fraction: 0.0 (skyrl_entrypoint pid=1617955) transformer_config_kwargs: (skyrl_entrypoint pid=1617955) recompute_granularity: full (skyrl_entrypoint pid=1617955) recompute_modules: (skyrl_entrypoint pid=1617955) - core_attn (skyrl_entrypoint pid=1617955) recompute_method: uniform (skyrl_entrypoint pid=1617955) recompute_num_layers: 1 (skyrl_entrypoint pid=1617955) empty_cuda_cache: true (skyrl_entrypoint pid=1617955) ref: (skyrl_entrypoint pid=1617955) model: (skyrl_entrypoint pid=1617955) path: ${trainer.policy.model.path} (skyrl_entrypoint pid=1617955) sequence_parallel_size: 1 (skyrl_entrypoint pid=1617955) deepspeed_config: ${deepspeed_config.eval} (skyrl_entrypoint pid=1617955) fsdp_config: (skyrl_entrypoint pid=1617955) cpu_offload: false (skyrl_entrypoint pid=1617955) reshard_after_forward: true (skyrl_entrypoint pid=1617955) fsdp_size: 4 (skyrl_entrypoint pid=1617955) expert_model_parallel_size: 1 (skyrl_entrypoint pid=1617955) expert_tensor_parallel_size: 1 (skyrl_entrypoint pid=1617955) moe_token_dispatcher_type: alltoall (skyrl_entrypoint pid=1617955) moe_router_replay: false (skyrl_entrypoint pid=1617955) moe_grouped_gemm: false (skyrl_entrypoint pid=1617955) ep_comm_backend: torch (skyrl_entrypoint pid=1617955) deepep_num_sms: 20 (skyrl_entrypoint pid=1617955) deepep_token_chunk_size: null (skyrl_entrypoint pid=1617955) megatron_config: (skyrl_entrypoint pid=1617955) tensor_model_parallel_size: 1 (skyrl_entrypoint pid=1617955) pipeline_model_parallel_size: 1 (skyrl_entrypoint pid=1617955) context_parallel_size: 1 (skyrl_entrypoint pid=1617955) expert_model_parallel_size: 1 (skyrl_entrypoint pid=1617955) expert_tensor_parallel_size: 1 (skyrl_entrypoint pid=1617955) model_config_kwargs: {} (skyrl_entrypoint pid=1617955) transformer_config_kwargs: {} (skyrl_entrypoint pid=1617955) critic: (skyrl_entrypoint pid=1617955) model: (skyrl_entrypoint pid=1617955) path: null (skyrl_entrypoint pid=1617955) lora: (skyrl_entrypoint pid=1617955) rank: 0 (skyrl_entrypoint pid=1617955) alpha: 16 (skyrl_entrypoint pid=1617955) dropout: 0 (skyrl_entrypoint pid=1617955) target_modules: all-linear (skyrl_entrypoint pid=1617955) exclude_modules: null (skyrl_entrypoint pid=1617955) deepspeed_config: ${deepspeed_config.train} (skyrl_entrypoint pid=1617955) optimizer_config: (skyrl_entrypoint pid=1617955) optimizer: AdamW (skyrl_entrypoint pid=1617955) lr: 5.0e-06 (skyrl_entrypoint pid=1617955) adam_betas: (skyrl_entrypoint pid=1617955) - 0.9 (skyrl_entrypoint pid=1617955) - 0.999 (skyrl_entrypoint pid=1617955) weight_decay: 0.01 (skyrl_entrypoint pid=1617955) max_grad_norm: 1.0 (skyrl_entrypoint pid=1617955) offload_after_step: true (skyrl_entrypoint pid=1617955) num_warmup_steps: 0 (skyrl_entrypoint pid=1617955) scheduler: constant_with_warmup (skyrl_entrypoint pid=1617955) optimizer_kwargs: {} (skyrl_entrypoint pid=1617955) fsdp_config: (skyrl_entrypoint pid=1617955) cpu_offload: false (skyrl_entrypoint pid=1617955) reshard_after_forward: true (skyrl_entrypoint pid=1617955) fsdp_size: -1 (skyrl_entrypoint pid=1617955) expert_model_parallel_size: 1 (skyrl_entrypoint pid=1617955) expert_tensor_parallel_size: 1 (skyrl_entrypoint pid=1617955) moe_token_dispatcher_type: alltoall (skyrl_entrypoint pid=1617955) moe_router_replay: false (skyrl_entrypoint pid=1617955) moe_grouped_gemm: false (skyrl_entrypoint pid=1617955) ep_comm_backend: torch (skyrl_entrypoint pid=1617955) deepep_num_sms: 20 (skyrl_entrypoint pid=1617955) deepep_token_chunk_size: null (skyrl_entrypoint pid=1617955) sequence_parallel_size: 1 (skyrl_entrypoint pid=1617955) algorithm: (skyrl_entrypoint pid=1617955) advantage_estimator: rloo_n (skyrl_entrypoint pid=1617955) kl_ctrl: (skyrl_entrypoint pid=1617955) type: fixed (skyrl_entrypoint pid=1617955) kl_target: 0.1 (skyrl_entrypoint pid=1617955) horizon: 10000 (skyrl_entrypoint pid=1617955) kl_estimator_type: k3 (skyrl_entrypoint pid=1617955) use_kl_estimator_k3: false (skyrl_entrypoint pid=1617955) use_abs_kl: false (skyrl_entrypoint pid=1617955) use_kl_in_reward: false (skyrl_entrypoint pid=1617955) use_kl_loss: false (skyrl_entrypoint pid=1617955) kl_loss_coef: 0.0 (skyrl_entrypoint pid=1617955) use_entropy_loss: false (skyrl_entrypoint pid=1617955) entropy_loss_coef: 0.01 (skyrl_entrypoint pid=1617955) advantage_batch_normalize: false (skyrl_entrypoint pid=1617955) value_head_prefix: value_head (skyrl_entrypoint pid=1617955) policy_loss_type: regular (skyrl_entrypoint pid=1617955) loss_reduction: sequence_mean (skyrl_entrypoint pid=1617955) global_loss_denom: null (skyrl_entrypoint pid=1617955) grpo_norm_by_std: true (skyrl_entrypoint pid=1617955) rloo_n_min_group_size: 4 (skyrl_entrypoint pid=1617955) rloo_n_filter_zero_reward_groups: true (skyrl_entrypoint pid=1617955) lambd: 1.0 (skyrl_entrypoint pid=1617955) gamma: 1.0 (skyrl_entrypoint pid=1617955) eps_clip_low: 0.2 (skyrl_entrypoint pid=1617955) eps_clip_high: 0.05 (skyrl_entrypoint pid=1617955) clip_ratio_c: 3.0 (skyrl_entrypoint pid=1617955) tis_imp_ratio_cap: 2.0 (skyrl_entrypoint pid=1617955) use_tis: true (skyrl_entrypoint pid=1617955) sapo: (skyrl_entrypoint pid=1617955) tau_pos: 1.0 (skyrl_entrypoint pid=1617955) tau_neg: 1.05 (skyrl_entrypoint pid=1617955) value_clip: 0.2 (skyrl_entrypoint pid=1617955) dynamic_sampling: (skyrl_entrypoint pid=1617955) type: null (skyrl_entrypoint pid=1617955) max_sample_batches: 30 (skyrl_entrypoint pid=1617955) min_replace_ratio: 0.3 (skyrl_entrypoint pid=1617955) clip_cov: (skyrl_entrypoint pid=1617955) clip_ratio: 0.0002 (skyrl_entrypoint pid=1617955) clip_cov_lb: 1.0 (skyrl_entrypoint pid=1617955) clip_cov_ub: 5.0 (skyrl_entrypoint pid=1617955) kl_cov: (skyrl_entrypoint pid=1617955) kl_cov_frac: 0.2 (skyrl_entrypoint pid=1617955) ppo_kl_coef: 1.0 (skyrl_entrypoint pid=1617955) cispo: (skyrl_entrypoint pid=1617955) cispo_eps_clip_low: 0 (skyrl_entrypoint pid=1617955) cispo_eps_clip_high: 5 (skyrl_entrypoint pid=1617955) z_clip: (skyrl_entrypoint pid=1617955) enabled: false (skyrl_entrypoint pid=1617955) alpha: 0.97 (skyrl_entrypoint pid=1617955) z_thresh: 2.5 (skyrl_entrypoint pid=1617955) warmup_steps: 25 (skyrl_entrypoint pid=1617955) mode: zscore (skyrl_entrypoint pid=1617955) clip_option: adaptive_scaling (skyrl_entrypoint pid=1617955) clip_factor: 1.0 (skyrl_entrypoint pid=1617955) skip_update_on_spike: false (skyrl_entrypoint pid=1617955) stale_clip: (skyrl_entrypoint pid=1617955) enabled: false (skyrl_entrypoint pid=1617955) alpha: 0.3 (skyrl_entrypoint pid=1617955) entropy_threshold: 0.15 (skyrl_entrypoint pid=1617955) entropy_window: 10 (skyrl_entrypoint pid=1617955) min_lr_scale: 0.1 (skyrl_entrypoint pid=1617955) max_seq_len: 1004095 (skyrl_entrypoint pid=1617955) fully_async: (skyrl_entrypoint pid=1617955) max_staleness_steps: 16 (skyrl_entrypoint pid=1617955) num_parallel_generation_workers: 338 (skyrl_entrypoint pid=1617955) gradient_checkpointing: true (skyrl_entrypoint pid=1617955) gradient_checkpointing_use_reentrant: false (skyrl_entrypoint pid=1617955) seed: 42 (skyrl_entrypoint pid=1617955) resume_mode: latest (skyrl_entrypoint pid=1617955) resume_path: null (skyrl_entrypoint pid=1617955) ckpt_path: /e/data1/datasets/playground/ot-baf/ablation-pymethods2test-seqmean-arm0-tis/ablation-pymethods2test-seqmean-arm0-tis/checkpoints (skyrl_entrypoint pid=1617955) max_ckpts_to_keep: -1 (skyrl_entrypoint pid=1617955) ckpt_interval: 2 (skyrl_entrypoint pid=1617955) hf_save_interval: 5 (skyrl_entrypoint pid=1617955) hf_upload_mode: latest (skyrl_entrypoint pid=1617955) export_path: /e/data1/datasets/playground/ot-baf/ablation-pymethods2test-seqmean-arm0-tis/ablation-pymethods2test-seqmean-arm0-tis/exports (skyrl_entrypoint pid=1617955) bf16: true (skyrl_entrypoint pid=1617955) epochs: 2 (skyrl_entrypoint pid=1617955) max_steps: 80 (skyrl_entrypoint pid=1617955) update_epochs_per_batch: 1 (skyrl_entrypoint pid=1617955) train_batch_size: 64 (skyrl_entrypoint pid=1617955) policy_mini_batch_size: 64 (skyrl_entrypoint pid=1617955) critic_mini_batch_size: 256 (skyrl_entrypoint pid=1617955) micro_train_batch_size_per_gpu: 1 (skyrl_entrypoint pid=1617955) micro_forward_batch_size_per_gpu: 4 (skyrl_entrypoint pid=1617955) update_ref_every_epoch: false (skyrl_entrypoint pid=1617955) use_sample_packing: true (skyrl_entrypoint pid=1617955) eval_batch_size: 64 (skyrl_entrypoint pid=1617955) eval_before_train: false (skyrl_entrypoint pid=1617955) eval_interval: 999999 (skyrl_entrypoint pid=1617955) max_prompt_length: 999999 (skyrl_entrypoint pid=1617955) flash_attn: true (skyrl_entrypoint pid=1617955) disable_fast_tokenizer: false (skyrl_entrypoint pid=1617955) target_modules: null (skyrl_entrypoint pid=1617955) exclude_modules: null (skyrl_entrypoint pid=1617955) project_name: OpenThoughts-Agent (skyrl_entrypoint pid=1617955) run_name: ablation-pymethods2test-seqmean-arm0-tis (skyrl_entrypoint pid=1617955) logger: console (skyrl_entrypoint pid=1617955) tracker_commit_each_step: true (skyrl_entrypoint pid=1617955) dump_data_batch: false (skyrl_entrypoint pid=1617955) dump_eval_results: true (skyrl_entrypoint pid=1617955) log_level: INFO (skyrl_entrypoint pid=1617955) rope_scaling: null (skyrl_entrypoint pid=1617955) rope_theta: null (skyrl_entrypoint pid=1617955) step_wise_training: false (skyrl_entrypoint pid=1617955) hf_hub_repo_id: laion/ablation-pymethods2test-seqmean-arm0-tis (skyrl_entrypoint pid=1617955) hf_hub_private: false (skyrl_entrypoint pid=1617955) hf_hub_revision: main (skyrl_entrypoint pid=1617955) enable_db_registration: false (skyrl_entrypoint pid=1617955) generator: (skyrl_entrypoint pid=1617955) model_name: ${trainer.policy.model.path} (skyrl_entrypoint pid=1617955) model_dtype: bfloat16 (skyrl_entrypoint pid=1617955) timeout_multiplier: 1.0 (skyrl_entrypoint pid=1617955) run_engines_locally: true (skyrl_entrypoint pid=1617955) num_inference_engines: 48 (skyrl_entrypoint pid=1617955) backend: vllm (skyrl_entrypoint pid=1617955) weight_sync_backend: nccl (skyrl_entrypoint pid=1617955) fuse_weights: false (skyrl_entrypoint pid=1617955) weight_transfer_threshold_cuda_ipc_GB: 1.0 (skyrl_entrypoint pid=1617955) inference_engine_tensor_parallel_size: 1 (skyrl_entrypoint pid=1617955) inference_engine_pipeline_parallel_size: 1 (skyrl_entrypoint pid=1617955) inference_engine_expert_parallel_size: 1 (skyrl_entrypoint pid=1617955) inference_engine_data_parallel_size: 1 (skyrl_entrypoint pid=1617955) n_samples_per_prompt: 8 (skyrl_entrypoint pid=1617955) async_engine: true (skyrl_entrypoint pid=1617955) batched: false (skyrl_entrypoint pid=1617955) max_input_length: ${trainer.max_prompt_length} (skyrl_entrypoint pid=1617955) vllm_v1_disable_multiproc: true (skyrl_entrypoint pid=1617955) enable_prefix_caching: true (skyrl_entrypoint pid=1617955) enable_chunked_prefill: true (skyrl_entrypoint pid=1617955) max_num_batched_tokens: 65536 (skyrl_entrypoint pid=1617955) enforce_eager: true (skyrl_entrypoint pid=1617955) fully_sharded_loras: false (skyrl_entrypoint pid=1617955) enable_ray_prometheus_stats: false (skyrl_entrypoint pid=1617955) vllm_stats_interval: 1 (skyrl_entrypoint pid=1617955) gpu_memory_utilization: 0.75 (skyrl_entrypoint pid=1617955) max_num_seqs: 24 (skyrl_entrypoint pid=1617955) remote_inference_engine_urls: (skyrl_entrypoint pid=1617955) - 127.0.0.1:8001 (skyrl_entrypoint pid=1617955) enable_http_endpoint: true (skyrl_entrypoint pid=1617955) http_endpoint_host: 127.0.0.1 (skyrl_entrypoint pid=1617955) http_endpoint_port: 8000 (skyrl_entrypoint pid=1617955) max_turns: 999999 (skyrl_entrypoint pid=1617955) chat_template: (skyrl_entrypoint pid=1617955) source: name (skyrl_entrypoint pid=1617955) name_or_path: null (skyrl_entrypoint pid=1617955) chat_template_kwargs: {} (skyrl_entrypoint pid=1617955) engine_init_kwargs: (skyrl_entrypoint pid=1617955) max_model_len: 32768 (skyrl_entrypoint pid=1617955) custom_chat_template_chat_completion_path: chat_templates/qwen3_thinking_acc.jinja2 (skyrl_entrypoint pid=1617955) served_model_name: 0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6 (skyrl_entrypoint pid=1617955) override_existing_update_group: disable (skyrl_entrypoint pid=1617955) sampling_params: (skyrl_entrypoint pid=1617955) max_generate_length: 4096 (skyrl_entrypoint pid=1617955) repetition_penalty: 1.0 (skyrl_entrypoint pid=1617955) temperature: 0.7 (skyrl_entrypoint pid=1617955) top_p: 0.95 (skyrl_entrypoint pid=1617955) min_p: 0.0 (skyrl_entrypoint pid=1617955) top_k: 20 (skyrl_entrypoint pid=1617955) logprobs: 0 (skyrl_entrypoint pid=1617955) stop: null (skyrl_entrypoint pid=1617955) use_conversation_multi_turn: true (skyrl_entrypoint pid=1617955) append_eos_token_after_stop_str_in_multi_turn: true (skyrl_entrypoint pid=1617955) eval_sampling_params: (skyrl_entrypoint pid=1617955) max_generate_length: ${generator.sampling_params.max_generate_length} (skyrl_entrypoint pid=1617955) repetition_penalty: 1.0 (skyrl_entrypoint pid=1617955) temperature: 0.0 (skyrl_entrypoint pid=1617955) top_p: 1.0 (skyrl_entrypoint pid=1617955) min_p: 0.0 (skyrl_entrypoint pid=1617955) top_k: -1 (skyrl_entrypoint pid=1617955) logprobs: null (skyrl_entrypoint pid=1617955) stop: null (skyrl_entrypoint pid=1617955) eval_n_samples_per_prompt: 8 (skyrl_entrypoint pid=1617955) zero_reward_on_non_stop: false (skyrl_entrypoint pid=1617955) apply_overlong_filtering: false (skyrl_entrypoint pid=1617955) rope_scaling: ${trainer.rope_scaling} (skyrl_entrypoint pid=1617955) rope_theta: ${trainer.rope_theta} (skyrl_entrypoint pid=1617955) teacher: (skyrl_entrypoint pid=1617955) model_path: null (skyrl_entrypoint pid=1617955) top_k_logprobs: 256 (skyrl_entrypoint pid=1617955) num_inference_engines: 1 (skyrl_entrypoint pid=1617955) inference_engine_tensor_parallel_size: 1 (skyrl_entrypoint pid=1617955) inference_engine_pipeline_parallel_size: 1 (skyrl_entrypoint pid=1617955) gpu_memory_utilization: 0.9 (skyrl_entrypoint pid=1617955) enforce_eager: false (skyrl_entrypoint pid=1617955) backend: vllm (skyrl_entrypoint pid=1617955) engine_init_kwargs: {} (skyrl_entrypoint pid=1617955) environment: (skyrl_entrypoint pid=1617955) env_class: gsm8k (skyrl_entrypoint pid=1617955) skyrl_gym: (skyrl_entrypoint pid=1617955) max_env_workers: 32 (skyrl_entrypoint pid=1617955) text2sql: (skyrl_entrypoint pid=1617955) db_path: /home/ray/default/sql_data (skyrl_entrypoint pid=1617955) llm_as_a_judge: (skyrl_entrypoint pid=1617955) model: gpt-4o-mini (skyrl_entrypoint pid=1617955) base_url: null (skyrl_entrypoint pid=1617955) search: (skyrl_entrypoint pid=1617955) log_requests: false (skyrl_entrypoint pid=1617955) search_url: http://127.0.0.1:8000/retrieve (skyrl_entrypoint pid=1617955) topk: 3 (skyrl_entrypoint pid=1617955) timeout: 30 (skyrl_entrypoint pid=1617955) rollout: (skyrl_entrypoint pid=1617955) fanout: (skyrl_entrypoint pid=1617955) enabled: true (skyrl_entrypoint pid=1617955) num_coordinators: 4 (skyrl_entrypoint pid=1617955) cpus_per_coordinator: 8 (skyrl_entrypoint pid=1617955) deepspeed_config: (skyrl_entrypoint pid=1617955) train: (skyrl_entrypoint pid=1617955) zero_optimization: (skyrl_entrypoint pid=1617955) stage: 3 (skyrl_entrypoint pid=1617955) offload_param: (skyrl_entrypoint pid=1617955) device: none (skyrl_entrypoint pid=1617955) offload_optimizer: (skyrl_entrypoint pid=1617955) device: none (skyrl_entrypoint pid=1617955) pin_memory: true (skyrl_entrypoint pid=1617955) sub_group_size: auto (skyrl_entrypoint pid=1617955) reduce_bucket_size: auto (skyrl_entrypoint pid=1617955) stage3_param_persistence_threshold: auto (skyrl_entrypoint pid=1617955) stage3_prefetch_bucket_size: auto (skyrl_entrypoint pid=1617955) stage3_max_live_parameters: auto (skyrl_entrypoint pid=1617955) stage3_max_reuse_distance: auto (skyrl_entrypoint pid=1617955) round_robin_gradients: true (skyrl_entrypoint pid=1617955) zero_hpz_partition_size: 1 (skyrl_entrypoint pid=1617955) zero_quantized_weights: false (skyrl_entrypoint pid=1617955) zero_quantized_gradients: false (skyrl_entrypoint pid=1617955) torch_autocast: (skyrl_entrypoint pid=1617955) enabled: true (skyrl_entrypoint pid=1617955) dtype: bfloat16 (skyrl_entrypoint pid=1617955) disable_trace_cache: false (skyrl_entrypoint pid=1617955) data_types: (skyrl_entrypoint pid=1617955) grad_accum_dtype: fp32 (skyrl_entrypoint pid=1617955) gradient_clipping: 1.0 (skyrl_entrypoint pid=1617955) wall_clock_breakdown: false (skyrl_entrypoint pid=1617955) prescale_gradient: false (skyrl_entrypoint pid=1617955) eval: (skyrl_entrypoint pid=1617955) zero_optimization: (skyrl_entrypoint pid=1617955) stage: 3 (skyrl_entrypoint pid=1617955) stage3_param_persistence_threshold: auto (skyrl_entrypoint pid=1617955) offload_param: (skyrl_entrypoint pid=1617955) device: cpu (skyrl_entrypoint pid=1617955) pin_memory: true (skyrl_entrypoint pid=1617955) torch_autocast: (skyrl_entrypoint pid=1617955) enabled: true (skyrl_entrypoint pid=1617955) dtype: bfloat16 (skyrl_entrypoint pid=1617955) gradient_clipping: 1.0 (skyrl_entrypoint pid=1617955) prescale_gradient: false (skyrl_entrypoint pid=1617955) wall_clock_breakdown: false (skyrl_entrypoint pid=1617955) terminal_bench_config: (skyrl_entrypoint pid=1617955) trials_dir: /e/data1/datasets/playground/ot-baf/ablation-pymethods2test-seqmean-arm0-tis/ablation-pymethods2test-seqmean-arm0-tis/trace_jobs (skyrl_entrypoint pid=1617955) harbor: (skyrl_entrypoint pid=1617955) name: terminus-2 (skyrl_entrypoint pid=1617955) max_episodes: 999999 (skyrl_entrypoint pid=1617955) enable_summarize: false (skyrl_entrypoint pid=1617955) store_all_messages: true (skyrl_entrypoint pid=1617955) trajectory_config: (skyrl_entrypoint pid=1617955) raw_content: true (skyrl_entrypoint pid=1617955) enable_episode_logging: false (skyrl_entrypoint pid=1617955) record_terminal_session: false (skyrl_entrypoint pid=1617955) enable_pane_logging: false (skyrl_entrypoint pid=1617955) strict_json_parser: true (skyrl_entrypoint pid=1617955) interleaved_thinking: true (skyrl_entrypoint pid=1617955) extra_body: (skyrl_entrypoint pid=1617955) chat_template_kwargs: (skyrl_entrypoint pid=1617955) enable_thinking: true (skyrl_entrypoint pid=1617955) override_timeout_sec: 900 (skyrl_entrypoint pid=1617955) override_cpus: 1 (skyrl_entrypoint pid=1617955) override_memory_mb: 2048 (skyrl_entrypoint pid=1617955) override_storage_mb: 2048 (skyrl_entrypoint pid=1617955) auto_snapshot: true (skyrl_entrypoint pid=1617955) verifier_override_timeout_sec: 120 (skyrl_entrypoint pid=1617955) max_retries: 3 (skyrl_entrypoint pid=1617955) min_wait_sec: 60.0 (skyrl_entrypoint pid=1617955) max_wait_sec: 600.0 (skyrl_entrypoint pid=1617955) wait_multiplier: 2.0 (skyrl_entrypoint pid=1617955) exclude_exceptions: (skyrl_entrypoint pid=1617955) - VerifierTimeoutError (skyrl_entrypoint pid=1617955) - VerifierRuntimeError (skyrl_entrypoint pid=1617955) - RewardFileNotFoundError (skyrl_entrypoint pid=1617955) - RewardFileEmptyError (skyrl_entrypoint pid=1617955) - VerifierOutputParseError (skyrl_entrypoint pid=1617955) n_concurrent_trials: 675 (skyrl_entrypoint pid=1617955) log_level: INFO (skyrl_entrypoint pid=1617955) enable_reward_shaping: false (skyrl_entrypoint pid=1617955) collect_rollout_details: true (skyrl_entrypoint pid=1617955) enable_error_classification: true (skyrl_entrypoint pid=1617955) mask_exceptions: (skyrl_entrypoint pid=1617955) - DaytonaError (skyrl_entrypoint pid=1617955) - EnvironmentStartTimeoutError (skyrl_entrypoint pid=1617955) - NetworkError (skyrl_entrypoint pid=1617955) - ConnectionError (skyrl_entrypoint pid=1617955) - RewardFileNotFoundError (skyrl_entrypoint pid=1617955) - RewardFileEmptyError (skyrl_entrypoint pid=1617955) - AgentEnvironmentTimeoutError (skyrl_entrypoint pid=1617955) - ContextLengthExceededError (skyrl_entrypoint pid=1617955) default_error_treatment: zero (skyrl_entrypoint pid=1617955) passthrough_exceptions: (skyrl_entrypoint pid=1617955) - AgentTimeoutError (skyrl_entrypoint pid=1617955) zero_exceptions: [] (skyrl_entrypoint pid=1617955) model_info: (skyrl_entrypoint pid=1617955) max_input_tokens: 32000 (skyrl_entrypoint pid=1617955) max_output_tokens: 4096 (skyrl_entrypoint pid=1617955) archiving: (skyrl_entrypoint pid=1617955) enabled: false (skyrl_entrypoint pid=1617955) trace_upload: (skyrl_entrypoint pid=1617955) enabled: true (skyrl_entrypoint pid=1617955) repo_org: DCAgent (skyrl_entrypoint pid=1617955) episodes: last (skyrl_entrypoint pid=1617955) dataset_type: SFT (skyrl_entrypoint pid=1617955) cleanup: true (skyrl_entrypoint pid=1617955)  (RegistryActor pid=1354131, ip=10.128.16.45) [2026-06-07 15:37:38,501 E 1354131 1354171] core_worker_process.cc:842: Failed to establish connection to the metrics exporter agent. Metrics will not be exported. Exporter agent status: RpcError: Running out of retries to initialize the metrics agent. rpc_code: 14 (skyrl_entrypoint pid=1617955) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/vllm/__init__.py:7: RuntimeWarning: Failed to read commit hash: (skyrl_entrypoint pid=1617955) No module named 'vllm._version' (skyrl_entrypoint pid=1617955) from .version import __version__, __version_tuple__ # isort:skip (raylet) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e [repeated 12x across cluster] (skyrl_entrypoint pid=1617955) [2026-06-07 15:37:40,589 E 1617955 1617998] core_worker_process.cc:842: Failed to establish connection to the metrics exporter agent. Metrics will not be exported. Exporter agent status: RpcError: Running out of retries to initialize the metrics agent. rpc_code: 14 [repeated 2x across cluster] (skyrl_entrypoint pid=1617955) W0607 15:42:21.650000 1617955 envs/rl/lib/python3.12/site-packages/torch/utils/cpp_extension.py:117] No CUDA runtime is found, using CUDA_HOME='/e/software/default/stages/2026/software/CUDA/13' (raylet) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e (get_all_env_variables pid=1618274) [2026-06-07 15:44:18,568 E 1618274 1618314] core_worker_process.cc:842: Failed to establish connection to the metrics exporter agent. Metrics will not be exported. Exporter agent status: RpcError: Running out of retries to initialize the metrics agent. rpc_code: 14 (raylet, ip=10.128.16.45) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e [repeated 29x across cluster] (pid=1354512, ip=10.128.16.45) [2026-06-07 15:44:20,995 E 1354512 1354615] core_worker_process.cc:842: Failed to establish connection to the metrics exporter agent. Metrics will not be exported. Exporter agent status: RpcError: Running out of retries to initialize the metrics agent. rpc_code: 14 [repeated 4x across cluster] (raylet, ip=10.128.16.42) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e [repeated 2x across cluster] (raylet, ip=10.128.16.42) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e [repeated 10x across cluster] (pid=1354512, ip=10.128.16.45) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/vllm/__init__.py:7: RuntimeWarning: Failed to read commit hash: (pid=1354512, ip=10.128.16.45) No module named 'vllm._version' (pid=1354512, ip=10.128.16.45) from .version import __version__, __version_tuple__ # isort:skip (get_addr_port pid=1516023, ip=10.128.16.42) [2026-06-07 15:45:41,920 E 1516023 1516098] core_worker_process.cc:842: Failed to establish connection to the metrics exporter agent. Metrics will not be exported. Exporter agent status: RpcError: Running out of retries to initialize the metrics agent. rpc_code: 14 (raylet, ip=10.128.16.42) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e [repeated 2x across cluster] (pid=1516024, ip=10.128.16.42) [2026-06-07 15:45:41,966 E 1516024 1516126] core_worker_process.cc:842: Failed to establish connection to the metrics exporter agent. Metrics will not be exported. Exporter agent status: RpcError: Running out of retries to initialize the metrics agent. rpc_code: 14 (pid=1516170, ip=10.128.16.42) [2026-06-07 15:47:31,328 E 1516170 1516245] core_worker_process.cc:842: Failed to establish connection to the metrics exporter agent. Metrics will not be exported. Exporter agent status: RpcError: Running out of retries to initialize the metrics agent. rpc_code: 14 (raylet, ip=10.128.16.37) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e [repeated 22x across cluster] (pid=1496715, ip=10.128.16.37) [2026-06-07 15:47:31,482 E 1496715 1496817] core_worker_process.cc:842: Failed to establish connection to the metrics exporter agent. Metrics will not be exported. Exporter agent status: RpcError: Running out of retries to initialize the metrics agent. rpc_code: 14 [repeated 3x across cluster] (raylet, ip=10.128.16.37) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e [repeated 3x across cluster] (pid=1497002, ip=10.128.16.37) [2026-06-07 15:52:22,486 E 1497002 1497139] core_worker_process.cc:842: Failed to establish connection to the metrics exporter agent. Metrics will not be exported. Exporter agent status: RpcError: Running out of retries to initialize the metrics agent. rpc_code: 14 (raylet, ip=10.128.16.37) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e [repeated 15x across cluster] (pid=1497001, ip=10.128.16.37) [2026-06-07 15:52:22,439 E 1497001 1497111] core_worker_process.cc:842: Failed to establish connection to the metrics exporter agent. Metrics will not be exported. Exporter agent status: RpcError: Running out of retries to initialize the metrics agent. rpc_code: 14 (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) 2026-06-07 15:53:25.238 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:152 - setup_envvars_for_vllm: distributed_executor_backend=uni, SKYRL_ENABLE_NUMA_AFFINITY=1, CUDA_VISIBLE_DEVICES=0, VLLM_ENABLE_V1_MULTIPROCESSING= (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) 2026-06-07 15:53:25.239 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:162 - setup_envvars_for_vllm: numa_enabled=True (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) 2026-06-07 15:53:25.239 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:165 - setup_envvars_for_vllm: set VLLM_ENABLE_V1_MULTIPROCESSING=0 for NUMA affinity (pid=1497003, ip=10.128.16.37) [2026-06-07 15:52:22,695 E 1497003 1497167] core_worker_process.cc:842: Failed to establish connection to the metrics exporter agent. Metrics will not be exported. Exporter agent status: RpcError: Running out of retries to initialize the metrics agent. rpc_code: 14 (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) 2026-06-07 15:53:25.845 | INFO | skyrl_train.utils.numa:set_numa_affinity_for_gpu:340 - NUMA affinity: bound process to CPUs 0-71 (NUMA node 0) for GPU 0 (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) 2026-06-07 15:53:26.289 | DEBUG | skyrl_train.utils.numa:_set_membind_via_libnuma:375 - NUMA affinity: set memory preferred to NUMA node 0 (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) 2026-06-07 15:53:26.289 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:613 - BaseVLLMInferenceEngine: vllm_v1_disable_multiproc=True, vllm.__version__=dev, VLLM_ENABLE_V1_MULTIPROCESSING=0 (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) 2026-06-07 15:53:26.289 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:621 - BaseVLLMInferenceEngine: set VLLM_ENABLE_V1_MULTIPROCESSING=0 (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) 2026-06-07 15:53:26.289 | WARNING | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1187 - OpenAI API sampling params overridden: temperature=0.7, top_p=0.95, top_k=20 (raylet, ip=10.128.16.41) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e (pid=1486338, ip=10.128.16.41) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/vllm/__init__.py:7: RuntimeWarning: Failed to read commit hash: (pid=1486338, ip=10.128.16.41) No module named 'vllm._version' (pid=1486338, ip=10.128.16.41) from .version import __version__, __version_tuple__ # isort:skip (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) 2026-06-07 15:53:50.076 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:152 - setup_envvars_for_vllm: distributed_executor_backend=uni, SKYRL_ENABLE_NUMA_AFFINITY=1, CUDA_VISIBLE_DEVICES=2, VLLM_ENABLE_V1_MULTIPROCESSING= (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) 2026-06-07 15:53:50.077 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:162 - setup_envvars_for_vllm: numa_enabled=True (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) 2026-06-07 15:53:50.077 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:165 - setup_envvars_for_vllm: set VLLM_ENABLE_V1_MULTIPROCESSING=0 for NUMA affinity (raylet, ip=10.128.16.42) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e [repeated 53x across cluster] (pid=1354981, ip=10.128.16.45) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/vllm/__init__.py:7: RuntimeWarning: Failed to read commit hash: [repeated 6x across cluster] (pid=1354981, ip=10.128.16.45) No module named 'vllm._version' [repeated 6x across cluster] (pid=1354981, ip=10.128.16.45) from .version import __version__, __version_tuple__ # isort:skip [repeated 6x across cluster] (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) 2026-06-07 15:53:51.035 | INFO | skyrl_train.utils.numa:set_numa_affinity_for_gpu:340 - NUMA affinity: bound process to CPUs 144-215 (NUMA node 2) for GPU 2 (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) 2026-06-07 15:53:51.038 | DEBUG | skyrl_train.utils.numa:_set_membind_via_libnuma:375 - NUMA affinity: set memory preferred to NUMA node 2 (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) 2026-06-07 15:53:51.038 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:613 - BaseVLLMInferenceEngine: vllm_v1_disable_multiproc=True, vllm.__version__=dev, VLLM_ENABLE_V1_MULTIPROCESSING=0 (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) 2026-06-07 15:53:51.038 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:621 - BaseVLLMInferenceEngine: set VLLM_ENABLE_V1_MULTIPROCESSING=0 (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) 2026-06-07 15:53:51.038 | WARNING | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1187 - OpenAI API sampling params overridden: temperature=0.7, top_p=0.95, top_k=20 (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) 2026-06-07 15:53:52.539 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1239 - Engine startup stagger: sleeping 2.11s (attempt 1/5) to avoid port collisions (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored. (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored. (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) The tokenizer you are loading from '/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6' with an incorrect regex pattern: https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503/discussions/84#69121093e8b480e709447d5e. This will lead to incorrect tokenization. You should set the `fix_mistral_regex=True` flag when loading this tokenizer to fix this issue. (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) 2026-06-07 15:53:50.076 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:152 - setup_envvars_for_vllm: distributed_executor_backend=uni, SKYRL_ENABLE_NUMA_AFFINITY=1, CUDA_VISIBLE_DEVICES=3, VLLM_ENABLE_V1_MULTIPROCESSING= [repeated 2x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) 2026-06-07 15:53:50.077 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:162 - setup_envvars_for_vllm: numa_enabled=True [repeated 2x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) 2026-06-07 15:53:50.077 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:165 - setup_envvars_for_vllm: set VLLM_ENABLE_V1_MULTIPROCESSING=0 for NUMA affinity [repeated 2x across cluster] (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/vllm/__init__.py:7: RuntimeWarning: Failed to read commit hash: (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) No module named 'vllm._version' (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) from .version import __version__, __version_tuple__ # isort:skip (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/vllm/__init__.py:7: RuntimeWarning: Failed to read commit hash: (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) No module named 'vllm._version' (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) from .version import __version__, __version_tuple__ # isort:skip (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) 2026-06-07 15:53:51.025 | INFO | skyrl_train.utils.numa:set_numa_affinity_for_gpu:340 - NUMA affinity: bound process to CPUs 216-287 (NUMA node 3) for GPU 3 [repeated 2x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) 2026-06-07 15:53:51.029 | DEBUG | skyrl_train.utils.numa:_set_membind_via_libnuma:375 - NUMA affinity: set memory preferred to NUMA node 3 [repeated 2x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) 2026-06-07 15:53:51.029 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:613 - BaseVLLMInferenceEngine: vllm_v1_disable_multiproc=True, vllm.__version__=dev, VLLM_ENABLE_V1_MULTIPROCESSING=0 [repeated 2x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) 2026-06-07 15:53:51.029 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:621 - BaseVLLMInferenceEngine: set VLLM_ENABLE_V1_MULTIPROCESSING=0 [repeated 2x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) 2026-06-07 15:53:51.030 | WARNING | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1187 - OpenAI API sampling params overridden: temperature=0.7, top_p=0.95, top_k=20 [repeated 2x across cluster] (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) [2026-06-07 15:54:10,417 E 1354979 1355118] core_worker_process.cc:842: Failed to establish connection to the metrics exporter agent. Metrics will not be exported. Exporter agent status: RpcError: Running out of retries to initialize the metrics agent. rpc_code: 14 (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) 2026-06-07 15:53:52.539 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1239 - Engine startup stagger: sleeping 2.70s (attempt 1/5) to avoid port collisions [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored. [repeated 6x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) The tokenizer you are loading from '/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6' with an incorrect regex pattern: https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503/discussions/84#69121093e8b480e709447d5e. This will lead to incorrect tokenization. You should set the `fix_mistral_regex=True` flag when loading this tokenizer to fix this issue. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/vllm/__init__.py:7: RuntimeWarning: Failed to read commit hash: [repeated 2x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) No module named 'vllm._version' [repeated 2x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) from .version import __version__, __version_tuple__ # isort:skip [repeated 2x across cluster] (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) (EngineCore_DP0 pid=1355204) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/__init__.py:1617: UserWarning: Please use the new API settings to control TF32 behavior, such as torch.backends.cudnn.conv.fp32_precision = 'tf32' or torch.backends.cuda.matmul.fp32_precision = 'ieee'. Old settings, e.g, torch.backends.cuda.matmul.allow_tf32 = True, torch.backends.cudnn.allow_tf32 = True, allowTF32CuDNN() and allowTF32CuBLAS() will be deprecated after Pytorch 2.9. Please see https://pytorch.org/docs/main/notes/cuda.html#tensorfloat-32-tf32-on-ampere-and-later-devices (Triggered internally at /pytorch/aten/src/ATen/Context.cpp:80.) (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) (EngineCore_DP0 pid=1355204) _C._set_float32_matmul_precision(precision) (pid=1486805, ip=10.128.16.41) [2026-06-07 15:54:10,667 E 1486805 1486970] core_worker_process.cc:842: Failed to establish connection to the metrics exporter agent. Metrics will not be exported. Exporter agent status: RpcError: Running out of retries to initialize the metrics agent. rpc_code: 14 [repeated 8x across cluster] (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) (EngineCore_DP0 pid=1355199) _C._set_float32_matmul_precision(precision) (AsyncVLLMInferenceEngine pid=1354980, ip=10.128.16.45) (EngineCore_DP0 pid=1355215) _C._set_float32_matmul_precision(precision) (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) (EngineCore_DP0 pid=1355219) _C._set_float32_matmul_precision(precision) (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) [W607 15:55:02.757090169 socket.cpp:767] [c10d] The client socket cannot be initialized to connect to [jpbo-001-45-interconnect-1.jupiter.internal]:51011 (errno: 97 - Address family not supported by protocol). (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) [W607 15:55:02.760273491 Utils.hpp:166] Warning: Environment variable NCCL_BLOCKING_WAIT is deprecated; use TORCH_NCCL_BLOCKING_WAIT instead (function operator()) (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) [rank0]:[W607 15:55:02.766688838 Utils.hpp:137] Warning: Environment variable TORCH_NCCL_TRACE_BUFFER_SIZE is deprecated; use TORCH_FR_BUFFER_SIZE instead (function operator()) (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) (EngineCore_DP0 pid=1355204) Loading safetensors checkpoint shards: 0% Completed | 0/4 [00:00 (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) 2026-06-07 16:03:12.661 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:162 - setup_envvars_for_vllm: numa_enabled=True (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) 2026-06-07 16:03:12.661 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:165 - setup_envvars_for_vllm: set VLLM_ENABLE_V1_MULTIPROCESSING=0 for NUMA affinity (pid=1538603, ip=10.128.16.47) [2026-06-07 16:00:47,874 E 1538603 1538768] core_worker_process.cc:842: Failed to establish connection to the metrics exporter agent. Metrics will not be exported. Exporter agent status: RpcError: Running out of retries to initialize the metrics agent. rpc_code: 14 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) 2026-06-07 16:03:13.898 | INFO | skyrl_train.utils.numa:set_numa_affinity_for_gpu:340 - NUMA affinity: bound process to CPUs 0-71 (NUMA node 0) for GPU 0 (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) 2026-06-07 16:03:14.213 | DEBUG | skyrl_train.utils.numa:_set_membind_via_libnuma:375 - NUMA affinity: set memory preferred to NUMA node 0 (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) 2026-06-07 16:03:14.213 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:613 - BaseVLLMInferenceEngine: vllm_v1_disable_multiproc=True, vllm.__version__=dev, VLLM_ENABLE_V1_MULTIPROCESSING=0 (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) 2026-06-07 16:03:14.214 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:621 - BaseVLLMInferenceEngine: set VLLM_ENABLE_V1_MULTIPROCESSING=0 (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) 2026-06-07 16:03:14.214 | WARNING | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1187 - OpenAI API sampling params overridden: temperature=0.7, top_p=0.95, top_k=20 (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) 2026-06-07 16:03:14.287 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1239 - Engine startup stagger: sleeping 1.95s (attempt 1/5) to avoid port collisions (raylet, ip=10.128.16.43) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) 2026-06-07 16:03:12.660 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:152 - setup_envvars_for_vllm: distributed_executor_backend=uni, SKYRL_ENABLE_NUMA_AFFINITY=1, CUDA_VISIBLE_DEVICES=3, VLLM_ENABLE_V1_MULTIPROCESSING= [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) 2026-06-07 16:03:12.661 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:162 - setup_envvars_for_vllm: numa_enabled=True [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) 2026-06-07 16:03:12.661 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:165 - setup_envvars_for_vllm: set VLLM_ENABLE_V1_MULTIPROCESSING=0 for NUMA affinity [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) 2026-06-07 16:03:13.898 | INFO | skyrl_train.utils.numa:set_numa_affinity_for_gpu:340 - NUMA affinity: bound process to CPUs 216-287 (NUMA node 3) for GPU 3 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) 2026-06-07 16:03:14.213 | DEBUG | skyrl_train.utils.numa:_set_membind_via_libnuma:375 - NUMA affinity: set memory preferred to NUMA node 3 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) 2026-06-07 16:03:14.213 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:613 - BaseVLLMInferenceEngine: vllm_v1_disable_multiproc=True, vllm.__version__=dev, VLLM_ENABLE_V1_MULTIPROCESSING=0 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) 2026-06-07 16:03:14.213 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:621 - BaseVLLMInferenceEngine: set VLLM_ENABLE_V1_MULTIPROCESSING=0 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) 2026-06-07 16:03:14.213 | WARNING | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1187 - OpenAI API sampling params overridden: temperature=0.7, top_p=0.95, top_k=20 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) 2026-06-07 16:03:14.287 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1239 - Engine startup stagger: sleeping 1.51s (attempt 1/5) to avoid port collisions [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored. (raylet, ip=10.128.16.43) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e [repeated 11x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored. [repeated 3x across cluster] (raylet, ip=10.128.16.36) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e [repeated 2x across cluster] (get_addr_port pid=1204669, ip=10.128.16.43) [2026-06-07 16:04:20,488 E 1204669 1204744] core_worker_process.cc:842: Failed to establish connection to the metrics exporter agent. Metrics will not be exported. Exporter agent status: RpcError: Running out of retries to initialize the metrics agent. rpc_code: 14 (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored. (raylet, ip=10.128.16.36) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e [repeated 22x across cluster] (pid=1204670, ip=10.128.16.43) [2026-06-07 16:04:20,535 E 1204670 1204772] core_worker_process.cc:842: Failed to establish connection to the metrics exporter agent. Metrics will not be exported. Exporter agent status: RpcError: Running out of retries to initialize the metrics agent. rpc_code: 14 (AsyncVLLMInferenceEngine pid=1486803, ip=10.128.16.41) The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored. (pid=1204670, ip=10.128.16.43) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/vllm/__init__.py:7: RuntimeWarning: Failed to read commit hash: (pid=1204670, ip=10.128.16.43) No module named 'vllm._version' (pid=1204670, ip=10.128.16.43) from .version import __version__, __version_tuple__ # isort:skip (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored. [repeated 2x across cluster] (pid=1204805, ip=10.128.16.43) [2026-06-07 16:04:48,441 E 1204805 1204857] core_worker_process.cc:842: Failed to establish connection to the metrics exporter agent. Metrics will not be exported. Exporter agent status: RpcError: Running out of retries to initialize the metrics agent. rpc_code: 14 (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) The tokenizer you are loading from '/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6' with an incorrect regex pattern: https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503/discussions/84#69121093e8b480e709447d5e. This will lead to incorrect tokenization. You should set the `fix_mistral_regex=True` flag when loading this tokenizer to fix this issue. (pid=1204805, ip=10.128.16.43) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/vllm/__init__.py:7: RuntimeWarning: Failed to read commit hash: [repeated 2x across cluster] (pid=1204805, ip=10.128.16.43) No module named 'vllm._version' [repeated 2x across cluster] (pid=1204805, ip=10.128.16.43) from .version import __version__, __version_tuple__ # isort:skip [repeated 2x across cluster] (pid=1495134, ip=10.128.16.36) [2026-06-07 16:04:48,804 E 1495134 1495237] core_worker_process.cc:842: Failed to establish connection to the metrics exporter agent. Metrics will not be exported. Exporter agent status: RpcError: Running out of retries to initialize the metrics agent. rpc_code: 14 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) (EngineCore_DP0 pid=1487375) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/__init__.py:1617: UserWarning: Please use the new API settings to control TF32 behavior, such as torch.backends.cudnn.conv.fp32_precision = 'tf32' or torch.backends.cuda.matmul.fp32_precision = 'ieee'. Old settings, e.g, torch.backends.cuda.matmul.allow_tf32 = True, torch.backends.cudnn.allow_tf32 = True, allowTF32CuDNN() and allowTF32CuBLAS() will be deprecated after Pytorch 2.9. Please see https://pytorch.org/docs/main/notes/cuda.html#tensorfloat-32-tf32-on-ampere-and-later-devices (Triggered internally at /pytorch/aten/src/ATen/Context.cpp:80.) (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) (EngineCore_DP0 pid=1487375) _C._set_float32_matmul_precision(precision) (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) The tokenizer you are loading from '/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6' with an incorrect regex pattern: https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503/discussions/84#69121093e8b480e709447d5e. This will lead to incorrect tokenization. You should set the `fix_mistral_regex=True` flag when loading this tokenizer to fix this issue. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/vllm/__init__.py:7: RuntimeWarning: Failed to read commit hash: [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) No module named 'vllm._version' [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) from .version import __version__, __version_tuple__ # isort:skip [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e [repeated 6x across cluster] (AsyncVLLMInferenceEngine pid=1486803, ip=10.128.16.41) (EngineCore_DP0 pid=1487383) _C._set_float32_matmul_precision(precision) (AsyncVLLMInferenceEngine pid=1486804, ip=10.128.16.41) (EngineCore_DP0 pid=1487374) _C._set_float32_matmul_precision(precision) (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) (EngineCore_DP0 pid=1487379) _C._set_float32_matmul_precision(precision) (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) [W607 16:05:55.641875132 socket.cpp:767] [c10d] The client socket cannot be initialized to connect to [jpbo-001-41-interconnect-1.jupiter.internal]:48789 (errno: 97 - Address family not supported by protocol). (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) [W607 16:05:55.643733206 Utils.hpp:166] Warning: Environment variable NCCL_BLOCKING_WAIT is deprecated; use TORCH_NCCL_BLOCKING_WAIT instead (function operator()) (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) [rank0]:[W607 16:05:55.649849626 Utils.hpp:137] Warning: Environment variable TORCH_NCCL_TRACE_BUFFER_SIZE is deprecated; use TORCH_FR_BUFFER_SIZE instead (function operator()) (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) (EngineCore_DP0 pid=1487375) Loading safetensors checkpoint shards: 0% Completed | 0/4 [00:00 (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) 2026-06-07 16:11:32.007 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:162 - setup_envvars_for_vllm: numa_enabled=True (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) 2026-06-07 16:11:32.007 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:165 - setup_envvars_for_vllm: set VLLM_ENABLE_V1_MULTIPROCESSING=0 for NUMA affinity (pid=1561915, ip=10.128.16.46) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/vllm/__init__.py:7: RuntimeWarning: Failed to read commit hash: [repeated 2x across cluster] (pid=1561915, ip=10.128.16.46) No module named 'vllm._version' [repeated 2x across cluster] (pid=1561915, ip=10.128.16.46) from .version import __version__, __version_tuple__ # isort:skip [repeated 2x across cluster] (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) 2026-06-07 16:11:33.258 | INFO | skyrl_train.utils.numa:set_numa_affinity_for_gpu:340 - NUMA affinity: bound process to CPUs 0-71 (NUMA node 0) for GPU 0 (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) 2026-06-07 16:11:33.747 | DEBUG | skyrl_train.utils.numa:_set_membind_via_libnuma:375 - NUMA affinity: set memory preferred to NUMA node 0 (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) 2026-06-07 16:11:33.747 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:613 - BaseVLLMInferenceEngine: vllm_v1_disable_multiproc=True, vllm.__version__=dev, VLLM_ENABLE_V1_MULTIPROCESSING=0 (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) 2026-06-07 16:11:33.747 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:621 - BaseVLLMInferenceEngine: set VLLM_ENABLE_V1_MULTIPROCESSING=0 (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) 2026-06-07 16:11:33.747 | WARNING | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1187 - OpenAI API sampling params overridden: temperature=0.7, top_p=0.95, top_k=20 (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) 2026-06-07 16:11:33.788 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1239 - Engine startup stagger: sleeping 2.84s (attempt 1/5) to avoid port collisions (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored. (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored. (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) The tokenizer you are loading from '/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6' with an incorrect regex pattern: https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503/discussions/84#69121093e8b480e709447d5e. This will lead to incorrect tokenization. You should set the `fix_mistral_regex=True` flag when loading this tokenizer to fix this issue. (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) 2026-06-07 16:11:32.006 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:152 - setup_envvars_for_vllm: distributed_executor_backend=uni, SKYRL_ENABLE_NUMA_AFFINITY=1, CUDA_VISIBLE_DEVICES=3, VLLM_ENABLE_V1_MULTIPROCESSING= [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) 2026-06-07 16:11:32.007 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:162 - setup_envvars_for_vllm: numa_enabled=True [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) 2026-06-07 16:11:32.007 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:165 - setup_envvars_for_vllm: set VLLM_ENABLE_V1_MULTIPROCESSING=0 for NUMA affinity [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/vllm/__init__.py:7: RuntimeWarning: Failed to read commit hash: (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) No module named 'vllm._version' (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) from .version import __version__, __version_tuple__ # isort:skip (AsyncVLLMInferenceEngine pid=1538601, ip=10.128.16.47) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/vllm/__init__.py:7: RuntimeWarning: Failed to read commit hash: (AsyncVLLMInferenceEngine pid=1538601, ip=10.128.16.47) No module named 'vllm._version' (AsyncVLLMInferenceEngine pid=1538601, ip=10.128.16.47) from .version import __version__, __version_tuple__ # isort:skip (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) 2026-06-07 16:11:33.258 | INFO | skyrl_train.utils.numa:set_numa_affinity_for_gpu:340 - NUMA affinity: bound process to CPUs 216-287 (NUMA node 3) for GPU 3 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) 2026-06-07 16:11:33.746 | DEBUG | skyrl_train.utils.numa:_set_membind_via_libnuma:375 - NUMA affinity: set memory preferred to NUMA node 3 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) 2026-06-07 16:11:33.746 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:613 - BaseVLLMInferenceEngine: vllm_v1_disable_multiproc=True, vllm.__version__=dev, VLLM_ENABLE_V1_MULTIPROCESSING=0 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) 2026-06-07 16:11:33.746 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:621 - BaseVLLMInferenceEngine: set VLLM_ENABLE_V1_MULTIPROCESSING=0 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) 2026-06-07 16:11:33.746 | WARNING | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1187 - OpenAI API sampling params overridden: temperature=0.7, top_p=0.95, top_k=20 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) 2026-06-07 16:11:33.787 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1239 - Engine startup stagger: sleeping 1.53s (attempt 1/5) to avoid port collisions [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored. [repeated 6x across cluster] (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) The tokenizer you are loading from '/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6' with an incorrect regex pattern: https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503/discussions/84#69121093e8b480e709447d5e. This will lead to incorrect tokenization. You should set the `fix_mistral_regex=True` flag when loading this tokenizer to fix this issue. [repeated 3x across cluster] (raylet, ip=10.128.16.38) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e [repeated 9x across cluster] (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/vllm/__init__.py:7: RuntimeWarning: Failed to read commit hash: [repeated 2x across cluster] (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) No module named 'vllm._version' [repeated 2x across cluster] (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) from .version import __version__, __version_tuple__ # isort:skip [repeated 2x across cluster] (get_addr_port pid=1446435, ip=10.128.16.44) [2026-06-07 16:12:25,922 E 1446435 1446509] core_worker_process.cc:842: Failed to establish connection to the metrics exporter agent. Metrics will not be exported. Exporter agent status: RpcError: Running out of retries to initialize the metrics agent. rpc_code: 14 (raylet, ip=10.128.16.38) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e [repeated 27x across cluster] (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) (EngineCore_DP0 pid=1539238) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/__init__.py:1617: UserWarning: Please use the new API settings to control TF32 behavior, such as torch.backends.cudnn.conv.fp32_precision = 'tf32' or torch.backends.cuda.matmul.fp32_precision = 'ieee'. Old settings, e.g, torch.backends.cuda.matmul.allow_tf32 = True, torch.backends.cudnn.allow_tf32 = True, allowTF32CuDNN() and allowTF32CuBLAS() will be deprecated after Pytorch 2.9. Please see https://pytorch.org/docs/main/notes/cuda.html#tensorfloat-32-tf32-on-ampere-and-later-devices (Triggered internally at /pytorch/aten/src/ATen/Context.cpp:80.) (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) (EngineCore_DP0 pid=1539238) _C._set_float32_matmul_precision(precision) (pid=1211472, ip=10.128.16.38) [2026-06-07 16:12:26,041 E 1211472 1211634] core_worker_process.cc:842: Failed to establish connection to the metrics exporter agent. Metrics will not be exported. Exporter agent status: RpcError: Running out of retries to initialize the metrics agent. rpc_code: 14 [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) (EngineCore_DP0 pid=1539220) _C._set_float32_matmul_precision(precision) (AsyncVLLMInferenceEngine pid=1538601, ip=10.128.16.47) (EngineCore_DP0 pid=1539228) _C._set_float32_matmul_precision(precision) (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) [W607 16:13:11.271167533 socket.cpp:767] [c10d] The client socket cannot be initialized to connect to [jpbo-001-47.jupiter.internal]:59713 (errno: 97 - Address family not supported by protocol). (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) [W607 16:13:11.272737868 Utils.hpp:166] Warning: Environment variable NCCL_BLOCKING_WAIT is deprecated; use TORCH_NCCL_BLOCKING_WAIT instead (function operator()) (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) [rank0]:[W607 16:13:11.277554377 Utils.hpp:137] Warning: Environment variable TORCH_NCCL_TRACE_BUFFER_SIZE is deprecated; use TORCH_FR_BUFFER_SIZE instead (function operator()) (AsyncVLLMInferenceEngine pid=1538602, ip=10.128.16.47) (EngineCore_DP0 pid=1539233) _C._set_float32_matmul_precision(precision) (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) (EngineCore_DP0 pid=1539238) Loading safetensors checkpoint shards: 0% Completed | 0/4 [00:00 (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) 2026-06-07 16:16:20.562 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:162 - setup_envvars_for_vllm: numa_enabled=True (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) 2026-06-07 16:16:20.562 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:165 - setup_envvars_for_vllm: set VLLM_ENABLE_V1_MULTIPROCESSING=0 for NUMA affinity (pid=2087855, ip=10.128.16.133) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/vllm/__init__.py:7: RuntimeWarning: Failed to read commit hash: [repeated 3x across cluster] (pid=2087855, ip=10.128.16.133) No module named 'vllm._version' [repeated 3x across cluster] (pid=2087855, ip=10.128.16.133) from .version import __version__, __version_tuple__ # isort:skip [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) 2026-06-07 16:16:21.512 | INFO | skyrl_train.utils.numa:set_numa_affinity_for_gpu:340 - NUMA affinity: bound process to CPUs 0-71 (NUMA node 0) for GPU 0 (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) 2026-06-07 16:16:21.808 | DEBUG | skyrl_train.utils.numa:_set_membind_via_libnuma:375 - NUMA affinity: set memory preferred to NUMA node 0 (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) 2026-06-07 16:16:21.808 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:613 - BaseVLLMInferenceEngine: vllm_v1_disable_multiproc=True, vllm.__version__=dev, VLLM_ENABLE_V1_MULTIPROCESSING=0 (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) 2026-06-07 16:16:21.808 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:621 - BaseVLLMInferenceEngine: set VLLM_ENABLE_V1_MULTIPROCESSING=0 (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) 2026-06-07 16:16:21.808 | WARNING | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1187 - OpenAI API sampling params overridden: temperature=0.7, top_p=0.95, top_k=20 (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) 2026-06-07 16:16:21.855 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1239 - Engine startup stagger: sleeping 1.69s (attempt 1/5) to avoid port collisions (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored. (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored. (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) 2026-06-07 16:16:20.562 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:152 - setup_envvars_for_vllm: distributed_executor_backend=uni, SKYRL_ENABLE_NUMA_AFFINITY=1, CUDA_VISIBLE_DEVICES=2, VLLM_ENABLE_V1_MULTIPROCESSING= [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) 2026-06-07 16:16:20.563 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:162 - setup_envvars_for_vllm: numa_enabled=True [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) 2026-06-07 16:16:20.563 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:165 - setup_envvars_for_vllm: set VLLM_ENABLE_V1_MULTIPROCESSING=0 for NUMA affinity [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) 2026-06-07 16:16:21.988 | INFO | skyrl_train.utils.numa:set_numa_affinity_for_gpu:340 - NUMA affinity: bound process to CPUs 144-215 (NUMA node 2) for GPU 2 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) 2026-06-07 16:16:21.992 | DEBUG | skyrl_train.utils.numa:_set_membind_via_libnuma:375 - NUMA affinity: set memory preferred to NUMA node 2 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) 2026-06-07 16:16:21.992 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:613 - BaseVLLMInferenceEngine: vllm_v1_disable_multiproc=True, vllm.__version__=dev, VLLM_ENABLE_V1_MULTIPROCESSING=0 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) 2026-06-07 16:16:21.992 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:621 - BaseVLLMInferenceEngine: set VLLM_ENABLE_V1_MULTIPROCESSING=0 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) 2026-06-07 16:16:21.992 | WARNING | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1187 - OpenAI API sampling params overridden: temperature=0.7, top_p=0.95, top_k=20 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) 2026-06-07 16:16:22.018 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1239 - Engine startup stagger: sleeping 1.57s (attempt 1/5) to avoid port collisions [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) The tokenizer you are loading from '/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6' with an incorrect regex pattern: https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503/discussions/84#69121093e8b480e709447d5e. This will lead to incorrect tokenization. You should set the `fix_mistral_regex=True` flag when loading this tokenizer to fix this issue. (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/vllm/__init__.py:7: RuntimeWarning: Failed to read commit hash: (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) No module named 'vllm._version' (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) from .version import __version__, __version_tuple__ # isort:skip (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored. [repeated 6x across cluster] (AsyncVLLMInferenceEngine pid=1204670, ip=10.128.16.43) 2026-06-07 16:17:15.622 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:152 - setup_envvars_for_vllm: distributed_executor_backend=uni, SKYRL_ENABLE_NUMA_AFFINITY=1, CUDA_VISIBLE_DEVICES=0, VLLM_ENABLE_V1_MULTIPROCESSING= (AsyncVLLMInferenceEngine pid=1204670, ip=10.128.16.43) 2026-06-07 16:17:15.624 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:162 - setup_envvars_for_vllm: numa_enabled=True (AsyncVLLMInferenceEngine pid=1204670, ip=10.128.16.43) 2026-06-07 16:17:15.624 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:165 - setup_envvars_for_vllm: set VLLM_ENABLE_V1_MULTIPROCESSING=0 for NUMA affinity (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) The tokenizer you are loading from '/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6' with an incorrect regex pattern: https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503/discussions/84#69121093e8b480e709447d5e. This will lead to incorrect tokenization. You should set the `fix_mistral_regex=True` flag when loading this tokenizer to fix this issue. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e [repeated 6x across cluster] (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/vllm/__init__.py:7: RuntimeWarning: Failed to read commit hash: [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) No module named 'vllm._version' [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) from .version import __version__, __version_tuple__ # isort:skip [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1204806, ip=10.128.16.43) 2026-06-07 16:17:15.622 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:152 - setup_envvars_for_vllm: distributed_executor_backend=uni, SKYRL_ENABLE_NUMA_AFFINITY=1, CUDA_VISIBLE_DEVICES=2, VLLM_ENABLE_V1_MULTIPROCESSING= (AsyncVLLMInferenceEngine pid=1204806, ip=10.128.16.43) 2026-06-07 16:17:15.624 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:162 - setup_envvars_for_vllm: numa_enabled=True (AsyncVLLMInferenceEngine pid=1204806, ip=10.128.16.43) 2026-06-07 16:17:15.624 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:165 - setup_envvars_for_vllm: set VLLM_ENABLE_V1_MULTIPROCESSING=0 for NUMA affinity (AsyncVLLMInferenceEngine pid=1204670, ip=10.128.16.43) 2026-06-07 16:17:16.902 | INFO | skyrl_train.utils.numa:set_numa_affinity_for_gpu:340 - NUMA affinity: bound process to CPUs 0-71 (NUMA node 0) for GPU 0 (AsyncVLLMInferenceEngine pid=1204670, ip=10.128.16.43) 2026-06-07 16:17:17.271 | DEBUG | skyrl_train.utils.numa:_set_membind_via_libnuma:375 - NUMA affinity: set memory preferred to NUMA node 0 (AsyncVLLMInferenceEngine pid=1204670, ip=10.128.16.43) 2026-06-07 16:17:17.271 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:613 - BaseVLLMInferenceEngine: vllm_v1_disable_multiproc=True, vllm.__version__=dev, VLLM_ENABLE_V1_MULTIPROCESSING=0 (AsyncVLLMInferenceEngine pid=1204670, ip=10.128.16.43) 2026-06-07 16:17:17.271 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:621 - BaseVLLMInferenceEngine: set VLLM_ENABLE_V1_MULTIPROCESSING=0 (AsyncVLLMInferenceEngine pid=1204670, ip=10.128.16.43) 2026-06-07 16:17:17.271 | WARNING | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1187 - OpenAI API sampling params overridden: temperature=0.7, top_p=0.95, top_k=20 (AsyncVLLMInferenceEngine pid=1204670, ip=10.128.16.43) 2026-06-07 16:17:17.325 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1239 - Engine startup stagger: sleeping 2.27s (attempt 1/5) to avoid port collisions (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) 2026-06-07 16:17:58.363 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:152 - setup_envvars_for_vllm: distributed_executor_backend=uni, SKYRL_ENABLE_NUMA_AFFINITY=1, CUDA_VISIBLE_DEVICES=0, VLLM_ENABLE_V1_MULTIPROCESSING= [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) 2026-06-07 16:17:58.364 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:162 - setup_envvars_for_vllm: numa_enabled=True [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) 2026-06-07 16:17:58.364 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:165 - setup_envvars_for_vllm: set VLLM_ENABLE_V1_MULTIPROCESSING=0 for NUMA affinity [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) 2026-06-07 16:17:16.902 | INFO | skyrl_train.utils.numa:set_numa_affinity_for_gpu:340 - NUMA affinity: bound process to CPUs 216-287 (NUMA node 3) for GPU 3 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) 2026-06-07 16:17:17.270 | DEBUG | skyrl_train.utils.numa:_set_membind_via_libnuma:375 - NUMA affinity: set memory preferred to NUMA node 3 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) 2026-06-07 16:17:17.271 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:613 - BaseVLLMInferenceEngine: vllm_v1_disable_multiproc=True, vllm.__version__=dev, VLLM_ENABLE_V1_MULTIPROCESSING=0 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) 2026-06-07 16:17:17.271 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:621 - BaseVLLMInferenceEngine: set VLLM_ENABLE_V1_MULTIPROCESSING=0 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) 2026-06-07 16:17:17.271 | WARNING | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1187 - OpenAI API sampling params overridden: temperature=0.7, top_p=0.95, top_k=20 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) 2026-06-07 16:17:17.325 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1239 - Engine startup stagger: sleeping 2.88s (attempt 1/5) to avoid port collisions [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495336, ip=10.128.16.36) The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored. (AsyncVLLMInferenceEngine pid=1495336, ip=10.128.16.36) The tokenizer you are loading from '/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6' with an incorrect regex pattern: https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503/discussions/84#69121093e8b480e709447d5e. This will lead to incorrect tokenization. You should set the `fix_mistral_regex=True` flag when loading this tokenizer to fix this issue. (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/vllm/__init__.py:7: RuntimeWarning: Failed to read commit hash: (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) No module named 'vllm._version' (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) from .version import __version__, __version_tuple__ # isort:skip (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) 2026-06-07 16:17:58.363 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:152 - setup_envvars_for_vllm: distributed_executor_backend=uni, SKYRL_ENABLE_NUMA_AFFINITY=1, CUDA_VISIBLE_DEVICES=3, VLLM_ENABLE_V1_MULTIPROCESSING= [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) 2026-06-07 16:17:58.364 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:162 - setup_envvars_for_vllm: numa_enabled=True [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) 2026-06-07 16:17:58.364 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:165 - setup_envvars_for_vllm: set VLLM_ENABLE_V1_MULTIPROCESSING=0 for NUMA affinity [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) 2026-06-07 16:17:59.597 | INFO | skyrl_train.utils.numa:set_numa_affinity_for_gpu:340 - NUMA affinity: bound process to CPUs 216-287 (NUMA node 3) for GPU 3 [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) 2026-06-07 16:17:59.625 | DEBUG | skyrl_train.utils.numa:_set_membind_via_libnuma:375 - NUMA affinity: set memory preferred to NUMA node 3 [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) 2026-06-07 16:17:59.625 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:613 - BaseVLLMInferenceEngine: vllm_v1_disable_multiproc=True, vllm.__version__=dev, VLLM_ENABLE_V1_MULTIPROCESSING=0 [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) 2026-06-07 16:17:59.625 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:621 - BaseVLLMInferenceEngine: set VLLM_ENABLE_V1_MULTIPROCESSING=0 [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) 2026-06-07 16:17:59.626 | WARNING | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1187 - OpenAI API sampling params overridden: temperature=0.7, top_p=0.95, top_k=20 [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) 2026-06-07 16:17:59.672 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1239 - Engine startup stagger: sleeping 2.04s (attempt 1/5) to avoid port collisions [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1204670, ip=10.128.16.43) The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored. [repeated 9x across cluster] (AsyncVLLMInferenceEngine pid=1495335, ip=10.128.16.36) The tokenizer you are loading from '/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6' with an incorrect regex pattern: https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503/discussions/84#69121093e8b480e709447d5e. This will lead to incorrect tokenization. You should set the `fix_mistral_regex=True` flag when loading this tokenizer to fix this issue. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495335, ip=10.128.16.36) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e [repeated 6x across cluster] (AsyncVLLMInferenceEngine pid=1495335, ip=10.128.16.36) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/vllm/__init__.py:7: RuntimeWarning: Failed to read commit hash: [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495335, ip=10.128.16.36) No module named 'vllm._version' [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495335, ip=10.128.16.36) from .version import __version__, __version_tuple__ # isort:skip [repeated 3x across cluster] (pid=3137116, ip=10.128.16.135) Using blocking ray.get inside async actor. This blocks the event loop. Please use `await` on object ref with asyncio.gather if you want to yield execution to the event loop instead. (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored. [repeated 6x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) The tokenizer you are loading from '/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6' with an incorrect regex pattern: https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503/discussions/84#69121093e8b480e709447d5e. This will lead to incorrect tokenization. You should set the `fix_mistral_regex=True` flag when loading this tokenizer to fix this issue. [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e [repeated 8x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/vllm/__init__.py:7: RuntimeWarning: Failed to read commit hash: [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) No module named 'vllm._version' [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) from .version import __version__, __version_tuple__ # isort:skip [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1568416, ip=10.128.16.48) (EngineCore_DP0 pid=1568978) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/__init__.py:1617: UserWarning: Please use the new API settings to control TF32 behavior, such as torch.backends.cudnn.conv.fp32_precision = 'tf32' or torch.backends.cuda.matmul.fp32_precision = 'ieee'. Old settings, e.g, torch.backends.cuda.matmul.allow_tf32 = True, torch.backends.cudnn.allow_tf32 = True, allowTF32CuDNN() and allowTF32CuBLAS() will be deprecated after Pytorch 2.9. Please see https://pytorch.org/docs/main/notes/cuda.html#tensorfloat-32-tf32-on-ampere-and-later-devices (Triggered internally at /pytorch/aten/src/ATen/Context.cpp:80.) (AsyncVLLMInferenceEngine pid=1568416, ip=10.128.16.48) (EngineCore_DP0 pid=1568978) _C._set_float32_matmul_precision(precision) (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) (EngineCore_DP0 pid=1568976) _C._set_float32_matmul_precision(precision) (AsyncVLLMInferenceEngine pid=1568418, ip=10.128.16.48) (EngineCore_DP0 pid=1568977) _C._set_float32_matmul_precision(precision) (AsyncVLLMInferenceEngine pid=1568416, ip=10.128.16.48) [W607 16:20:38.111142654 socket.cpp:767] [c10d] The client socket cannot be initialized to connect to [jpbo-001-48-interconnect-1.jupiter.internal]:53891 (errno: 97 - Address family not supported by protocol). (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) [W607 16:20:38.111310683 Utils.hpp:166] Warning: Environment variable NCCL_BLOCKING_WAIT is deprecated; use TORCH_NCCL_BLOCKING_WAIT instead (function operator()) (AsyncVLLMInferenceEngine pid=1568418, ip=10.128.16.48) [rank0]:[W607 16:20:38.116523344 Utils.hpp:137] Warning: Environment variable TORCH_NCCL_TRACE_BUFFER_SIZE is deprecated; use TORCH_FR_BUFFER_SIZE instead (function operator()) (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) (EngineCore_DP0 pid=1568980) _C._set_float32_matmul_precision(precision) (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) (EngineCore_DP0 pid=1568976) Loading safetensors checkpoint shards: 0% Completed | 0/4 [00:00 (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) 2026-06-07 16:28:40.672 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:162 - setup_envvars_for_vllm: numa_enabled=True (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) 2026-06-07 16:28:40.672 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:165 - setup_envvars_for_vllm: set VLLM_ENABLE_V1_MULTIPROCESSING=0 for NUMA affinity (AsyncVLLMInferenceEngine pid=1204805, ip=10.128.16.43) (EngineCore_DP0 pid=1205529) The tokenizer you are loading from '/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6' with an incorrect regex pattern: https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503/discussions/84#69121093e8b480e709447d5e. This will lead to incorrect tokenization. You should set the `fix_mistral_regex=True` flag when loading this tokenizer to fix this issue. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1204670, ip=10.128.16.43) 2026-06-07 16:24:03.085 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1296 - Initializing OpenAIServingChat with custom_chat_template read from: chat_templates/qwen3_thinking_acc.jinja2 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) 2026-06-07 16:28:41.853 | INFO | skyrl_train.utils.numa:set_numa_affinity_for_gpu:340 - NUMA affinity: bound process to CPUs 0-71 (NUMA node 0) for GPU 0 (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) 2026-06-07 16:28:41.880 | DEBUG | skyrl_train.utils.numa:_set_membind_via_libnuma:375 - NUMA affinity: set memory preferred to NUMA node 0 (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) 2026-06-07 16:28:41.881 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:613 - BaseVLLMInferenceEngine: vllm_v1_disable_multiproc=True, vllm.__version__=dev, VLLM_ENABLE_V1_MULTIPROCESSING=0 (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) 2026-06-07 16:28:41.881 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:621 - BaseVLLMInferenceEngine: set VLLM_ENABLE_V1_MULTIPROCESSING=0 (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) 2026-06-07 16:28:41.881 | WARNING | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1187 - OpenAI API sampling params overridden: temperature=0.7, top_p=0.95, top_k=20 (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) 2026-06-07 16:28:41.928 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1239 - Engine startup stagger: sleeping 1.91s (attempt 1/5) to avoid port collisions (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored. (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored. (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) 2026-06-07 16:28:40.671 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:152 - setup_envvars_for_vllm: distributed_executor_backend=uni, SKYRL_ENABLE_NUMA_AFFINITY=1, CUDA_VISIBLE_DEVICES=2, VLLM_ENABLE_V1_MULTIPROCESSING= [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) 2026-06-07 16:28:40.672 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:162 - setup_envvars_for_vllm: numa_enabled=True [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) 2026-06-07 16:28:40.672 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:setup_envvars_for_vllm:165 - setup_envvars_for_vllm: set VLLM_ENABLE_V1_MULTIPROCESSING=0 for NUMA affinity [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) 2026-06-07 16:28:41.853 | INFO | skyrl_train.utils.numa:set_numa_affinity_for_gpu:340 - NUMA affinity: bound process to CPUs 144-215 (NUMA node 2) for GPU 2 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) 2026-06-07 16:28:41.881 | DEBUG | skyrl_train.utils.numa:_set_membind_via_libnuma:375 - NUMA affinity: set memory preferred to NUMA node 2 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) 2026-06-07 16:28:41.881 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:613 - BaseVLLMInferenceEngine: vllm_v1_disable_multiproc=True, vllm.__version__=dev, VLLM_ENABLE_V1_MULTIPROCESSING=0 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) 2026-06-07 16:28:41.881 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:__init__:621 - BaseVLLMInferenceEngine: set VLLM_ENABLE_V1_MULTIPROCESSING=0 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) 2026-06-07 16:28:41.881 | WARNING | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1187 - OpenAI API sampling params overridden: temperature=0.7, top_p=0.95, top_k=20 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) 2026-06-07 16:28:41.928 | INFO | skyrl_train.inference_engines.vllm.vllm_engine:_create_engine:1239 - Engine startup stagger: sleeping 1.63s (attempt 1/5) to avoid port collisions [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) The tokenizer you are loading from '/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6' with an incorrect regex pattern: https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503/discussions/84#69121093e8b480e709447d5e. This will lead to incorrect tokenization. You should set the `fix_mistral_regex=True` flag when loading this tokenizer to fix this issue. (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/vllm/__init__.py:7: RuntimeWarning: Failed to read commit hash: (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) No module named 'vllm._version' (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) from .version import __version__, __version_tuple__ # isort:skip (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) (EngineCore_DP0 pid=1447387) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/__init__.py:1617: UserWarning: Please use the new API settings to control TF32 behavior, such as torch.backends.cudnn.conv.fp32_precision = 'tf32' or torch.backends.cuda.matmul.fp32_precision = 'ieee'. Old settings, e.g, torch.backends.cuda.matmul.allow_tf32 = True, torch.backends.cudnn.allow_tf32 = True, allowTF32CuDNN() and allowTF32CuBLAS() will be deprecated after Pytorch 2.9. Please see https://pytorch.org/docs/main/notes/cuda.html#tensorfloat-32-tf32-on-ampere-and-later-devices (Triggered internally at /pytorch/aten/src/ATen/Context.cpp:80.) (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) (EngineCore_DP0 pid=1447387) _C._set_float32_matmul_precision(precision) (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) The argument `trust_remote_code` is to be used with Auto classes. It has no effect here and is ignored. [repeated 6x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) The tokenizer you are loading from '/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6' with an incorrect regex pattern: https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503/discussions/84#69121093e8b480e709447d5e. This will lead to incorrect tokenization. You should set the `fix_mistral_regex=True` flag when loading this tokenizer to fix this issue. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) [proxychains] DLL init: proxychains-ng 4.17-git-9-g78ead0e [repeated 6x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/vllm/__init__.py:7: RuntimeWarning: Failed to read commit hash: [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) No module named 'vllm._version' [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) from .version import __version__, __version_tuple__ # isort:skip [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446659, ip=10.128.16.44) (EngineCore_DP0 pid=1447391) _C._set_float32_matmul_precision(precision) (AsyncVLLMInferenceEngine pid=1446660, ip=10.128.16.44) (EngineCore_DP0 pid=1447388) _C._set_float32_matmul_precision(precision) (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) (EngineCore_DP0 pid=1447390) _C._set_float32_matmul_precision(precision) (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) [W607 16:33:02.639468297 socket.cpp:767] [c10d] The client socket cannot be initialized to connect to [jpbo-001-44-interconnect-1.jupiter.internal]:57389 (errno: 97 - Address family not supported by protocol). (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) [W607 16:33:02.642952955 Utils.hpp:166] Warning: Environment variable NCCL_BLOCKING_WAIT is deprecated; use TORCH_NCCL_BLOCKING_WAIT instead (function operator()) (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) [rank0]:[W607 16:33:02.648917654 Utils.hpp:137] Warning: Environment variable TORCH_NCCL_TRACE_BUFFER_SIZE is deprecated; use TORCH_FR_BUFFER_SIZE instead (function operator()) (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) (EngineCore_DP0 pid=1447387) Loading safetensors checkpoint shards: 0% Completed | 0/4 [00:00) File "/e/scratch/jureap59/feuer1/miniforge3/lib/python3.12/concurrent/futures/_base.py", line 449, in result return self.__get_result() ^^^^^^^^^^^^^^^^^^^ File "/e/scratch/jureap59/feuer1/miniforge3/lib/python3.12/concurrent/futures/_base.py", line 401, in __get_result raise self._exception ^^^^^^^^^^^^^^^^^^^^^ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/SkyRL/skyrl-train/skyrl_train/workers/worker.py", line 149, in init_worker_process_group torch.distributed.init_process_group( File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/distributed/c10d_logger.py", line 81, in wrapper return func(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^ File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/distributed/c10d_logger.py", line 95, in wrapper func_return = func(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^ File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/distributed/distributed_c10d.py", line 1762, in init_process_group store, rank, world_size = next(rendezvous_iterator) ^^^^^^^^^^^^^^^^^^^^^^^^^ File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/distributed/rendezvous.py", line 278, in _env_rendezvous_handler store = _create_c10d_store( ^^^^^^^^^^^^^^^^^^^ File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/distributed/rendezvous.py", line 198, in _create_c10d_store return TCPStore( ^^^^^^^^^ torch.distributed.DistStoreError: Timed out after 601 seconds waiting for clients. 4/8 clients joined. Exception raised from waitForWorkers at /pytorch/torch/csrc/distributed/c10d/TCPStore.cpp:396 (most recent call first): frame #0: c10::Error::Error(c10::SourceLocation, std::__cxx11::basic_string, std::allocator >) + 0xb0 (0x4000ab70c700 in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/lib/libc10.so) frame #1: + 0x5e9c9c0 (0x400b025fc9c0 in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/lib/libtorch_cpu.so) frame #2: c10d::TCPStore::waitForWorkers() + 0x350 (0x400b02691410 in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/lib/libtorch_cpu.so) frame #3: c10d::TCPStore::TCPStore(std::__cxx11::basic_string, std::allocator >, c10d::TCPStoreOptions const&) + 0x468 (0x400b026918c8 in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/lib/libtorch_cpu.so) frame #4: + 0x109a094 (0x400afc4fa094 in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/lib/libtorch_python.so) frame #5: + 0x113236c (0x400afc59236c in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/lib/libtorch_python.so) frame #6: + 0x5d6d60 (0x400afba36d60 in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/lib/libtorch_python.so) frame #7: + 0x1b7a38 (0xaaaad83d7a38 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #8: _PyObject_MakeTpCall + 0x98 (0xaaaad8385db8 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #9: + 0x169f50 (0xaaaad8389f50 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #10: + 0x1682e4 (0xaaaad83882e4 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #11: + 0x1e0ce8 (0xaaaad8400ce8 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #12: + 0x1d7ddc (0xaaaad83f7ddc in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #13: + 0x646b0c (0x400afbaa6b0c in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/lib/libtorch_python.so) frame #14: _PyObject_MakeTpCall + 0x98 (0xaaaad8385db8 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #15: _PyEval_EvalFrameDefault + 0x280c (0xaaaad848ae54 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #16: + 0x1808c0 (0xaaaad83a08c0 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #17: + 0x182bf8 (0xaaaad83a2bf8 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #18: + 0x25fd30 (0xaaaad847fd30 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #19: + 0x1b7d20 (0xaaaad83d7d20 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #20: PyObject_Vectorcall + 0x54 (0xaaaad83860e4 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #21: _PyEval_EvalFrameDefault + 0x280c (0xaaaad848ae54 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #22: + 0x1808c0 (0xaaaad83a08c0 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #23: + 0x1835c8 (0xaaaad83a35c8 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #24: + 0x84b708 (0x40002c15b708 in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/ray/_raylet.so) frame #25: + 0x865fe4 (0x40002c175fe4 in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/ray/_raylet.so) frame #26: + 0x823470 (0x40002c133470 in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/ray/_raylet.so) frame #27: + 0x830110 (0x40002c140110 in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/ray/_raylet.so) frame #28: + 0x174a60 (0xaaaad8394a60 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #29: PyObject_VectorcallMethod + 0xa4 (0xaaaad8386270 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #30: PyIter_Send + 0xbc (0xaaaad836b62c in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #31: + 0xaa04 (0x40002e3caa04 in /e/scratch/jureap59/feuer1/miniforge3/lib/python3.12/lib-dynload/_asyncio.cpython-312-aarch64-linux-gnu.so) frame #32: + 0xbcc4 (0x40002e3cbcc4 in /e/scratch/jureap59/feuer1/miniforge3/lib/python3.12/lib-dynload/_asyncio.cpython-312-aarch64-linux-gnu.so) frame #33: _PyObject_MakeTpCall + 0x98 (0xaaaad8385db8 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #34: + 0x28cfac (0xaaaad84acfac in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #35: + 0x1b7b68 (0xaaaad83d7b68 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #36: _PyEval_EvalFrameDefault + 0x52ac (0xaaaad848d8f4 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #37: + 0x83c564 (0x40002c14c564 in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/ray/_raylet.so) frame #38: _PyEval_EvalFrameDefault + 0x52ac (0xaaaad848d8f4 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #39: + 0x169f88 (0xaaaad8389f88 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #40: + 0x36002c (0xaaaad858002c in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #41: + 0x2e4024 (0xaaaad8504024 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #42: + 0x80e00 (0x40002b080e00 in /lib64/libc.so.6) frame #43: + 0xeb49c (0x40002b0eb49c in /lib64/libc.so.6) 2026-06-07 16:54:19.009 | INFO | __main__:main:137 - Shutting down Ray on head node... (FSDPPolicyWorkerBase pid=1620365) [W607 16:54:00.377853387 socket.cpp:767] [c10d] The client socket cannot be initialized to connect to [jpbo-003-39-interconnect-1.jupiter.internal]:54749 (errno: 97 - Address family not supported by protocol). [repeated 3x across cluster] (skyrl_entrypoint pid=1617955) [fd-monitor] Started monitoring (every 120s) (skyrl_entrypoint pid=1617955) [fd-monitor] [15:37:18] OK: 50 / 131,072 FDs open (0.0% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [15:37:18] OK: RSS 1.38 GiB | node mem 171.6/858.0 GiB used (20.0%), avail 686.3 GiB (skyrl_entrypoint pid=1617955) ⚙️ Running in WANDB offline mode (skyrl_entrypoint pid=1617955) [fd-monitor] [15:39:18] OK: 50 / 131,072 FDs open (0.0% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [15:39:18] OK: RSS 1.50 GiB | node mem 171.9/858.0 GiB used (20.0%), avail 686.0 GiB (skyrl_entrypoint pid=1617955) [fd-monitor] [15:41:18] OK: 51 / 131,072 FDs open (0.0% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [15:41:18] OK: RSS 1.58 GiB | node mem 172.0/858.0 GiB used (20.0%), avail 686.0 GiB (skyrl_entrypoint pid=1617955) INFO 06-07 15:42:21 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s (skyrl_entrypoint pid=1617955) [fd-monitor] [15:43:18] OK: 51 / 131,072 FDs open (0.0% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [15:43:18] OK: RSS 1.59 GiB | node mem 172.0/858.0 GiB used (20.0%), avail 686.0 GiB (skyrl_entrypoint pid=1617955) [fd-monitor] [15:45:18] OK: 62 / 131,072 FDs open (0.0% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [15:45:18] OK: RSS 1.63 GiB | node mem 172.2/858.0 GiB used (20.1%), avail 685.8 GiB (skyrl_entrypoint pid=1617955) [fd-monitor] [15:47:18] OK: 64 / 131,072 FDs open (0.0% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [15:47:18] OK: RSS 1.63 GiB | node mem 172.1/858.0 GiB used (20.1%), avail 685.8 GiB (skyrl_entrypoint pid=1617955) [fd-monitor] [15:49:18] OK: 64 / 131,072 FDs open (0.0% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [15:49:18] OK: RSS 1.63 GiB | node mem 172.2/858.0 GiB used (20.1%), avail 685.8 GiB (skyrl_entrypoint pid=1617955) [fd-monitor] [15:51:18] OK: 64 / 131,072 FDs open (0.0% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [15:51:18] OK: RSS 1.63 GiB | node mem 172.2/858.0 GiB used (20.1%), avail 685.8 GiB (skyrl_entrypoint pid=1617955) [fd-monitor] [15:53:18] OK: 64 / 131,072 FDs open (0.0% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [15:53:18] OK: RSS 1.63 GiB | node mem 172.2/858.0 GiB used (20.1%), avail 685.8 GiB (pid=1354512, ip=10.128.16.45) INFO 06-07 15:53:23 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s (pid=1354979, ip=10.128.16.45) INFO 06-07 15:53:49 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s (pid=1354980, ip=10.128.16.45) INFO 06-07 15:53:49 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) WARNING 06-07 15:53:54 [arg_utils.py:1256] The global random seed is set to 56. Since VLLM_ENABLE_V1_MULTIPROCESSING is set to False, this may affect the random state of the Python process that launched vLLM. (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) INFO 06-07 15:53:54 [model.py:529] Resolved architecture: Qwen3ForCausalLM (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) INFO 06-07 15:53:54 [model.py:1549] Using max model len 32768 (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) INFO 06-07 15:53:54 [arg_utils.py:1469] Using ray runtime env (env vars redacted): {'env_vars': {'NCCL_CUMEM_ENABLE': '***', 'NCCL_DEBUG': '***', 'NCCL_SOCKET_FAMILY': '***', 'NCCL_SOCKET_IFNAME': '***', 'RAY_ADDRESS': '***', 'VLLM_ALLOW_INSECURE_SERIALIZATION': '***', 'VLLM_ALLOW_RUNTIME_LORA_UPDATING': '***', 'VLLM_DISABLE_COMPILE_CACHE': '***', 'WANDB_API_KEY': '***'}} (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) INFO 06-07 15:53:54 [scheduler.py:224] Chunked prefill is enabled with max_num_batched_tokens=65536. (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) INFO 06-07 15:53:54 [vllm.py:690] Asynchronous scheduling is enabled. (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) WARNING 06-07 15:53:54 [vllm.py:728] Enforce eager set, overriding optimization level to -O0 (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) INFO 06-07 15:53:54 [vllm.py:846] Cudagraph is disabled under eager mode (pid=1354981, ip=10.128.16.45) INFO 06-07 15:53:49 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) WARNING 06-07 15:53:54 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) WARNING 06-07 15:53:54 [system_utils.py:140] We must use the `spawn` multiprocessing start method. Overriding VLLM_WORKER_MULTIPROC_METHOD to 'spawn'. See https://docs.vllm.ai/en/latest/usage/troubleshooting.html#python-multiprocessing for more information. Reasons: In a Ray actor and can only be spawned (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) INFO 06-07 15:54:01 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) WARNING 06-07 15:53:55 [arg_utils.py:1256] The global random seed is set to 57. Since VLLM_ENABLE_V1_MULTIPROCESSING is set to False, this may affect the random state of the Python process that launched vLLM. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) INFO 06-07 15:53:55 [model.py:529] Resolved architecture: Qwen3ForCausalLM [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) INFO 06-07 15:53:55 [model.py:1549] Using max model len 32768 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) INFO 06-07 15:53:55 [arg_utils.py:1469] Using ray runtime env (env vars redacted): {'env_vars': {'NCCL_CUMEM_ENABLE': '***', 'NCCL_DEBUG': '***', 'NCCL_SOCKET_FAMILY': '***', 'NCCL_SOCKET_IFNAME': '***', 'RAY_ADDRESS': '***', 'VLLM_ALLOW_INSECURE_SERIALIZATION': '***', 'VLLM_ALLOW_RUNTIME_LORA_UPDATING': '***', 'VLLM_DISABLE_COMPILE_CACHE': '***', 'WANDB_API_KEY': '***'}} [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) INFO 06-07 15:53:55 [scheduler.py:224] Chunked prefill is enabled with max_num_batched_tokens=65536. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) INFO 06-07 15:53:55 [vllm.py:690] Asynchronous scheduling is enabled. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) WARNING 06-07 15:53:55 [vllm.py:728] Enforce eager set, overriding optimization level to -O0 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) INFO 06-07 15:53:55 [vllm.py:846] Cudagraph is disabled under eager mode [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) WARNING 06-07 15:53:55 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) WARNING 06-07 15:53:55 [system_utils.py:140] We must use the `spawn` multiprocessing start method. Overriding VLLM_WORKER_MULTIPROC_METHOD to 'spawn'. See https://docs.vllm.ai/en/latest/usage/troubleshooting.html#python-multiprocessing for more information. Reasons: In a Ray actor and can only be spawned [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) (EngineCore_DP0 pid=1355199) INFO 06-07 15:54:02 [core.py:97] Initializing a V1 LLM engine (vdev) with config: model='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', speculative_config=None, tokenizer='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=32768, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False), seed=56, served_model_name=0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'level': None, 'mode': , 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'splitting_ops': [], 'compile_mm_encoder': False, 'compile_sizes': [], 'compile_ranges_split_points': [65536], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': False, 'fuse_act_quant': False, 'fuse_attn_quant': False, 'eliminate_noops': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': True, 'static_all_moe_layers': []} (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) (EngineCore_DP0 pid=1355204) INFO 06-07 15:55:01 [worker_base.py:289] Injected into for extended collective_rpc calls ['_apply_fp8_weight_loader_patches', '_is_fp8_model', '_quantize_weights_for_fp8', '_restore_param_subclasses', '_undo_param_subclasses', 'begin_weight_update', 'destroy_weights_update_group', 'end_weight_update', 'init_weight_update_communicator', 'load_weights', 'read_named_weights', 'set_numa_affinity', 'test_rpc'] (AsyncVLLMInferenceEngine pid=1354980, ip=10.128.16.45) INFO 06-07 15:54:04 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354980, ip=10.128.16.45) (EngineCore_DP0 pid=1355215) INFO 06-07 15:54:04 [core.py:97] Initializing a V1 LLM engine (vdev) with config: model='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', speculative_config=None, tokenizer='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=32768, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False), seed=55, served_model_name=0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'level': None, 'mode': , 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'splitting_ops': [], 'compile_mm_encoder': False, 'compile_sizes': [], 'compile_ranges_split_points': [65536], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': False, 'fuse_act_quant': False, 'fuse_attn_quant': False, 'eliminate_noops': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': True, 'static_all_moe_layers': []} [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) (EngineCore_DP0 pid=1355204) INFO 06-07 15:55:02 [parallel_state.py:1234] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://10.128.16.45:51011 backend=nccl (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) (EngineCore_DP0 pid=1355204) INFO 06-07 15:55:02 [parallel_state.py:1445] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A (skyrl_entrypoint pid=1617955) [fd-monitor] [15:55:18] OK: 66 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [15:55:18] OK: RSS 1.63 GiB | node mem 172.2/858.0 GiB used (20.1%), avail 685.8 GiB (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) (EngineCore_DP0 pid=1355219) INFO 06-07 15:55:01 [worker_base.py:289] Injected into for extended collective_rpc calls ['_apply_fp8_weight_loader_patches', '_is_fp8_model', '_quantize_weights_for_fp8', '_restore_param_subclasses', '_undo_param_subclasses', 'begin_weight_update', 'destroy_weights_update_group', 'end_weight_update', 'init_weight_update_communicator', 'load_weights', 'read_named_weights', 'set_numa_affinity', 'test_rpc'] [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) (EngineCore_DP0 pid=1355219) INFO 06-07 15:55:02 [parallel_state.py:1234] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://10.128.16.45:44505 backend=nccl [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) (EngineCore_DP0 pid=1355219) INFO 06-07 15:55:02 [parallel_state.py:1445] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) (EngineCore_DP0 pid=1355204) INFO 06-07 15:55:28 [gpu_model_runner.py:4125] Starting to load model /e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6... (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) (EngineCore_DP0 pid=1355204) INFO 06-07 15:55:56 [cuda.py:367] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION']. (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) (EngineCore_DP0 pid=1355219) INFO 06-07 15:55:28 [gpu_model_runner.py:4125] Starting to load model /e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6... [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) (EngineCore_DP0 pid=1355204) INFO 06-07 15:56:10 [default_loader.py:293] Loading weights took 14.22 seconds (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) (EngineCore_DP0 pid=1355219) INFO 06-07 15:55:56 [cuda.py:367] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION']. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) (EngineCore_DP0 pid=1355204) INFO 06-07 15:56:11 [gpu_model_runner.py:4222] Model loading took 15.27 GiB memory and 41.635916 seconds (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) (EngineCore_DP0 pid=1355204) INFO 06-07 15:56:13 [gpu_worker.py:373] Available KV cache memory: 49.32 GiB (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) (EngineCore_DP0 pid=1355204) INFO 06-07 15:56:13 [kv_cache_utils.py:1307] GPU KV cache size: 359,104 tokens (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) (EngineCore_DP0 pid=1355204) INFO 06-07 15:56:13 [kv_cache_utils.py:1312] Maximum concurrency for 32,768 tokens per request: 10.96x (AsyncVLLMInferenceEngine pid=1354979, ip=10.128.16.45) (EngineCore_DP0 pid=1355199) INFO 06-07 15:56:14 [kernel_warmup.py:44] Skipping FlashInfer autotune because it is disabled. (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) (EngineCore_DP0 pid=1355204) INFO 06-07 15:56:14 [core.py:278] init engine (profile, create kv cache, warmup model) took 3.24 seconds (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) (EngineCore_DP0 pid=1355204) WARNING 06-07 15:56:14 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) (EngineCore_DP0 pid=1355204) INFO 06-07 15:56:14 [vllm.py:690] Asynchronous scheduling is enabled. (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) (EngineCore_DP0 pid=1355204) WARNING 06-07 15:56:14 [vllm.py:735] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) (EngineCore_DP0 pid=1355204) INFO 06-07 15:56:14 [vllm.py:846] Cudagraph is disabled under eager mode (AsyncVLLMInferenceEngine pid=1354512, ip=10.128.16.45) WARNING 06-07 15:56:14 [model.py:1350] Default vLLM sampling parameters have been overridden by the model's `generation_config.json`: `{'temperature': 0.6, 'top_k': 20, 'top_p': 0.95}`. If this is not intended, please relaunch vLLM instance with `--generation-config vllm`. (skyrl_entrypoint pid=1617955) [fd-monitor] [15:57:18] OK: 66 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [15:57:18] OK: RSS 1.63 GiB | node mem 172.2/858.0 GiB used (20.1%), avail 685.8 GiB (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) (EngineCore_DP0 pid=1355219) INFO 06-07 15:56:10 [default_loader.py:293] Loading weights took 14.19 seconds [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) (EngineCore_DP0 pid=1355219) INFO 06-07 15:56:11 [gpu_model_runner.py:4222] Model loading took 15.27 GiB memory and 41.644509 seconds [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354980, ip=10.128.16.45) (EngineCore_DP0 pid=1355215) INFO 06-07 15:56:13 [gpu_worker.py:373] Available KV cache memory: 49.32 GiB [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354980, ip=10.128.16.45) (EngineCore_DP0 pid=1355215) INFO 06-07 15:56:13 [kv_cache_utils.py:1307] GPU KV cache size: 359,104 tokens [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354980, ip=10.128.16.45) (EngineCore_DP0 pid=1355215) INFO 06-07 15:56:13 [kv_cache_utils.py:1312] Maximum concurrency for 32,768 tokens per request: 10.96x [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) (EngineCore_DP0 pid=1355219) INFO 06-07 15:56:14 [kernel_warmup.py:44] Skipping FlashInfer autotune because it is disabled. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) (EngineCore_DP0 pid=1355219) INFO 06-07 15:56:14 [core.py:278] init engine (profile, create kv cache, warmup model) took 3.27 seconds [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354980, ip=10.128.16.45) (EngineCore_DP0 pid=1355215) WARNING 06-07 15:56:14 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354980, ip=10.128.16.45) (EngineCore_DP0 pid=1355215) INFO 06-07 15:56:14 [vllm.py:690] Asynchronous scheduling is enabled. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354980, ip=10.128.16.45) (EngineCore_DP0 pid=1355215) WARNING 06-07 15:56:14 [vllm.py:735] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354980, ip=10.128.16.45) (EngineCore_DP0 pid=1355215) INFO 06-07 15:56:14 [vllm.py:846] Cudagraph is disabled under eager mode [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1354981, ip=10.128.16.45) WARNING 06-07 15:56:15 [model.py:1350] Default vLLM sampling parameters have been overridden by the model's `generation_config.json`: `{'temperature': 0.6, 'top_k': 20, 'top_p': 0.95}`. If this is not intended, please relaunch vLLM instance with `--generation-config vllm`. [repeated 3x across cluster] (skyrl_entrypoint pid=1617955) [fd-monitor] [15:59:18] OK: 68 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [15:59:18] OK: RSS 1.63 GiB | node mem 172.2/858.0 GiB used (20.1%), avail 685.8 GiB (skyrl_entrypoint pid=1617955) [fd-monitor] [16:01:18] OK: 70 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:01:18] OK: RSS 1.63 GiB | node mem 172.2/858.0 GiB used (20.1%), avail 685.8 GiB (pid=1486338, ip=10.128.16.41) INFO 06-07 16:03:10 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s (skyrl_entrypoint pid=1617955) [fd-monitor] [16:03:18] OK: 70 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:03:18] OK: RSS 1.63 GiB | node mem 172.1/858.0 GiB used (20.1%), avail 685.8 GiB (pid=1486805, ip=10.128.16.41) INFO 06-07 16:03:10 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) WARNING 06-07 16:04:07 [arg_utils.py:1256] The global random seed is set to 50. Since VLLM_ENABLE_V1_MULTIPROCESSING is set to False, this may affect the random state of the Python process that launched vLLM. (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) INFO 06-07 16:04:32 [model.py:529] Resolved architecture: Qwen3ForCausalLM (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) INFO 06-07 16:04:32 [model.py:1549] Using max model len 32768 (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) INFO 06-07 16:04:32 [arg_utils.py:1469] Using ray runtime env (env vars redacted): {'env_vars': {'NCCL_CUMEM_ENABLE': '***', 'NCCL_DEBUG': '***', 'NCCL_SOCKET_FAMILY': '***', 'NCCL_SOCKET_IFNAME': '***', 'RAY_ADDRESS': '***', 'VLLM_ALLOW_INSECURE_SERIALIZATION': '***', 'VLLM_ALLOW_RUNTIME_LORA_UPDATING': '***', 'VLLM_DISABLE_COMPILE_CACHE': '***', 'WANDB_API_KEY': '***'}} (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) INFO 06-07 16:04:32 [scheduler.py:224] Chunked prefill is enabled with max_num_batched_tokens=65536. (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) INFO 06-07 16:04:32 [vllm.py:690] Asynchronous scheduling is enabled. (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) WARNING 06-07 16:04:32 [vllm.py:728] Enforce eager set, overriding optimization level to -O0 (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) INFO 06-07 16:04:32 [vllm.py:846] Cudagraph is disabled under eager mode (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) WARNING 06-07 16:04:07 [arg_utils.py:1256] The global random seed is set to 53. Since VLLM_ENABLE_V1_MULTIPROCESSING is set to False, this may affect the random state of the Python process that launched vLLM. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) WARNING 06-07 16:04:55 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) WARNING 06-07 16:04:55 [system_utils.py:140] We must use the `spawn` multiprocessing start method. Overriding VLLM_WORKER_MULTIPROC_METHOD to 'spawn'. See https://docs.vllm.ai/en/latest/usage/troubleshooting.html#python-multiprocessing for more information. Reasons: In a Ray actor and can only be spawned (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) INFO 06-07 16:04:32 [model.py:529] Resolved architecture: Qwen3ForCausalLM [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) INFO 06-07 16:04:32 [model.py:1549] Using max model len 32768 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) INFO 06-07 16:04:32 [arg_utils.py:1469] Using ray runtime env (env vars redacted): {'env_vars': {'NCCL_CUMEM_ENABLE': '***', 'NCCL_DEBUG': '***', 'NCCL_SOCKET_FAMILY': '***', 'NCCL_SOCKET_IFNAME': '***', 'RAY_ADDRESS': '***', 'VLLM_ALLOW_INSECURE_SERIALIZATION': '***', 'VLLM_ALLOW_RUNTIME_LORA_UPDATING': '***', 'VLLM_DISABLE_COMPILE_CACHE': '***', 'WANDB_API_KEY': '***'}} [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) INFO 06-07 16:04:32 [scheduler.py:224] Chunked prefill is enabled with max_num_batched_tokens=65536. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) INFO 06-07 16:04:32 [vllm.py:690] Asynchronous scheduling is enabled. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) WARNING 06-07 16:04:32 [vllm.py:728] Enforce eager set, overriding optimization level to -O0 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) INFO 06-07 16:04:32 [vllm.py:846] Cudagraph is disabled under eager mode [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) INFO 06-07 16:05:08 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) WARNING 06-07 16:04:55 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) WARNING 06-07 16:04:55 [system_utils.py:140] We must use the `spawn` multiprocessing start method. Overriding VLLM_WORKER_MULTIPROC_METHOD to 'spawn'. See https://docs.vllm.ai/en/latest/usage/troubleshooting.html#python-multiprocessing for more information. Reasons: In a Ray actor and can only be spawned [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) (EngineCore_DP0 pid=1487375) INFO 06-07 16:05:10 [core.py:97] Initializing a V1 LLM engine (vdev) with config: model='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', speculative_config=None, tokenizer='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=32768, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False), seed=50, served_model_name=0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'level': None, 'mode': , 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'splitting_ops': [], 'compile_mm_encoder': False, 'compile_sizes': [], 'compile_ranges_split_points': [65536], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': False, 'fuse_act_quant': False, 'fuse_attn_quant': False, 'eliminate_noops': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': True, 'static_all_moe_layers': []} (skyrl_entrypoint pid=1617955) [fd-monitor] [16:05:18] OK: 74 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:05:18] OK: RSS 1.63 GiB | node mem 172.0/858.0 GiB used (20.1%), avail 685.9 GiB (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) INFO 06-07 16:05:08 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) (EngineCore_DP0 pid=1487379) INFO 06-07 16:05:10 [core.py:97] Initializing a V1 LLM engine (vdev) with config: model='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', speculative_config=None, tokenizer='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=32768, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False), seed=53, served_model_name=0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'level': None, 'mode': , 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'splitting_ops': [], 'compile_mm_encoder': False, 'compile_sizes': [], 'compile_ranges_split_points': [65536], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': False, 'fuse_act_quant': False, 'fuse_attn_quant': False, 'eliminate_noops': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': True, 'static_all_moe_layers': []} [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) (EngineCore_DP0 pid=1487375) INFO 06-07 16:05:54 [worker_base.py:289] Injected into for extended collective_rpc calls ['_apply_fp8_weight_loader_patches', '_is_fp8_model', '_quantize_weights_for_fp8', '_restore_param_subclasses', '_undo_param_subclasses', 'begin_weight_update', 'destroy_weights_update_group', 'end_weight_update', 'init_weight_update_communicator', 'load_weights', 'read_named_weights', 'set_numa_affinity', 'test_rpc'] (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) (EngineCore_DP0 pid=1487375) INFO 06-07 16:05:55 [parallel_state.py:1234] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://10.128.16.41:48789 backend=nccl (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) (EngineCore_DP0 pid=1487375) INFO 06-07 16:05:55 [parallel_state.py:1445] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) (EngineCore_DP0 pid=1487375) INFO 06-07 16:06:01 [gpu_model_runner.py:4125] Starting to load model /e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6... (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) (EngineCore_DP0 pid=1487379) INFO 06-07 16:05:54 [worker_base.py:289] Injected into for extended collective_rpc calls ['_apply_fp8_weight_loader_patches', '_is_fp8_model', '_quantize_weights_for_fp8', '_restore_param_subclasses', '_undo_param_subclasses', 'begin_weight_update', 'destroy_weights_update_group', 'end_weight_update', 'init_weight_update_communicator', 'load_weights', 'read_named_weights', 'set_numa_affinity', 'test_rpc'] [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) (EngineCore_DP0 pid=1487379) INFO 06-07 16:05:55 [parallel_state.py:1234] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://10.128.16.41:48427 backend=nccl [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) (EngineCore_DP0 pid=1487379) INFO 06-07 16:05:55 [parallel_state.py:1445] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) (EngineCore_DP0 pid=1487375) INFO 06-07 16:06:03 [cuda.py:367] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION']. (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) (EngineCore_DP0 pid=1487375) INFO 06-07 16:06:19 [default_loader.py:293] Loading weights took 15.29 seconds (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) (EngineCore_DP0 pid=1487379) INFO 06-07 16:06:01 [gpu_model_runner.py:4125] Starting to load model /e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6... [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) (EngineCore_DP0 pid=1487379) INFO 06-07 16:06:03 [cuda.py:367] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION']. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) (EngineCore_DP0 pid=1487375) INFO 06-07 16:06:19 [gpu_model_runner.py:4222] Model loading took 15.27 GiB memory and 17.289778 seconds (AsyncVLLMInferenceEngine pid=1486804, ip=10.128.16.41) (EngineCore_DP0 pid=1487374) INFO 06-07 16:06:21 [gpu_worker.py:373] Available KV cache memory: 49.32 GiB (AsyncVLLMInferenceEngine pid=1486804, ip=10.128.16.41) (EngineCore_DP0 pid=1487374) INFO 06-07 16:06:21 [kv_cache_utils.py:1307] GPU KV cache size: 359,104 tokens (AsyncVLLMInferenceEngine pid=1486804, ip=10.128.16.41) (EngineCore_DP0 pid=1487374) INFO 06-07 16:06:21 [kv_cache_utils.py:1312] Maximum concurrency for 32,768 tokens per request: 10.96x (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) (EngineCore_DP0 pid=1487375) INFO 06-07 16:06:22 [kernel_warmup.py:44] Skipping FlashInfer autotune because it is disabled. (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) (EngineCore_DP0 pid=1487375) INFO 06-07 16:06:22 [core.py:278] init engine (profile, create kv cache, warmup model) took 3.01 seconds (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) (EngineCore_DP0 pid=1487375) WARNING 06-07 16:06:22 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) (EngineCore_DP0 pid=1487375) INFO 06-07 16:06:22 [vllm.py:690] Asynchronous scheduling is enabled. (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) (EngineCore_DP0 pid=1487375) WARNING 06-07 16:06:22 [vllm.py:735] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) (EngineCore_DP0 pid=1487375) INFO 06-07 16:06:22 [vllm.py:846] Cudagraph is disabled under eager mode (AsyncVLLMInferenceEngine pid=1486338, ip=10.128.16.41) WARNING 06-07 16:06:23 [model.py:1350] Default vLLM sampling parameters have been overridden by the model's `generation_config.json`: `{'temperature': 0.6, 'top_k': 20, 'top_p': 0.95}`. If this is not intended, please relaunch vLLM instance with `--generation-config vllm`. (skyrl_entrypoint pid=1617955) [fd-monitor] [16:07:18] OK: 76 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:07:18] OK: RSS 1.63 GiB | node mem 172.0/858.0 GiB used (20.1%), avail 685.9 GiB (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) (EngineCore_DP0 pid=1487379) INFO 06-07 16:06:19 [default_loader.py:293] Loading weights took 15.29 seconds [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) (EngineCore_DP0 pid=1487379) INFO 06-07 16:06:19 [gpu_model_runner.py:4222] Model loading took 15.27 GiB memory and 17.295963 seconds [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) (EngineCore_DP0 pid=1487379) INFO 06-07 16:06:22 [gpu_worker.py:373] Available KV cache memory: 49.32 GiB [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) (EngineCore_DP0 pid=1487379) INFO 06-07 16:06:22 [kv_cache_utils.py:1307] GPU KV cache size: 359,104 tokens [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) (EngineCore_DP0 pid=1487379) INFO 06-07 16:06:22 [kv_cache_utils.py:1312] Maximum concurrency for 32,768 tokens per request: 10.96x [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) (EngineCore_DP0 pid=1487379) INFO 06-07 16:06:22 [kernel_warmup.py:44] Skipping FlashInfer autotune because it is disabled. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) (EngineCore_DP0 pid=1487379) INFO 06-07 16:06:22 [core.py:278] init engine (profile, create kv cache, warmup model) took 3.04 seconds [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) (EngineCore_DP0 pid=1487379) WARNING 06-07 16:06:22 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) (EngineCore_DP0 pid=1487379) INFO 06-07 16:06:22 [vllm.py:690] Asynchronous scheduling is enabled. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) (EngineCore_DP0 pid=1487379) WARNING 06-07 16:06:22 [vllm.py:735] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486805, ip=10.128.16.41) (EngineCore_DP0 pid=1487379) INFO 06-07 16:06:22 [vllm.py:846] Cudagraph is disabled under eager mode [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1486803, ip=10.128.16.41) WARNING 06-07 16:06:23 [model.py:1350] Default vLLM sampling parameters have been overridden by the model's `generation_config.json`: `{'temperature': 0.6, 'top_k': 20, 'top_p': 0.95}`. If this is not intended, please relaunch vLLM instance with `--generation-config vllm`. [repeated 3x across cluster] (skyrl_entrypoint pid=1617955) [fd-monitor] [16:09:18] OK: 78 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:09:18] OK: RSS 1.63 GiB | node mem 172.1/858.0 GiB used (20.1%), avail 685.9 GiB (skyrl_entrypoint pid=1617955) [fd-monitor] [16:11:18] OK: 78 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:11:18] OK: RSS 1.63 GiB | node mem 172.1/858.0 GiB used (20.1%), avail 685.9 GiB (pid=1538388, ip=10.128.16.47) INFO 06-07 16:11:30 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) WARNING 06-07 16:11:35 [arg_utils.py:1256] The global random seed is set to 65. Since VLLM_ENABLE_V1_MULTIPROCESSING is set to False, this may affect the random state of the Python process that launched vLLM. (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) INFO 06-07 16:11:35 [model.py:529] Resolved architecture: Qwen3ForCausalLM (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) INFO 06-07 16:11:35 [model.py:1549] Using max model len 32768 (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) INFO 06-07 16:11:35 [arg_utils.py:1469] Using ray runtime env (env vars redacted): {'env_vars': {'NCCL_CUMEM_ENABLE': '***', 'NCCL_DEBUG': '***', 'NCCL_SOCKET_FAMILY': '***', 'NCCL_SOCKET_IFNAME': '***', 'RAY_ADDRESS': '***', 'VLLM_ALLOW_INSECURE_SERIALIZATION': '***', 'VLLM_ALLOW_RUNTIME_LORA_UPDATING': '***', 'VLLM_DISABLE_COMPILE_CACHE': '***', 'WANDB_API_KEY': '***'}} (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) INFO 06-07 16:11:35 [scheduler.py:224] Chunked prefill is enabled with max_num_batched_tokens=65536. (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) INFO 06-07 16:11:35 [vllm.py:690] Asynchronous scheduling is enabled. (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) WARNING 06-07 16:11:35 [vllm.py:728] Enforce eager set, overriding optimization level to -O0 (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) INFO 06-07 16:11:35 [vllm.py:846] Cudagraph is disabled under eager mode (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) WARNING 06-07 16:11:35 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) WARNING 06-07 16:11:35 [system_utils.py:140] We must use the `spawn` multiprocessing start method. Overriding VLLM_WORKER_MULTIPROC_METHOD to 'spawn'. See https://docs.vllm.ai/en/latest/usage/troubleshooting.html#python-multiprocessing for more information. Reasons: In a Ray actor and can only be spawned (pid=1538603, ip=10.128.16.47) INFO 06-07 16:11:30 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) WARNING 06-07 16:11:36 [arg_utils.py:1256] The global random seed is set to 62. Since VLLM_ENABLE_V1_MULTIPROCESSING is set to False, this may affect the random state of the Python process that launched vLLM. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) INFO 06-07 16:11:36 [model.py:529] Resolved architecture: Qwen3ForCausalLM [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) INFO 06-07 16:11:36 [model.py:1549] Using max model len 32768 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) INFO 06-07 16:11:36 [arg_utils.py:1469] Using ray runtime env (env vars redacted): {'env_vars': {'NCCL_CUMEM_ENABLE': '***', 'NCCL_DEBUG': '***', 'NCCL_SOCKET_FAMILY': '***', 'NCCL_SOCKET_IFNAME': '***', 'RAY_ADDRESS': '***', 'VLLM_ALLOW_INSECURE_SERIALIZATION': '***', 'VLLM_ALLOW_RUNTIME_LORA_UPDATING': '***', 'VLLM_DISABLE_COMPILE_CACHE': '***', 'WANDB_API_KEY': '***'}} [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) INFO 06-07 16:11:36 [scheduler.py:224] Chunked prefill is enabled with max_num_batched_tokens=65536. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) INFO 06-07 16:11:36 [vllm.py:690] Asynchronous scheduling is enabled. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) WARNING 06-07 16:11:36 [vllm.py:728] Enforce eager set, overriding optimization level to -O0 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) INFO 06-07 16:11:36 [vllm.py:846] Cudagraph is disabled under eager mode [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) WARNING 06-07 16:11:37 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) WARNING 06-07 16:11:37 [system_utils.py:140] We must use the `spawn` multiprocessing start method. Overriding VLLM_WORKER_MULTIPROC_METHOD to 'spawn'. See https://docs.vllm.ai/en/latest/usage/troubleshooting.html#python-multiprocessing for more information. Reasons: In a Ray actor and can only be spawned [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) INFO 06-07 16:11:45 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s (AsyncVLLMInferenceEngine pid=1538601, ip=10.128.16.47) INFO 06-07 16:11:45 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) (EngineCore_DP0 pid=1539238) INFO 06-07 16:11:46 [core.py:97] Initializing a V1 LLM engine (vdev) with config: model='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', speculative_config=None, tokenizer='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=32768, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False), seed=62, served_model_name=0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'level': None, 'mode': , 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'splitting_ops': [], 'compile_mm_encoder': False, 'compile_sizes': [], 'compile_ranges_split_points': [65536], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': False, 'fuse_act_quant': False, 'fuse_attn_quant': False, 'eliminate_noops': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': True, 'static_all_moe_layers': []} (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) (EngineCore_DP0 pid=1539238) INFO 06-07 16:13:10 [worker_base.py:289] Injected into for extended collective_rpc calls ['_apply_fp8_weight_loader_patches', '_is_fp8_model', '_quantize_weights_for_fp8', '_restore_param_subclasses', '_undo_param_subclasses', 'begin_weight_update', 'destroy_weights_update_group', 'end_weight_update', 'init_weight_update_communicator', 'load_weights', 'read_named_weights', 'set_numa_affinity', 'test_rpc'] (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) INFO 06-07 16:11:45 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s [repeated 2x across cluster] (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) (EngineCore_DP0 pid=1539220) INFO 06-07 16:11:46 [core.py:97] Initializing a V1 LLM engine (vdev) with config: model='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', speculative_config=None, tokenizer='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=32768, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False), seed=65, served_model_name=0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'level': None, 'mode': , 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'splitting_ops': [], 'compile_mm_encoder': False, 'compile_sizes': [], 'compile_ranges_split_points': [65536], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': False, 'fuse_act_quant': False, 'fuse_attn_quant': False, 'eliminate_noops': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': True, 'static_all_moe_layers': []} [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) (EngineCore_DP0 pid=1539238) INFO 06-07 16:13:11 [parallel_state.py:1234] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://10.128.16.47:59713 backend=nccl (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) (EngineCore_DP0 pid=1539238) INFO 06-07 16:13:11 [parallel_state.py:1445] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A (skyrl_entrypoint pid=1617955) [fd-monitor] [16:13:18] OK: 80 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:13:18] OK: RSS 1.63 GiB | node mem 172.0/858.0 GiB used (20.1%), avail 685.9 GiB (AsyncVLLMInferenceEngine pid=1538602, ip=10.128.16.47) (EngineCore_DP0 pid=1539233) INFO 06-07 16:13:11 [worker_base.py:289] Injected into for extended collective_rpc calls ['_apply_fp8_weight_loader_patches', '_is_fp8_model', '_quantize_weights_for_fp8', '_restore_param_subclasses', '_undo_param_subclasses', 'begin_weight_update', 'destroy_weights_update_group', 'end_weight_update', 'init_weight_update_communicator', 'load_weights', 'read_named_weights', 'set_numa_affinity', 'test_rpc'] [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538602, ip=10.128.16.47) (EngineCore_DP0 pid=1539233) INFO 06-07 16:13:11 [parallel_state.py:1234] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://10.128.16.47:60611 backend=nccl [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538602, ip=10.128.16.47) (EngineCore_DP0 pid=1539233) INFO 06-07 16:13:11 [parallel_state.py:1445] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) (EngineCore_DP0 pid=1539238) INFO 06-07 16:13:37 [gpu_model_runner.py:4125] Starting to load model /e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6... (AsyncVLLMInferenceEngine pid=1538602, ip=10.128.16.47) (EngineCore_DP0 pid=1539233) INFO 06-07 16:13:38 [cuda.py:367] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION']. (AsyncVLLMInferenceEngine pid=1538601, ip=10.128.16.47) (EngineCore_DP0 pid=1539228) INFO 06-07 16:13:53 [default_loader.py:293] Loading weights took 13.80 seconds (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) (EngineCore_DP0 pid=1539220) INFO 06-07 16:13:37 [gpu_model_runner.py:4125] Starting to load model /e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6... [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538601, ip=10.128.16.47) (EngineCore_DP0 pid=1539228) INFO 06-07 16:13:38 [cuda.py:367] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION']. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) (EngineCore_DP0 pid=1539238) INFO 06-07 16:13:53 [gpu_model_runner.py:4222] Model loading took 15.27 GiB memory and 15.105954 seconds (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) (EngineCore_DP0 pid=1539220) INFO 06-07 16:13:55 [gpu_worker.py:373] Available KV cache memory: 49.32 GiB (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) (EngineCore_DP0 pid=1539220) INFO 06-07 16:13:55 [kv_cache_utils.py:1307] GPU KV cache size: 359,104 tokens (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) (EngineCore_DP0 pid=1539220) INFO 06-07 16:13:55 [kv_cache_utils.py:1312] Maximum concurrency for 32,768 tokens per request: 10.96x (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) (EngineCore_DP0 pid=1539238) INFO 06-07 16:13:56 [kernel_warmup.py:44] Skipping FlashInfer autotune because it is disabled. (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) (EngineCore_DP0 pid=1539238) INFO 06-07 16:13:56 [core.py:278] init engine (profile, create kv cache, warmup model) took 2.79 seconds (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) (EngineCore_DP0 pid=1539238) WARNING 06-07 16:13:56 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) (EngineCore_DP0 pid=1539238) INFO 06-07 16:13:56 [vllm.py:690] Asynchronous scheduling is enabled. (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) (EngineCore_DP0 pid=1539238) WARNING 06-07 16:13:56 [vllm.py:735] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) (EngineCore_DP0 pid=1539238) INFO 06-07 16:13:56 [vllm.py:846] Cudagraph is disabled under eager mode (AsyncVLLMInferenceEngine pid=1538601, ip=10.128.16.47) WARNING 06-07 16:13:56 [model.py:1350] Default vLLM sampling parameters have been overridden by the model's `generation_config.json`: `{'temperature': 0.6, 'top_k': 20, 'top_p': 0.95}`. If this is not intended, please relaunch vLLM instance with `--generation-config vllm`. (skyrl_entrypoint pid=1617955) [fd-monitor] [16:15:18] OK: 87 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:15:18] OK: RSS 1.64 GiB | node mem 172.0/858.0 GiB used (20.0%), avail 686.0 GiB (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) (EngineCore_DP0 pid=1539238) INFO 06-07 16:13:53 [default_loader.py:293] Loading weights took 13.80 seconds [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) (EngineCore_DP0 pid=1539220) INFO 06-07 16:13:53 [gpu_model_runner.py:4222] Model loading took 15.27 GiB memory and 15.107068 seconds [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538602, ip=10.128.16.47) (EngineCore_DP0 pid=1539233) INFO 06-07 16:13:55 [gpu_worker.py:373] Available KV cache memory: 49.32 GiB [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538602, ip=10.128.16.47) (EngineCore_DP0 pid=1539233) INFO 06-07 16:13:55 [kv_cache_utils.py:1307] GPU KV cache size: 359,104 tokens [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538602, ip=10.128.16.47) (EngineCore_DP0 pid=1539233) INFO 06-07 16:13:55 [kv_cache_utils.py:1312] Maximum concurrency for 32,768 tokens per request: 10.96x [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538602, ip=10.128.16.47) (EngineCore_DP0 pid=1539233) INFO 06-07 16:13:56 [kernel_warmup.py:44] Skipping FlashInfer autotune because it is disabled. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) (EngineCore_DP0 pid=1539220) INFO 06-07 16:13:56 [core.py:278] init engine (profile, create kv cache, warmup model) took 2.79 seconds [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) (EngineCore_DP0 pid=1539220) WARNING 06-07 16:13:56 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) (EngineCore_DP0 pid=1539220) INFO 06-07 16:13:56 [vllm.py:690] Asynchronous scheduling is enabled. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) (EngineCore_DP0 pid=1539220) WARNING 06-07 16:13:56 [vllm.py:735] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538603, ip=10.128.16.47) (EngineCore_DP0 pid=1539220) INFO 06-07 16:13:56 [vllm.py:846] Cudagraph is disabled under eager mode [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1538388, ip=10.128.16.47) WARNING 06-07 16:13:56 [model.py:1350] Default vLLM sampling parameters have been overridden by the model's `generation_config.json`: `{'temperature': 0.6, 'top_k': 20, 'top_p': 0.95}`. If this is not intended, please relaunch vLLM instance with `--generation-config vllm`. [repeated 3x across cluster] (pid=1568271, ip=10.128.16.48) INFO 06-07 16:15:51 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) WARNING 06-07 16:16:49 [arg_utils.py:1256] The global random seed is set to 74. Since VLLM_ENABLE_V1_MULTIPROCESSING is set to False, this may affect the random state of the Python process that launched vLLM. (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) INFO 06-07 16:16:49 [model.py:529] Resolved architecture: Qwen3ForCausalLM (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) INFO 06-07 16:16:49 [model.py:1549] Using max model len 32768 (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) INFO 06-07 16:16:49 [arg_utils.py:1469] Using ray runtime env (env vars redacted): {'env_vars': {'NCCL_CUMEM_ENABLE': '***', 'NCCL_DEBUG': '***', 'NCCL_SOCKET_FAMILY': '***', 'NCCL_SOCKET_IFNAME': '***', 'RAY_ADDRESS': '***', 'VLLM_ALLOW_INSECURE_SERIALIZATION': '***', 'VLLM_ALLOW_RUNTIME_LORA_UPDATING': '***', 'VLLM_DISABLE_COMPILE_CACHE': '***', 'WANDB_API_KEY': '***'}} (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) INFO 06-07 16:16:49 [scheduler.py:224] Chunked prefill is enabled with max_num_batched_tokens=65536. (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) INFO 06-07 16:16:49 [vllm.py:690] Asynchronous scheduling is enabled. (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) WARNING 06-07 16:16:49 [vllm.py:728] Enforce eager set, overriding optimization level to -O0 (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) INFO 06-07 16:16:49 [vllm.py:846] Cudagraph is disabled under eager mode (pid=1568417, ip=10.128.16.48) INFO 06-07 16:15:51 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) WARNING 06-07 16:16:49 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) WARNING 06-07 16:16:50 [system_utils.py:140] We must use the `spawn` multiprocessing start method. Overriding VLLM_WORKER_MULTIPROC_METHOD to 'spawn'. See https://docs.vllm.ai/en/latest/usage/troubleshooting.html#python-multiprocessing for more information. Reasons: In a Ray actor and can only be spawned (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) WARNING 06-07 16:16:49 [arg_utils.py:1256] The global random seed is set to 76. Since VLLM_ENABLE_V1_MULTIPROCESSING is set to False, this may affect the random state of the Python process that launched vLLM. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) INFO 06-07 16:16:49 [model.py:529] Resolved architecture: Qwen3ForCausalLM [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) INFO 06-07 16:16:49 [model.py:1549] Using max model len 32768 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) INFO 06-07 16:16:49 [arg_utils.py:1469] Using ray runtime env (env vars redacted): {'env_vars': {'NCCL_CUMEM_ENABLE': '***', 'NCCL_DEBUG': '***', 'NCCL_SOCKET_FAMILY': '***', 'NCCL_SOCKET_IFNAME': '***', 'RAY_ADDRESS': '***', 'VLLM_ALLOW_INSECURE_SERIALIZATION': '***', 'VLLM_ALLOW_RUNTIME_LORA_UPDATING': '***', 'VLLM_DISABLE_COMPILE_CACHE': '***', 'WANDB_API_KEY': '***'}} [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) INFO 06-07 16:16:49 [scheduler.py:224] Chunked prefill is enabled with max_num_batched_tokens=65536. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) INFO 06-07 16:16:49 [vllm.py:690] Asynchronous scheduling is enabled. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) WARNING 06-07 16:16:49 [vllm.py:728] Enforce eager set, overriding optimization level to -O0 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) INFO 06-07 16:16:49 [vllm.py:846] Cudagraph is disabled under eager mode [repeated 3x across cluster] (pid=1204670, ip=10.128.16.43) INFO 06-07 16:17:13 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) WARNING 06-07 16:16:49 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) WARNING 06-07 16:16:49 [system_utils.py:140] We must use the `spawn` multiprocessing start method. Overriding VLLM_WORKER_MULTIPROC_METHOD to 'spawn'. See https://docs.vllm.ai/en/latest/usage/troubleshooting.html#python-multiprocessing for more information. Reasons: In a Ray actor and can only be spawned [repeated 3x across cluster] (pid=1204806, ip=10.128.16.43) INFO 06-07 16:17:13 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s (skyrl_entrypoint pid=1617955) [fd-monitor] [16:17:18] OK: 87 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:17:18] OK: RSS 1.64 GiB | node mem 172.0/858.0 GiB used (20.0%), avail 686.0 GiB (pid=1495134, ip=10.128.16.36) INFO 06-07 16:17:56 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495336, ip=10.128.16.36) WARNING 06-07 16:18:01 [arg_utils.py:1256] The global random seed is set to 72. Since VLLM_ENABLE_V1_MULTIPROCESSING is set to False, this may affect the random state of the Python process that launched vLLM. (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) INFO 06-07 16:18:01 [model.py:529] Resolved architecture: Qwen3ForCausalLM (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) INFO 06-07 16:18:01 [model.py:1549] Using max model len 32768 (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) INFO 06-07 16:18:01 [arg_utils.py:1469] Using ray runtime env (env vars redacted): {'env_vars': {'NCCL_CUMEM_ENABLE': '***', 'NCCL_DEBUG': '***', 'NCCL_SOCKET_FAMILY': '***', 'NCCL_SOCKET_IFNAME': '***', 'RAY_ADDRESS': '***', 'VLLM_ALLOW_INSECURE_SERIALIZATION': '***', 'VLLM_ALLOW_RUNTIME_LORA_UPDATING': '***', 'VLLM_DISABLE_COMPILE_CACHE': '***', 'WANDB_API_KEY': '***'}} (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) INFO 06-07 16:18:01 [scheduler.py:224] Chunked prefill is enabled with max_num_batched_tokens=65536. (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) INFO 06-07 16:18:01 [vllm.py:690] Asynchronous scheduling is enabled. (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) WARNING 06-07 16:18:01 [vllm.py:728] Enforce eager set, overriding optimization level to -O0 (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) INFO 06-07 16:18:01 [vllm.py:846] Cudagraph is disabled under eager mode (pid=1495334, ip=10.128.16.36) INFO 06-07 16:17:56 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495336, ip=10.128.16.36) WARNING 06-07 16:18:01 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 (AsyncVLLMInferenceEngine pid=1495336, ip=10.128.16.36) WARNING 06-07 16:18:01 [system_utils.py:140] We must use the `spawn` multiprocessing start method. Overriding VLLM_WORKER_MULTIPROC_METHOD to 'spawn'. See https://docs.vllm.ai/en/latest/usage/troubleshooting.html#python-multiprocessing for more information. Reasons: In a Ray actor and can only be spawned (AsyncVLLMInferenceEngine pid=1204670, ip=10.128.16.43) WARNING 06-07 16:18:11 [arg_utils.py:1256] The global random seed is set to 67. Since VLLM_ENABLE_V1_MULTIPROCESSING is set to False, this may affect the random state of the Python process that launched vLLM. [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1495335, ip=10.128.16.36) INFO 06-07 16:18:02 [model.py:529] Resolved architecture: Qwen3ForCausalLM [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495335, ip=10.128.16.36) INFO 06-07 16:18:02 [model.py:1549] Using max model len 32768 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495335, ip=10.128.16.36) INFO 06-07 16:18:02 [arg_utils.py:1469] Using ray runtime env (env vars redacted): {'env_vars': {'NCCL_CUMEM_ENABLE': '***', 'NCCL_DEBUG': '***', 'NCCL_SOCKET_FAMILY': '***', 'NCCL_SOCKET_IFNAME': '***', 'RAY_ADDRESS': '***', 'VLLM_ALLOW_INSECURE_SERIALIZATION': '***', 'VLLM_ALLOW_RUNTIME_LORA_UPDATING': '***', 'VLLM_DISABLE_COMPILE_CACHE': '***', 'WANDB_API_KEY': '***'}} [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495335, ip=10.128.16.36) INFO 06-07 16:18:02 [scheduler.py:224] Chunked prefill is enabled with max_num_batched_tokens=65536. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495335, ip=10.128.16.36) INFO 06-07 16:18:02 [vllm.py:690] Asynchronous scheduling is enabled. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495335, ip=10.128.16.36) WARNING 06-07 16:18:02 [vllm.py:728] Enforce eager set, overriding optimization level to -O0 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495335, ip=10.128.16.36) INFO 06-07 16:18:02 [vllm.py:846] Cudagraph is disabled under eager mode [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495335, ip=10.128.16.36) WARNING 06-07 16:18:02 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495335, ip=10.128.16.36) WARNING 06-07 16:18:02 [system_utils.py:140] We must use the `spawn` multiprocessing start method. Overriding VLLM_WORKER_MULTIPROC_METHOD to 'spawn'. See https://docs.vllm.ai/en/latest/usage/troubleshooting.html#python-multiprocessing for more information. Reasons: In a Ray actor and can only be spawned [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) INFO 06-07 16:19:13 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) WARNING 06-07 16:18:11 [arg_utils.py:1256] The global random seed is set to 78. Since VLLM_ENABLE_V1_MULTIPROCESSING is set to False, this may affect the random state of the Python process that launched vLLM. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) INFO 06-07 16:18:11 [model.py:529] Resolved architecture: Qwen3ForCausalLM [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) INFO 06-07 16:18:11 [model.py:1549] Using max model len 32768 [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) INFO 06-07 16:18:11 [arg_utils.py:1469] Using ray runtime env (env vars redacted): {'env_vars': {'NCCL_CUMEM_ENABLE': '***', 'NCCL_DEBUG': '***', 'NCCL_SOCKET_FAMILY': '***', 'NCCL_SOCKET_IFNAME': '***', 'RAY_ADDRESS': '***', 'VLLM_ALLOW_INSECURE_SERIALIZATION': '***', 'VLLM_ALLOW_RUNTIME_LORA_UPDATING': '***', 'VLLM_DISABLE_COMPILE_CACHE': '***', 'WANDB_API_KEY': '***'}} [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) INFO 06-07 16:18:11 [scheduler.py:224] Chunked prefill is enabled with max_num_batched_tokens=65536. [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) INFO 06-07 16:18:11 [vllm.py:690] Asynchronous scheduling is enabled. [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) WARNING 06-07 16:18:11 [vllm.py:728] Enforce eager set, overriding optimization level to -O0 [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) INFO 06-07 16:18:11 [vllm.py:846] Cudagraph is disabled under eager mode [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) WARNING 06-07 16:18:12 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) WARNING 06-07 16:18:12 [system_utils.py:140] We must use the `spawn` multiprocessing start method. Overriding VLLM_WORKER_MULTIPROC_METHOD to 'spawn'. See https://docs.vllm.ai/en/latest/usage/troubleshooting.html#python-multiprocessing for more information. Reasons: In a Ray actor and can only be spawned [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) (EngineCore_DP0 pid=1568980) INFO 06-07 16:19:14 [core.py:97] Initializing a V1 LLM engine (vdev) with config: model='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', speculative_config=None, tokenizer='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=32768, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False), seed=74, served_model_name=0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'level': None, 'mode': , 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'splitting_ops': [], 'compile_mm_encoder': False, 'compile_sizes': [], 'compile_ranges_split_points': [65536], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': False, 'fuse_act_quant': False, 'fuse_attn_quant': False, 'eliminate_noops': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': True, 'static_all_moe_layers': []} (skyrl_entrypoint pid=1617955) [fd-monitor] [16:19:18] OK: 87 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:19:18] OK: RSS 1.64 GiB | node mem 171.9/858.0 GiB used (20.0%), avail 686.0 GiB (AsyncVLLMInferenceEngine pid=1204670, ip=10.128.16.43) INFO 06-07 16:20:08 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) (EngineCore_DP0 pid=1568976) INFO 06-07 16:19:14 [core.py:97] Initializing a V1 LLM engine (vdev) with config: model='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', speculative_config=None, tokenizer='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=32768, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False), seed=76, served_model_name=0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'level': None, 'mode': , 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'splitting_ops': [], 'compile_mm_encoder': False, 'compile_sizes': [], 'compile_ranges_split_points': [65536], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': False, 'fuse_act_quant': False, 'fuse_attn_quant': False, 'eliminate_noops': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': True, 'static_all_moe_layers': []} [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568416, ip=10.128.16.48) (EngineCore_DP0 pid=1568978) INFO 06-07 16:20:37 [worker_base.py:289] Injected into for extended collective_rpc calls ['_apply_fp8_weight_loader_patches', '_is_fp8_model', '_quantize_weights_for_fp8', '_restore_param_subclasses', '_undo_param_subclasses', 'begin_weight_update', 'destroy_weights_update_group', 'end_weight_update', 'init_weight_update_communicator', 'load_weights', 'read_named_weights', 'set_numa_affinity', 'test_rpc'] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) INFO 06-07 16:20:08 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) (EngineCore_DP0 pid=1205528) INFO 06-07 16:20:10 [core.py:97] Initializing a V1 LLM engine (vdev) with config: model='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', speculative_config=None, tokenizer='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=32768, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False), seed=78, served_model_name=0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'level': None, 'mode': , 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'splitting_ops': [], 'compile_mm_encoder': False, 'compile_sizes': [], 'compile_ranges_split_points': [65536], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': False, 'fuse_act_quant': False, 'fuse_attn_quant': False, 'eliminate_noops': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': True, 'static_all_moe_layers': []} [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1568418, ip=10.128.16.48) (EngineCore_DP0 pid=1568977) INFO 06-07 16:20:38 [parallel_state.py:1234] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://10.128.16.48:55611 backend=nccl (AsyncVLLMInferenceEngine pid=1568418, ip=10.128.16.48) (EngineCore_DP0 pid=1568977) INFO 06-07 16:20:38 [parallel_state.py:1445] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A (AsyncVLLMInferenceEngine pid=1568418, ip=10.128.16.48) (EngineCore_DP0 pid=1568977) INFO 06-07 16:20:38 [gpu_model_runner.py:4125] Starting to load model /e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6... (AsyncVLLMInferenceEngine pid=1568418, ip=10.128.16.48) (EngineCore_DP0 pid=1568977) INFO 06-07 16:20:40 [cuda.py:367] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION']. (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) (EngineCore_DP0 pid=1568980) INFO 06-07 16:20:38 [worker_base.py:289] Injected into for extended collective_rpc calls ['_apply_fp8_weight_loader_patches', '_is_fp8_model', '_quantize_weights_for_fp8', '_restore_param_subclasses', '_undo_param_subclasses', 'begin_weight_update', 'destroy_weights_update_group', 'end_weight_update', 'init_weight_update_communicator', 'load_weights', 'read_named_weights', 'set_numa_affinity', 'test_rpc'] [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) INFO 06-07 16:20:51 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) (EngineCore_DP0 pid=1568980) INFO 06-07 16:20:38 [parallel_state.py:1234] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://10.128.16.48:34821 backend=nccl [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) (EngineCore_DP0 pid=1568980) INFO 06-07 16:20:38 [parallel_state.py:1445] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) (EngineCore_DP0 pid=1568980) INFO 06-07 16:20:39 [gpu_model_runner.py:4125] Starting to load model /e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6... [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) (EngineCore_DP0 pid=1568980) INFO 06-07 16:20:40 [cuda.py:367] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION']. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495335, ip=10.128.16.36) INFO 06-07 16:20:51 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) (EngineCore_DP0 pid=1495916) INFO 06-07 16:20:52 [core.py:97] Initializing a V1 LLM engine (vdev) with config: model='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', speculative_config=None, tokenizer='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=32768, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False), seed=70, served_model_name=0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'level': None, 'mode': , 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'splitting_ops': [], 'compile_mm_encoder': False, 'compile_sizes': [], 'compile_ranges_split_points': [65536], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': False, 'fuse_act_quant': False, 'fuse_attn_quant': False, 'eliminate_noops': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': True, 'static_all_moe_layers': []} (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) (EngineCore_DP0 pid=1568980) INFO 06-07 16:20:54 [default_loader.py:293] Loading weights took 13.69 seconds (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) (EngineCore_DP0 pid=1568976) INFO 06-07 16:20:54 [gpu_model_runner.py:4222] Model loading took 15.27 GiB memory and 14.966695 seconds (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) (EngineCore_DP0 pid=1568976) INFO 06-07 16:20:56 [gpu_worker.py:373] Available KV cache memory: 49.32 GiB (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) (EngineCore_DP0 pid=1568976) INFO 06-07 16:20:56 [kv_cache_utils.py:1307] GPU KV cache size: 359,104 tokens (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) (EngineCore_DP0 pid=1568976) INFO 06-07 16:20:56 [kv_cache_utils.py:1312] Maximum concurrency for 32,768 tokens per request: 10.96x (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) INFO 06-07 16:20:51 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s [repeated 2x across cluster] (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) (EngineCore_DP0 pid=1568976) INFO 06-07 16:20:57 [kernel_warmup.py:44] Skipping FlashInfer autotune because it is disabled. (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) (EngineCore_DP0 pid=1568976) INFO 06-07 16:20:57 [core.py:278] init engine (profile, create kv cache, warmup model) took 2.80 seconds (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) (EngineCore_DP0 pid=1568976) WARNING 06-07 16:20:57 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) (EngineCore_DP0 pid=1568976) INFO 06-07 16:20:57 [vllm.py:690] Asynchronous scheduling is enabled. (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) (EngineCore_DP0 pid=1568976) WARNING 06-07 16:20:57 [vllm.py:735] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) (EngineCore_DP0 pid=1568976) INFO 06-07 16:20:57 [vllm.py:846] Cudagraph is disabled under eager mode (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) (EngineCore_DP0 pid=1495920) INFO 06-07 16:20:52 [core.py:97] Initializing a V1 LLM engine (vdev) with config: model='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', speculative_config=None, tokenizer='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=32768, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False), seed=73, served_model_name=0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'level': None, 'mode': , 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'splitting_ops': [], 'compile_mm_encoder': False, 'compile_sizes': [], 'compile_ranges_split_points': [65536], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': False, 'fuse_act_quant': False, 'fuse_attn_quant': False, 'eliminate_noops': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': True, 'static_all_moe_layers': []} [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568271, ip=10.128.16.48) WARNING 06-07 16:20:57 [model.py:1350] Default vLLM sampling parameters have been overridden by the model's `generation_config.json`: `{'temperature': 0.6, 'top_k': 20, 'top_p': 0.95}`. If this is not intended, please relaunch vLLM instance with `--generation-config vllm`. (skyrl_entrypoint pid=1617955) [fd-monitor] [16:21:18] OK: 87 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:21:18] OK: RSS 1.64 GiB | node mem 171.9/858.0 GiB used (20.0%), avail 686.1 GiB (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) (EngineCore_DP0 pid=1568976) INFO 06-07 16:20:54 [default_loader.py:293] Loading weights took 13.74 seconds [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568416, ip=10.128.16.48) (EngineCore_DP0 pid=1568978) INFO 06-07 16:20:54 [gpu_model_runner.py:4222] Model loading took 15.27 GiB memory and 14.958987 seconds [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568416, ip=10.128.16.48) (EngineCore_DP0 pid=1568978) INFO 06-07 16:20:56 [gpu_worker.py:373] Available KV cache memory: 49.32 GiB [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568416, ip=10.128.16.48) (EngineCore_DP0 pid=1568978) INFO 06-07 16:20:56 [kv_cache_utils.py:1307] GPU KV cache size: 359,104 tokens [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568416, ip=10.128.16.48) (EngineCore_DP0 pid=1568978) INFO 06-07 16:20:56 [kv_cache_utils.py:1312] Maximum concurrency for 32,768 tokens per request: 10.96x [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568416, ip=10.128.16.48) (EngineCore_DP0 pid=1568978) INFO 06-07 16:20:57 [kernel_warmup.py:44] Skipping FlashInfer autotune because it is disabled. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568416, ip=10.128.16.48) (EngineCore_DP0 pid=1568978) INFO 06-07 16:20:57 [core.py:278] init engine (profile, create kv cache, warmup model) took 2.83 seconds [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568416, ip=10.128.16.48) (EngineCore_DP0 pid=1568978) WARNING 06-07 16:20:57 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568416, ip=10.128.16.48) (EngineCore_DP0 pid=1568978) INFO 06-07 16:20:57 [vllm.py:690] Asynchronous scheduling is enabled. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568416, ip=10.128.16.48) (EngineCore_DP0 pid=1568978) WARNING 06-07 16:20:57 [vllm.py:735] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568416, ip=10.128.16.48) (EngineCore_DP0 pid=1568978) INFO 06-07 16:20:57 [vllm.py:846] Cudagraph is disabled under eager mode [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1568417, ip=10.128.16.48) WARNING 06-07 16:20:57 [model.py:1350] Default vLLM sampling parameters have been overridden by the model's `generation_config.json`: `{'temperature': 0.6, 'top_k': 20, 'top_p': 0.95}`. If this is not intended, please relaunch vLLM instance with `--generation-config vllm`. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) (EngineCore_DP0 pid=1495916) INFO 06-07 16:22:41 [worker_base.py:289] Injected into for extended collective_rpc calls ['_apply_fp8_weight_loader_patches', '_is_fp8_model', '_quantize_weights_for_fp8', '_restore_param_subclasses', '_undo_param_subclasses', 'begin_weight_update', 'destroy_weights_update_group', 'end_weight_update', 'init_weight_update_communicator', 'load_weights', 'read_named_weights', 'set_numa_affinity', 'test_rpc'] (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) (EngineCore_DP0 pid=1495916) INFO 06-07 16:22:42 [parallel_state.py:1234] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://10.128.16.36:36041 backend=nccl (AsyncVLLMInferenceEngine pid=1495335, ip=10.128.16.36) (EngineCore_DP0 pid=1495936) INFO 06-07 16:22:42 [parallel_state.py:1445] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) (EngineCore_DP0 pid=1495916) INFO 06-07 16:22:42 [gpu_model_runner.py:4125] Starting to load model /e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6... (AsyncVLLMInferenceEngine pid=1204670, ip=10.128.16.43) (EngineCore_DP0 pid=1205535) INFO 06-07 16:22:51 [worker_base.py:289] Injected into for extended collective_rpc calls ['_apply_fp8_weight_loader_patches', '_is_fp8_model', '_quantize_weights_for_fp8', '_restore_param_subclasses', '_undo_param_subclasses', 'begin_weight_update', 'destroy_weights_update_group', 'end_weight_update', 'init_weight_update_communicator', 'load_weights', 'read_named_weights', 'set_numa_affinity', 'test_rpc'] [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) (EngineCore_DP0 pid=1495920) INFO 06-07 16:22:42 [parallel_state.py:1234] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://10.128.16.36:36069 backend=nccl [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) (EngineCore_DP0 pid=1495916) INFO 06-07 16:22:42 [parallel_state.py:1445] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) (EngineCore_DP0 pid=1495920) INFO 06-07 16:22:42 [gpu_model_runner.py:4125] Starting to load model /e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6... [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) (EngineCore_DP0 pid=1495916) INFO 06-07 16:23:09 [cuda.py:367] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION']. (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) (EngineCore_DP0 pid=1205528) INFO 06-07 16:22:51 [worker_base.py:289] Injected into for extended collective_rpc calls ['_apply_fp8_weight_loader_patches', '_is_fp8_model', '_quantize_weights_for_fp8', '_restore_param_subclasses', '_undo_param_subclasses', 'begin_weight_update', 'destroy_weights_update_group', 'end_weight_update', 'init_weight_update_communicator', 'load_weights', 'read_named_weights', 'set_numa_affinity', 'test_rpc'] [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) (EngineCore_DP0 pid=1205528) INFO 06-07 16:22:51 [parallel_state.py:1234] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://10.128.16.43:50323 backend=nccl [repeated 4x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) (EngineCore_DP0 pid=1205528) INFO 06-07 16:22:51 [parallel_state.py:1445] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A [repeated 4x across cluster] (skyrl_entrypoint pid=1617955) [fd-monitor] [16:23:18] OK: 87 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:23:18] OK: RSS 1.64 GiB | node mem 171.9/858.0 GiB used (20.0%), avail 686.1 GiB (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) (EngineCore_DP0 pid=1495920) INFO 06-07 16:23:09 [cuda.py:367] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION']. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1204670, ip=10.128.16.43) (EngineCore_DP0 pid=1205535) INFO 06-07 16:23:18 [gpu_model_runner.py:4125] Starting to load model /e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6... (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) (EngineCore_DP0 pid=1495916) INFO 06-07 16:23:23 [default_loader.py:293] Loading weights took 13.44 seconds (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) (EngineCore_DP0 pid=1205528) INFO 06-07 16:23:18 [gpu_model_runner.py:4125] Starting to load model /e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6... [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) (EngineCore_DP0 pid=1495916) INFO 06-07 16:23:23 [gpu_model_runner.py:4222] Model loading took 15.27 GiB memory and 40.381533 seconds (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) (EngineCore_DP0 pid=1495916) INFO 06-07 16:23:26 [gpu_worker.py:373] Available KV cache memory: 49.32 GiB (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) (EngineCore_DP0 pid=1495916) INFO 06-07 16:23:26 [kv_cache_utils.py:1307] GPU KV cache size: 359,104 tokens (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) (EngineCore_DP0 pid=1495916) INFO 06-07 16:23:26 [kv_cache_utils.py:1312] Maximum concurrency for 32,768 tokens per request: 10.96x (AsyncVLLMInferenceEngine pid=1495335, ip=10.128.16.36) (EngineCore_DP0 pid=1495936) INFO 06-07 16:23:26 [kernel_warmup.py:44] Skipping FlashInfer autotune because it is disabled. (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) (EngineCore_DP0 pid=1495916) INFO 06-07 16:23:26 [core.py:278] init engine (profile, create kv cache, warmup model) took 2.81 seconds (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) (EngineCore_DP0 pid=1495916) WARNING 06-07 16:23:26 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) (EngineCore_DP0 pid=1495916) INFO 06-07 16:23:26 [vllm.py:690] Asynchronous scheduling is enabled. (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) (EngineCore_DP0 pid=1495916) WARNING 06-07 16:23:26 [vllm.py:735] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. (AsyncVLLMInferenceEngine pid=1495134, ip=10.128.16.36) (EngineCore_DP0 pid=1495916) INFO 06-07 16:23:26 [vllm.py:846] Cudagraph is disabled under eager mode (AsyncVLLMInferenceEngine pid=1495335, ip=10.128.16.36) WARNING 06-07 16:23:27 [model.py:1350] Default vLLM sampling parameters have been overridden by the model's `generation_config.json`: `{'temperature': 0.6, 'top_k': 20, 'top_p': 0.95}`. If this is not intended, please relaunch vLLM instance with `--generation-config vllm`. (AsyncVLLMInferenceEngine pid=1204670, ip=10.128.16.43) (EngineCore_DP0 pid=1205535) INFO 06-07 16:23:45 [cuda.py:367] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION']. (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) (EngineCore_DP0 pid=1495920) INFO 06-07 16:23:23 [default_loader.py:293] Loading weights took 13.45 seconds [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) (EngineCore_DP0 pid=1495920) INFO 06-07 16:23:23 [gpu_model_runner.py:4222] Model loading took 15.27 GiB memory and 40.383910 seconds [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) (EngineCore_DP0 pid=1495920) INFO 06-07 16:23:25 [gpu_worker.py:373] Available KV cache memory: 49.32 GiB [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) (EngineCore_DP0 pid=1495920) INFO 06-07 16:23:25 [kv_cache_utils.py:1307] GPU KV cache size: 359,104 tokens [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) (EngineCore_DP0 pid=1495920) INFO 06-07 16:23:25 [kv_cache_utils.py:1312] Maximum concurrency for 32,768 tokens per request: 10.96x [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) (EngineCore_DP0 pid=1495920) INFO 06-07 16:23:26 [kernel_warmup.py:44] Skipping FlashInfer autotune because it is disabled. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) (EngineCore_DP0 pid=1495920) INFO 06-07 16:23:26 [core.py:278] init engine (profile, create kv cache, warmup model) took 2.81 seconds [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) (EngineCore_DP0 pid=1495920) WARNING 06-07 16:23:26 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) (EngineCore_DP0 pid=1495920) INFO 06-07 16:23:26 [vllm.py:690] Asynchronous scheduling is enabled. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) (EngineCore_DP0 pid=1495920) WARNING 06-07 16:23:26 [vllm.py:735] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) (EngineCore_DP0 pid=1495920) INFO 06-07 16:23:26 [vllm.py:846] Cudagraph is disabled under eager mode [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1495334, ip=10.128.16.36) WARNING 06-07 16:23:27 [model.py:1350] Default vLLM sampling parameters have been overridden by the model's `generation_config.json`: `{'temperature': 0.6, 'top_k': 20, 'top_p': 0.95}`. If this is not intended, please relaunch vLLM instance with `--generation-config vllm`. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) (EngineCore_DP0 pid=1205528) INFO 06-07 16:23:45 [cuda.py:367] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION']. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1204670, ip=10.128.16.43) (EngineCore_DP0 pid=1205535) INFO 06-07 16:23:59 [default_loader.py:293] Loading weights took 13.87 seconds (AsyncVLLMInferenceEngine pid=1204806, ip=10.128.16.43) (EngineCore_DP0 pid=1205537) INFO 06-07 16:23:59 [default_loader.py:293] Loading weights took 13.88 seconds (AsyncVLLMInferenceEngine pid=1204670, ip=10.128.16.43) (EngineCore_DP0 pid=1205535) INFO 06-07 16:23:59 [gpu_model_runner.py:4222] Model loading took 15.27 GiB memory and 40.839385 seconds (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) (EngineCore_DP0 pid=1205528) INFO 06-07 16:24:01 [gpu_worker.py:373] Available KV cache memory: 49.32 GiB (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) (EngineCore_DP0 pid=1205528) INFO 06-07 16:24:01 [kv_cache_utils.py:1307] GPU KV cache size: 359,104 tokens (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) (EngineCore_DP0 pid=1205528) INFO 06-07 16:24:01 [kv_cache_utils.py:1312] Maximum concurrency for 32,768 tokens per request: 10.96x (AsyncVLLMInferenceEngine pid=1204805, ip=10.128.16.43) (EngineCore_DP0 pid=1205529) INFO 06-07 16:24:02 [kernel_warmup.py:44] Skipping FlashInfer autotune because it is disabled. (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) (EngineCore_DP0 pid=1205528) INFO 06-07 16:24:02 [core.py:278] init engine (profile, create kv cache, warmup model) took 2.71 seconds (AsyncVLLMInferenceEngine pid=1204670, ip=10.128.16.43) (EngineCore_DP0 pid=1205535) WARNING 06-07 16:24:02 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 (AsyncVLLMInferenceEngine pid=1204670, ip=10.128.16.43) (EngineCore_DP0 pid=1205535) INFO 06-07 16:24:02 [vllm.py:690] Asynchronous scheduling is enabled. (AsyncVLLMInferenceEngine pid=1204670, ip=10.128.16.43) (EngineCore_DP0 pid=1205535) WARNING 06-07 16:24:02 [vllm.py:735] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. (AsyncVLLMInferenceEngine pid=1204670, ip=10.128.16.43) (EngineCore_DP0 pid=1205535) INFO 06-07 16:24:02 [vllm.py:846] Cudagraph is disabled under eager mode (AsyncVLLMInferenceEngine pid=1204806, ip=10.128.16.43) WARNING 06-07 16:24:03 [model.py:1350] Default vLLM sampling parameters have been overridden by the model's `generation_config.json`: `{'temperature': 0.6, 'top_k': 20, 'top_p': 0.95}`. If this is not intended, please relaunch vLLM instance with `--generation-config vllm`. (skyrl_entrypoint pid=1617955) [fd-monitor] [16:25:18] OK: 87 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:25:18] OK: RSS 1.64 GiB | node mem 171.9/858.0 GiB used (20.0%), avail 686.1 GiB (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) (EngineCore_DP0 pid=1205528) INFO 06-07 16:23:59 [default_loader.py:293] Loading weights took 13.87 seconds [repeated 2x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) (EngineCore_DP0 pid=1205528) INFO 06-07 16:23:59 [gpu_model_runner.py:4222] Model loading took 15.27 GiB memory and 40.851068 seconds [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1204806, ip=10.128.16.43) (EngineCore_DP0 pid=1205537) INFO 06-07 16:24:02 [gpu_worker.py:373] Available KV cache memory: 49.32 GiB [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1204806, ip=10.128.16.43) (EngineCore_DP0 pid=1205537) INFO 06-07 16:24:02 [kv_cache_utils.py:1307] GPU KV cache size: 359,104 tokens [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1204806, ip=10.128.16.43) (EngineCore_DP0 pid=1205537) INFO 06-07 16:24:02 [kv_cache_utils.py:1312] Maximum concurrency for 32,768 tokens per request: 10.96x [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1204806, ip=10.128.16.43) (EngineCore_DP0 pid=1205537) INFO 06-07 16:24:02 [kernel_warmup.py:44] Skipping FlashInfer autotune because it is disabled. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1204805, ip=10.128.16.43) (EngineCore_DP0 pid=1205529) INFO 06-07 16:24:02 [core.py:278] init engine (profile, create kv cache, warmup model) took 2.76 seconds [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) (EngineCore_DP0 pid=1205528) WARNING 06-07 16:24:02 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) (EngineCore_DP0 pid=1205528) INFO 06-07 16:24:02 [vllm.py:690] Asynchronous scheduling is enabled. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) (EngineCore_DP0 pid=1205528) WARNING 06-07 16:24:02 [vllm.py:735] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1205099, ip=10.128.16.43) (EngineCore_DP0 pid=1205528) INFO 06-07 16:24:02 [vllm.py:846] Cudagraph is disabled under eager mode [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1204670, ip=10.128.16.43) WARNING 06-07 16:24:03 [model.py:1350] Default vLLM sampling parameters have been overridden by the model's `generation_config.json`: `{'temperature': 0.6, 'top_k': 20, 'top_p': 0.95}`. If this is not intended, please relaunch vLLM instance with `--generation-config vllm`. [repeated 3x across cluster] (skyrl_entrypoint pid=1617955) [fd-monitor] [16:27:18] OK: 87 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:27:18] OK: RSS 1.64 GiB | node mem 171.9/858.0 GiB used (20.0%), avail 686.1 GiB (pid=1446434, ip=10.128.16.44) INFO 06-07 16:27:46 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) WARNING 06-07 16:29:09 [arg_utils.py:1256] The global random seed is set to 86. Since VLLM_ENABLE_V1_MULTIPROCESSING is set to False, this may affect the random state of the Python process that launched vLLM. (pid=1446661, ip=10.128.16.44) INFO 06-07 16:27:46 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) INFO 06-07 16:29:09 [model.py:529] Resolved architecture: Qwen3ForCausalLM (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) INFO 06-07 16:29:09 [model.py:1549] Using max model len 32768 (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) INFO 06-07 16:29:09 [arg_utils.py:1469] Using ray runtime env (env vars redacted): {'env_vars': {'NCCL_CUMEM_ENABLE': '***', 'NCCL_DEBUG': '***', 'NCCL_SOCKET_FAMILY': '***', 'NCCL_SOCKET_IFNAME': '***', 'RAY_ADDRESS': '***', 'VLLM_ALLOW_INSECURE_SERIALIZATION': '***', 'VLLM_ALLOW_RUNTIME_LORA_UPDATING': '***', 'VLLM_DISABLE_COMPILE_CACHE': '***', 'WANDB_API_KEY': '***'}} (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) INFO 06-07 16:29:09 [scheduler.py:224] Chunked prefill is enabled with max_num_batched_tokens=65536. (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) INFO 06-07 16:29:09 [vllm.py:690] Asynchronous scheduling is enabled. (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) WARNING 06-07 16:29:09 [vllm.py:728] Enforce eager set, overriding optimization level to -O0 (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) INFO 06-07 16:29:09 [vllm.py:846] Cudagraph is disabled under eager mode (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) WARNING 06-07 16:29:10 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) WARNING 06-07 16:29:10 [system_utils.py:140] We must use the `spawn` multiprocessing start method. Overriding VLLM_WORKER_MULTIPROC_METHOD to 'spawn'. See https://docs.vllm.ai/en/latest/usage/troubleshooting.html#python-multiprocessing for more information. Reasons: In a Ray actor and can only be spawned (skyrl_entrypoint pid=1617955) [fd-monitor] [16:29:18] OK: 87 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:29:18] OK: RSS 1.64 GiB | node mem 171.9/858.0 GiB used (20.0%), avail 686.1 GiB (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) WARNING 06-07 16:29:09 [arg_utils.py:1256] The global random seed is set to 88. Since VLLM_ENABLE_V1_MULTIPROCESSING is set to False, this may affect the random state of the Python process that launched vLLM. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) INFO 06-07 16:29:09 [model.py:529] Resolved architecture: Qwen3ForCausalLM [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) INFO 06-07 16:29:09 [model.py:1549] Using max model len 32768 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) INFO 06-07 16:29:09 [arg_utils.py:1469] Using ray runtime env (env vars redacted): {'env_vars': {'NCCL_CUMEM_ENABLE': '***', 'NCCL_DEBUG': '***', 'NCCL_SOCKET_FAMILY': '***', 'NCCL_SOCKET_IFNAME': '***', 'RAY_ADDRESS': '***', 'VLLM_ALLOW_INSECURE_SERIALIZATION': '***', 'VLLM_ALLOW_RUNTIME_LORA_UPDATING': '***', 'VLLM_DISABLE_COMPILE_CACHE': '***', 'WANDB_API_KEY': '***'}} [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) INFO 06-07 16:29:09 [scheduler.py:224] Chunked prefill is enabled with max_num_batched_tokens=65536. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) INFO 06-07 16:29:09 [vllm.py:690] Asynchronous scheduling is enabled. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) WARNING 06-07 16:29:09 [vllm.py:728] Enforce eager set, overriding optimization level to -O0 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) INFO 06-07 16:29:09 [vllm.py:846] Cudagraph is disabled under eager mode [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) WARNING 06-07 16:29:10 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) WARNING 06-07 16:29:10 [system_utils.py:140] We must use the `spawn` multiprocessing start method. Overriding VLLM_WORKER_MULTIPROC_METHOD to 'spawn'. See https://docs.vllm.ai/en/latest/usage/troubleshooting.html#python-multiprocessing for more information. Reasons: In a Ray actor and can only be spawned [repeated 3x across cluster] (skyrl_entrypoint pid=1617955) [fd-monitor] [16:31:18] OK: 87 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:31:18] OK: RSS 1.64 GiB | node mem 171.9/858.0 GiB used (20.0%), avail 686.1 GiB (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) INFO 06-07 16:32:28 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) (EngineCore_DP0 pid=1447387) INFO 06-07 16:32:29 [core.py:97] Initializing a V1 LLM engine (vdev) with config: model='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', speculative_config=None, tokenizer='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=32768, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False), seed=86, served_model_name=0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'level': None, 'mode': , 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'splitting_ops': [], 'compile_mm_encoder': False, 'compile_sizes': [], 'compile_ranges_split_points': [65536], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': False, 'fuse_act_quant': False, 'fuse_attn_quant': False, 'eliminate_noops': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': True, 'static_all_moe_layers': []} (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) (EngineCore_DP0 pid=1447387) INFO 06-07 16:33:01 [worker_base.py:289] Injected into for extended collective_rpc calls ['_apply_fp8_weight_loader_patches', '_is_fp8_model', '_quantize_weights_for_fp8', '_restore_param_subclasses', '_undo_param_subclasses', 'begin_weight_update', 'destroy_weights_update_group', 'end_weight_update', 'init_weight_update_communicator', 'load_weights', 'read_named_weights', 'set_numa_affinity', 'test_rpc'] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) INFO 06-07 16:32:28 [pynccl.py:178] pynccl trace buffer enabled: size=10000 dump_dir=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps flush_interval=0 s [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) (EngineCore_DP0 pid=1447390) INFO 06-07 16:32:29 [core.py:97] Initializing a V1 LLM engine (vdev) with config: model='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', speculative_config=None, tokenizer='/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.bfloat16, max_seq_len=32768, download_dir=None, load_format=auto, tensor_parallel_size=1, pipeline_parallel_size=1, data_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=True, enable_return_routed_experts=False, kv_cache_dtype=auto, device_config=cuda, structured_outputs_config=StructuredOutputsConfig(backend='auto', disable_fallback=False, disable_any_whitespace=False, disable_additional_properties=False, reasoning_parser='', reasoning_parser_plugin='', enable_in_reasoning=False), observability_config=ObservabilityConfig(show_hidden_metrics_for_version=None, otlp_traces_endpoint=None, collect_detailed_traces=None, kv_cache_metrics=False, kv_cache_metrics_sample=0.01, cudagraph_metrics=False, enable_layerwise_nvtx_tracing=False, enable_mfu_metrics=False, enable_mm_processor_stats=False, enable_logging_iteration_details=False), seed=88, served_model_name=0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6, enable_prefix_caching=True, enable_chunked_prefill=True, pooler_config=None, compilation_config={'level': None, 'mode': , 'debug_dump_path': None, 'cache_dir': '', 'compile_cache_save_format': 'binary', 'backend': 'inductor', 'custom_ops': ['all'], 'splitting_ops': [], 'compile_mm_encoder': False, 'compile_sizes': [], 'compile_ranges_split_points': [65536], 'inductor_compile_config': {'enable_auto_functionalized_v2': False, 'combo_kernels': True, 'benchmark_combo_kernel': True}, 'inductor_passes': {}, 'cudagraph_mode': , 'cudagraph_num_of_warmups': 0, 'cudagraph_capture_sizes': [], 'cudagraph_copy_inputs': False, 'cudagraph_specialize_lora': True, 'use_inductor_graph_partition': False, 'pass_config': {'fuse_norm_quant': False, 'fuse_act_quant': False, 'fuse_attn_quant': False, 'eliminate_noops': False, 'enable_sp': False, 'fuse_gemm_comms': False, 'fuse_allreduce_rms': False, 'fuse_act_padding': False}, 'max_cudagraph_capture_size': 0, 'dynamic_shapes_config': {'type': , 'evaluate_guards': False, 'assume_32_bit_indexing': False}, 'local_cache_dir': None, 'fast_moe_cold_start': True, 'static_all_moe_layers': []} [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) (EngineCore_DP0 pid=1447387) INFO 06-07 16:33:02 [parallel_state.py:1234] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://10.128.16.44:57389 backend=nccl (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) (EngineCore_DP0 pid=1447387) INFO 06-07 16:33:02 [parallel_state.py:1445] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A (AsyncVLLMInferenceEngine pid=1446659, ip=10.128.16.44) (EngineCore_DP0 pid=1447391) INFO 06-07 16:33:02 [gpu_model_runner.py:4125] Starting to load model /e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6... (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) (EngineCore_DP0 pid=1447387) INFO 06-07 16:33:04 [cuda.py:367] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION']. (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) (EngineCore_DP0 pid=1447387) INFO 06-07 16:33:18 [default_loader.py:293] Loading weights took 13.72 seconds (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) (EngineCore_DP0 pid=1447390) INFO 06-07 16:33:01 [worker_base.py:289] Injected into for extended collective_rpc calls ['_apply_fp8_weight_loader_patches', '_is_fp8_model', '_quantize_weights_for_fp8', '_restore_param_subclasses', '_undo_param_subclasses', 'begin_weight_update', 'destroy_weights_update_group', 'end_weight_update', 'init_weight_update_communicator', 'load_weights', 'read_named_weights', 'set_numa_affinity', 'test_rpc'] [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) (EngineCore_DP0 pid=1447390) INFO 06-07 16:33:02 [parallel_state.py:1234] world_size=1 rank=0 local_rank=0 distributed_init_method=tcp://10.128.16.44:48017 backend=nccl [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) (EngineCore_DP0 pid=1447390) INFO 06-07 16:33:02 [parallel_state.py:1445] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, PCP rank 0, TP rank 0, EP rank N/A [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) (EngineCore_DP0 pid=1447390) INFO 06-07 16:33:02 [gpu_model_runner.py:4125] Starting to load model /e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6... [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) (EngineCore_DP0 pid=1447390) INFO 06-07 16:33:04 [cuda.py:367] Using FLASH_ATTN attention backend out of potential backends: ['FLASH_ATTN', 'FLASHINFER', 'TRITON_ATTN', 'FLEX_ATTENTION']. [repeated 3x across cluster] (skyrl_entrypoint pid=1617955) [fd-monitor] [16:33:18] OK: 87 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:33:18] OK: RSS 1.64 GiB | node mem 171.8/858.0 GiB used (20.0%), avail 686.2 GiB (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) (EngineCore_DP0 pid=1447387) INFO 06-07 16:33:18 [gpu_model_runner.py:4222] Model loading took 15.27 GiB memory and 15.118078 seconds (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) (EngineCore_DP0 pid=1447387) INFO 06-07 16:33:20 [gpu_worker.py:373] Available KV cache memory: 49.32 GiB (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) (EngineCore_DP0 pid=1447387) INFO 06-07 16:33:20 [kv_cache_utils.py:1307] GPU KV cache size: 359,104 tokens (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) (EngineCore_DP0 pid=1447387) INFO 06-07 16:33:20 [kv_cache_utils.py:1312] Maximum concurrency for 32,768 tokens per request: 10.96x (AsyncVLLMInferenceEngine pid=1446659, ip=10.128.16.44) (EngineCore_DP0 pid=1447391) INFO 06-07 16:33:21 [kernel_warmup.py:44] Skipping FlashInfer autotune because it is disabled. (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) (EngineCore_DP0 pid=1447387) INFO 06-07 16:33:21 [core.py:278] init engine (profile, create kv cache, warmup model) took 2.80 seconds (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) (EngineCore_DP0 pid=1447387) WARNING 06-07 16:33:21 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) (EngineCore_DP0 pid=1447387) INFO 06-07 16:33:21 [vllm.py:690] Asynchronous scheduling is enabled. (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) (EngineCore_DP0 pid=1447387) WARNING 06-07 16:33:21 [vllm.py:735] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. (AsyncVLLMInferenceEngine pid=1446434, ip=10.128.16.44) (EngineCore_DP0 pid=1447387) INFO 06-07 16:33:21 [vllm.py:846] Cudagraph is disabled under eager mode (AsyncVLLMInferenceEngine pid=1446659, ip=10.128.16.44) WARNING 06-07 16:33:21 [model.py:1350] Default vLLM sampling parameters have been overridden by the model's `generation_config.json`: `{'temperature': 0.6, 'top_k': 20, 'top_p': 0.95}`. If this is not intended, please relaunch vLLM instance with `--generation-config vllm`. (skyrl_entrypoint pid=1617955) [fd-monitor] [16:35:18] OK: 87 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:35:18] OK: RSS 1.64 GiB | node mem 171.8/858.0 GiB used (20.0%), avail 686.2 GiB (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) (EngineCore_DP0 pid=1447390) INFO 06-07 16:33:18 [default_loader.py:293] Loading weights took 13.71 seconds [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) (EngineCore_DP0 pid=1447390) INFO 06-07 16:33:18 [gpu_model_runner.py:4222] Model loading took 15.27 GiB memory and 15.118351 seconds [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446660, ip=10.128.16.44) (EngineCore_DP0 pid=1447388) INFO 06-07 16:33:20 [gpu_worker.py:373] Available KV cache memory: 49.32 GiB [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446660, ip=10.128.16.44) (EngineCore_DP0 pid=1447388) INFO 06-07 16:33:20 [kv_cache_utils.py:1307] GPU KV cache size: 359,104 tokens [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446660, ip=10.128.16.44) (EngineCore_DP0 pid=1447388) INFO 06-07 16:33:20 [kv_cache_utils.py:1312] Maximum concurrency for 32,768 tokens per request: 10.96x [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) (EngineCore_DP0 pid=1447390) INFO 06-07 16:33:21 [kernel_warmup.py:44] Skipping FlashInfer autotune because it is disabled. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) (EngineCore_DP0 pid=1447390) INFO 06-07 16:33:21 [core.py:278] init engine (profile, create kv cache, warmup model) took 2.81 seconds [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) (EngineCore_DP0 pid=1447390) WARNING 06-07 16:33:21 [serial_utils.py:57] Allowing insecure serialization using pickle due to VLLM_ALLOW_INSECURE_SERIALIZATION=1 [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) (EngineCore_DP0 pid=1447390) INFO 06-07 16:33:21 [vllm.py:690] Asynchronous scheduling is enabled. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) (EngineCore_DP0 pid=1447390) WARNING 06-07 16:33:21 [vllm.py:735] Inductor compilation was disabled by user settings, optimizations settings that are only active during inductor compilation will be ignored. [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) (EngineCore_DP0 pid=1447390) INFO 06-07 16:33:21 [vllm.py:846] Cudagraph is disabled under eager mode [repeated 3x across cluster] (AsyncVLLMInferenceEngine pid=1446661, ip=10.128.16.44) WARNING 06-07 16:33:22 [model.py:1350] Default vLLM sampling parameters have been overridden by the model's `generation_config.json`: `{'temperature': 0.6, 'top_k': 20, 'top_p': 0.95}`. If this is not intended, please relaunch vLLM instance with `--generation-config vllm`. [repeated 3x across cluster] (skyrl_entrypoint pid=1617955) [fd-monitor] [16:37:18] OK: 85 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:37:18] OK: RSS 1.64 GiB | node mem 171.8/858.0 GiB used (20.0%), avail 686.2 GiB (skyrl_entrypoint pid=1617955) [fd-monitor] [16:39:18] OK: 85 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:39:18] OK: RSS 1.64 GiB | node mem 171.8/858.0 GiB used (20.0%), avail 686.2 GiB (skyrl_entrypoint pid=1617955) [fd-monitor] [16:41:18] OK: 85 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:41:18] OK: RSS 1.64 GiB | node mem 171.8/858.0 GiB used (20.0%), avail 686.2 GiB (skyrl_entrypoint pid=1617955) [fd-monitor] [16:43:18] OK: 85 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:43:18] OK: RSS 1.64 GiB | node mem 171.7/858.0 GiB used (20.0%), avail 686.3 GiB (pid=3137116, ip=10.128.16.135) ⚙️ Running in WANDB offline mode (skyrl_entrypoint pid=1617955) [fd-monitor] [16:45:18] OK: 80 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:45:18] OK: RSS 1.64 GiB | node mem 173.1/858.0 GiB used (20.2%), avail 684.9 GiB (pid=3138189, ip=10.128.16.135) ⚙️ Running in WANDB offline mode (skyrl_entrypoint pid=1617955) [fd-monitor] [16:47:18] OK: 81 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:47:18] OK: RSS 1.64 GiB | node mem 173.1/858.0 GiB used (20.2%), avail 684.9 GiB (pid=3138187, ip=10.128.16.135) ⚙️ Running in WANDB offline mode [repeated 2x across cluster] (skyrl_entrypoint pid=1617955) [fd-monitor] [16:49:18] OK: 81 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:49:18] OK: RSS 1.64 GiB | node mem 173.1/858.0 GiB used (20.2%), avail 684.8 GiB (skyrl_entrypoint pid=1617955) [fd-monitor] [16:51:18] OK: 82 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:51:18] OK: RSS 1.64 GiB | node mem 173.2/858.0 GiB used (20.2%), avail 684.8 GiB (skyrl_entrypoint pid=1617955) [fd-monitor] [16:53:18] OK: 81 / 131,072 FDs open (0.1% of soft limit, hard limit: 131,072) (skyrl_entrypoint pid=1617955) [fd-monitor] [16:53:18] OK: RSS 1.64 GiB | node mem 173.2/858.0 GiB used (20.2%), avail 684.7 GiB (pid=1620366) ⚙️ Running in WANDB offline mode (pid=1620365) ⚙️ Running in WANDB offline mode [repeated 3x across cluster] Error executing job with overrides: ['+terminal_bench_config=terminal_bench', 'trainer.strategy=fsdp2', 'trainer.algorithm.advantage_estimator=rloo_n', 'trainer.algorithm.use_kl_loss=false', 'trainer.algorithm.kl_loss_coef=0.0', 'trainer.algorithm.eps_clip_low=0.2', 'trainer.algorithm.eps_clip_high=0.05', 'trainer.algorithm.loss_reduction=sequence_mean', 'trainer.algorithm.use_tis=true', 'trainer.algorithm.tis_imp_ratio_cap=2.0', 'trainer.epochs=2', 'trainer.max_steps=80', 'trainer.update_epochs_per_batch=1', 'trainer.train_batch_size=64', 'trainer.policy_mini_batch_size=64', 'trainer.eval_batch_size=64', 'trainer.micro_forward_batch_size_per_gpu=4', 'trainer.micro_train_batch_size_per_gpu=1', 'trainer.max_prompt_length=999999', 'trainer.eval_interval=999999', 'trainer.eval_before_train=false', 'trainer.ckpt_interval=2', 'trainer.resume_mode=latest', 'trainer.hf_save_interval=5', '++trainer.hf_hub_repo_id=laion/ablation-pymethods2test-seqmean-arm0-tis', '++trainer.hf_hub_private=false', '++trainer.hf_hub_revision=main', '++trainer.enable_db_registration=false', 'trainer.project_name=OpenThoughts-Agent', 'trainer.log_level=INFO', 'trainer.tracker_commit_each_step=true', 'trainer.logger=console', 'trainer.run_name=ablation-pymethods2test-seqmean-arm0-tis', 'trainer.ckpt_path=/e/data1/datasets/playground/ot-baf/ablation-pymethods2test-seqmean-arm0-tis/ablation-pymethods2test-seqmean-arm0-tis/checkpoints', 'trainer.export_path=/e/data1/datasets/playground/ot-baf/ablation-pymethods2test-seqmean-arm0-tis/ablation-pymethods2test-seqmean-arm0-tis/exports', 'trainer.policy.optimizer_config.lr=8e-6', 'trainer.policy.optimizer_config.weight_decay=0.0', 'trainer.policy.optimizer_config.adam_betas=[0.9,0.999]', 'trainer.policy.optimizer_config.max_grad_norm=0.9', 'trainer.policy.fsdp_config.cpu_offload=false', 'trainer.policy.fsdp_config.reshard_after_forward=true', 'trainer.policy.fsdp_config.fsdp_size=4', 'trainer.policy.model.path=/e/data1/datasets/playground/ot-baf/hf_hub/models--laion--GLM-4_7-swesmith-sandboxes-with_tests-oracle_verified_120s-maxeps-131k-fixthink/snapshots/0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', 'trainer.ref.fsdp_config.cpu_offload=false', 'trainer.ref.fsdp_config.reshard_after_forward=true', 'trainer.ref.fsdp_config.fsdp_size=4', 'trainer.placement.colocate_all=false', 'trainer.placement.policy_num_nodes=2', 'trainer.placement.ref_num_nodes=2', 'trainer.placement.policy_num_gpus_per_node=4', 'trainer.placement.ref_num_gpus_per_node=4', 'trainer.fully_async.max_staleness_steps=16', 'trainer.fully_async.num_parallel_generation_workers=338', 'generator.backend=vllm', 'generator.timeout_multiplier=1.0', 'generator.model_dtype=bfloat16', 'generator.inference_engine_tensor_parallel_size=1', 'generator.num_inference_engines=48', 'generator.n_samples_per_prompt=8', 'generator.eval_n_samples_per_prompt=8', 'generator.gpu_memory_utilization=0.75', 'generator.max_num_seqs=24', 'generator.max_num_batched_tokens=65536', 'generator.enable_prefix_caching=true', 'generator.enable_chunked_prefill=true', 'generator.run_engines_locally=true', 'generator.weight_sync_backend=nccl', 'generator.async_engine=true', 'generator.batched=false', 'generator.enable_http_endpoint=true', 'generator.enable_ray_prometheus_stats=false', 'generator.vllm_stats_interval=1', 'generator.append_eos_token_after_stop_str_in_multi_turn=true', 'generator.max_turns=999999', 'generator.sampling_params.max_generate_length=4096', 'generator.sampling_params.temperature=0.7', 'generator.sampling_params.top_p=0.95', 'generator.sampling_params.top_k=20', '++generator.engine_init_kwargs.max_model_len=32768', '++generator.engine_init_kwargs.custom_chat_template_chat_completion_path=chat_templates/qwen3_thinking_acc.jinja2', '++generator.engine_init_kwargs.served_model_name=0e3bff0c4e51f6b9ec0713b98b9eec36efb91cc6', 'data.train_data=["/e/scratch/jureap59/feuer1/tasks/exp_rpt_pymethods2test-large"]', 'data.val_data=[]', '+terminal_bench_config.trials_dir=/e/data1/datasets/playground/ot-baf/ablation-pymethods2test-seqmean-arm0-tis/ablation-pymethods2test-seqmean-arm0-tis/trace_jobs', '+terminal_bench_config.harbor.name=terminus-2', '+terminal_bench_config.harbor.max_episodes=999999', '+terminal_bench_config.harbor.enable_summarize=false', '+terminal_bench_config.harbor.store_all_messages=true', '+terminal_bench_config.harbor.trajectory_config.raw_content=true', '+terminal_bench_config.harbor.enable_episode_logging=false', '+terminal_bench_config.harbor.record_terminal_session=false', '+terminal_bench_config.harbor.enable_pane_logging=false', '+terminal_bench_config.harbor.strict_json_parser=true', '+terminal_bench_config.harbor.interleaved_thinking=true', '+terminal_bench_config.harbor.extra_body.chat_template_kwargs.enable_thinking=true', '+terminal_bench_config.harbor.override_timeout_sec=900', '+terminal_bench_config.harbor.override_cpus=1', '+terminal_bench_config.harbor.override_memory_mb=2048', '+terminal_bench_config.harbor.override_storage_mb=2048', '+terminal_bench_config.harbor.auto_snapshot=true', '+terminal_bench_config.harbor.verifier_override_timeout_sec=120', '+terminal_bench_config.harbor.max_retries=3', '+terminal_bench_config.harbor.min_wait_sec=60.0', '+terminal_bench_config.harbor.max_wait_sec=600.0', '+terminal_bench_config.harbor.wait_multiplier=2.0', '+terminal_bench_config.harbor.exclude_exceptions=["VerifierTimeoutError","VerifierRuntimeError","RewardFileNotFoundError","RewardFileEmptyError","VerifierOutputParseError"]', '+terminal_bench_config.harbor.n_concurrent_trials=675', '+terminal_bench_config.harbor.log_level=INFO', '+terminal_bench_config.harbor.enable_reward_shaping=false', '+terminal_bench_config.harbor.collect_rollout_details=true', '+terminal_bench_config.harbor.enable_error_classification=true', '+terminal_bench_config.harbor.mask_exceptions=["DaytonaError","EnvironmentStartTimeoutError","NetworkError","ConnectionError","RewardFileNotFoundError","RewardFileEmptyError","AgentEnvironmentTimeoutError","ContextLengthExceededError"]', '+terminal_bench_config.harbor.default_error_treatment=zero', '+terminal_bench_config.harbor.passthrough_exceptions=["AgentTimeoutError"]', '+terminal_bench_config.harbor.zero_exceptions=[]', '+terminal_bench_config.model_info.max_input_tokens=32000', '+terminal_bench_config.model_info.max_output_tokens=4096', '+terminal_bench_config.archiving.enabled=false', '+terminal_bench_config.trace_upload.enabled=true', '+terminal_bench_config.trace_upload.repo_org=DCAgent', '+terminal_bench_config.trace_upload.episodes=last', '+terminal_bench_config.trace_upload.dataset_type=SFT', '+terminal_bench_config.trace_upload.cleanup=true'] Traceback (most recent call last): File "", line 198, in _run_module_as_main File "", line 88, in _run_code File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/SkyRL/skyrl-train/examples/terminal_bench/entrypoints/main_tbench.py", line 142, in main() File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/hydra/main.py", line 94, in decorated_main _run_hydra( File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/hydra/_internal/utils.py", line 394, in _run_hydra _run_app( File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/hydra/_internal/utils.py", line 457, in _run_app run_and_report( File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/hydra/_internal/utils.py", line 223, in run_and_report raise ex File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/hydra/_internal/utils.py", line 220, in run_and_report return func() ^^^^^^ File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/hydra/_internal/utils.py", line 458, in lambda: hydra.run( ^^^^^^^^^^ File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/hydra/_internal/hydra.py", line 132, in run _ = ret.return_value ^^^^^^^^^^^^^^^^ File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/hydra/core/utils.py", line 260, in return_value raise self._return_value File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/hydra/core/utils.py", line 186, in run_job ret.return_value = task_function(task_cfg) ^^^^^^^^^^^^^^^^^^^^^^^ File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/SkyRL/skyrl-train/examples/terminal_bench/entrypoints/main_tbench.py", line 132, in main ray.get(skyrl_entrypoint.remote(cfg)) File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/ray/_private/auto_init_hook.py", line 22, in auto_init_wrapper return fn(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^ File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/ray/_private/client_mode_hook.py", line 104, in wrapper return func(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^ File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/ray/_private/worker.py", line 2967, in get values, debugger_breakpoint = worker.get_objects( ^^^^^^^^^^^^^^^^^^^ File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/ray/_private/worker.py", line 1015, in get_objects raise value.as_instanceof_cause() ray.exceptions.RayTaskError(DistStoreError): ray::skyrl_entrypoint() (pid=1617955, ip=10.128.16.35) File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/SkyRL/skyrl-train/examples/terminal_bench/entrypoints/main_tbench.py", line 105, in skyrl_entrypoint exp.run() File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/SkyRL/skyrl-train/skyrl_train/entrypoints/main_base.py", line 481, in run trainer = self._setup_trainer() ^^^^^^^^^^^^^^^^^^^^^ File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/SkyRL/skyrl-train/skyrl_train/entrypoints/main_base.py", line 450, in _setup_trainer trainer.build_models(PolicyWorker, CriticWorker, RefWorker, policy_pg=self.policy_pg) File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/SkyRL/skyrl-train/skyrl_train/trainer.py", line 757, in build_models File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/SkyRL/skyrl-train/skyrl_train/workers/worker.py", line 511, in __init__ self._initiate_actors(pg, num_gpus_per_actor) File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/SkyRL/skyrl-train/skyrl_train/workers/worker.py", line 637, in _initiate_actors ray.get([actor.init_worker_process_group.remote() for actor in self._actor_handlers]) ^^^^^^^^^^^^^^^^^^^ ^^^^^^^^^^^^^^^^^^^^^ ^^^^^^^^^^^^^^^^^^^ ray.exceptions.RayTaskError(DistStoreError): ray::FSDPPolicyWorkerBase.init_worker_process_group() (pid=3137116, ip=10.128.16.135, actor_id=538f00c3532d78f16f4d391d02000000, repr=) File "/e/scratch/jureap59/feuer1/miniforge3/lib/python3.12/concurrent/futures/_base.py", line 449, in result return self.__get_result() ^^^^^^^^^^^^^^^^^^^ File "/e/scratch/jureap59/feuer1/miniforge3/lib/python3.12/concurrent/futures/_base.py", line 401, in __get_result raise self._exception ^^^^^^^^^^^^^^^^^^^^^ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/SkyRL/skyrl-train/skyrl_train/workers/worker.py", line 149, in init_worker_process_group torch.distributed.init_process_group( File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/distributed/c10d_logger.py", line 81, in wrapper return func(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^ File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/distributed/c10d_logger.py", line 95, in wrapper func_return = func(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^ File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/distributed/distributed_c10d.py", line 1762, in init_process_group store, rank, world_size = next(rendezvous_iterator) ^^^^^^^^^^^^^^^^^^^^^^^^^ File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/distributed/rendezvous.py", line 278, in _env_rendezvous_handler store = _create_c10d_store( ^^^^^^^^^^^^^^^^^^^ File "/e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/distributed/rendezvous.py", line 198, in _create_c10d_store return TCPStore( ^^^^^^^^^ torch.distributed.DistStoreError: Timed out after 601 seconds waiting for clients. 4/8 clients joined. Exception raised from waitForWorkers at /pytorch/torch/csrc/distributed/c10d/TCPStore.cpp:396 (most recent call first): frame #0: c10::Error::Error(c10::SourceLocation, std::__cxx11::basic_string, std::allocator >) + 0xb0 (0x4000ab70c700 in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/lib/libc10.so) frame #1: + 0x5e9c9c0 (0x400b025fc9c0 in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/lib/libtorch_cpu.so) frame #2: c10d::TCPStore::waitForWorkers() + 0x350 (0x400b02691410 in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/lib/libtorch_cpu.so) frame #3: c10d::TCPStore::TCPStore(std::__cxx11::basic_string, std::allocator >, c10d::TCPStoreOptions const&) + 0x468 (0x400b026918c8 in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/lib/libtorch_cpu.so) frame #4: + 0x109a094 (0x400afc4fa094 in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/lib/libtorch_python.so) frame #5: + 0x113236c (0x400afc59236c in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/lib/libtorch_python.so) frame #6: + 0x5d6d60 (0x400afba36d60 in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/lib/libtorch_python.so) frame #7: + 0x1b7a38 (0xaaaad83d7a38 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #8: _PyObject_MakeTpCall + 0x98 (0xaaaad8385db8 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #9: + 0x169f50 (0xaaaad8389f50 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #10: + 0x1682e4 (0xaaaad83882e4 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #11: + 0x1e0ce8 (0xaaaad8400ce8 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #12: + 0x1d7ddc (0xaaaad83f7ddc in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #13: + 0x646b0c (0x400afbaa6b0c in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/torch/lib/libtorch_python.so) frame #14: _PyObject_MakeTpCall + 0x98 (0xaaaad8385db8 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #15: _PyEval_EvalFrameDefault + 0x280c (0xaaaad848ae54 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #16: + 0x1808c0 (0xaaaad83a08c0 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #17: + 0x182bf8 (0xaaaad83a2bf8 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #18: + 0x25fd30 (0xaaaad847fd30 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #19: + 0x1b7d20 (0xaaaad83d7d20 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #20: PyObject_Vectorcall + 0x54 (0xaaaad83860e4 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #21: _PyEval_EvalFrameDefault + 0x280c (0xaaaad848ae54 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #22: + 0x1808c0 (0xaaaad83a08c0 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #23: + 0x1835c8 (0xaaaad83a35c8 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #24: + 0x84b708 (0x40002c15b708 in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/ray/_raylet.so) frame #25: + 0x865fe4 (0x40002c175fe4 in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/ray/_raylet.so) frame #26: + 0x823470 (0x40002c133470 in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/ray/_raylet.so) frame #27: + 0x830110 (0x40002c140110 in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/ray/_raylet.so) frame #28: + 0x174a60 (0xaaaad8394a60 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #29: PyObject_VectorcallMethod + 0xa4 (0xaaaad8386270 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #30: PyIter_Send + 0xbc (0xaaaad836b62c in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #31: + 0xaa04 (0x40002e3caa04 in /e/scratch/jureap59/feuer1/miniforge3/lib/python3.12/lib-dynload/_asyncio.cpython-312-aarch64-linux-gnu.so) frame #32: + 0xbcc4 (0x40002e3cbcc4 in /e/scratch/jureap59/feuer1/miniforge3/lib/python3.12/lib-dynload/_asyncio.cpython-312-aarch64-linux-gnu.so) frame #33: _PyObject_MakeTpCall + 0x98 (0xaaaad8385db8 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #34: + 0x28cfac (0xaaaad84acfac in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #35: + 0x1b7b68 (0xaaaad83d7b68 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #36: _PyEval_EvalFrameDefault + 0x52ac (0xaaaad848d8f4 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #37: + 0x83c564 (0x40002c14c564 in /e/scratch/jureap59/feuer1/OpenThoughts-Agent/envs/rl/lib/python3.12/site-packages/ray/_raylet.so) frame #38: _PyEval_EvalFrameDefault + 0x52ac (0xaaaad848d8f4 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #39: + 0x169f88 (0xaaaad8389f88 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #40: + 0x36002c (0xaaaad858002c in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #41: + 0x2e4024 (0xaaaad8504024 in ray::FSDPPolicyWorkerBase.init_worker_process_group) frame #42: + 0x80e00 (0x40002b080e00 in /lib64/libc.so.6) frame #43: + 0xeb49c (0x40002b0eb49c in /lib64/libc.so.6) Stopping Ray cluster... Warning: Failed to stop Ray on jpbo-001-42: Command '['srun', '--export=ALL,WANDB_MODE=offline,GLOO_USE_IPV6=0,NCCL_SOCKET_FAMILY=AF_INET,VLLM_FORCE_IPV4=1,VLLM_SKIP_FLAG_DISCOVERY=1,SKYRL_ENABLE_NUMA_AFFINITY=1,DISABLE_AIOHTTP_TRANSPORT=True,VLLM_ALLREDUCE_USE_SYMM_MEM=0,TORCH_CUDNN_SDPA_ENABLED=0,PYTHONFAULTHANDLER=1,TORCH_NCCL_ASYNC_ERROR_HANDLING=1,TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC=1800,TORCH_NCCL_BLOCKING_WAIT_TIMEOUT_MS=1800000,VLLM_MQ_MAX_CHUNKS=240,OT_AGENT_RAY_LOG_DIR=/e/data1/datasets/playground/ot-baf/experiments/_ray_logs,TORCH_NCCL_TRACE_BUFFER_SIZE=10000,TORCH_FR_BUFFER_SIZE=10000,TORCH_NCCL_DESYNC_DEBUG=1,TORCH_NCCL_DEBUG_INFO_TEMP_FILE=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps/nccl_trace,VLLM_PYNCCL_TRACE_BUFFER_SIZE=10000,VLLM_PYNCCL_TRACE_DUMP_DIR=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps/nccl_trace,VLLM_RAY_EXTRA_ENV_VAR_PREFIXES_TO_COPY=TORCH_NCCL_,TORCH_FR_,VLLM_PYNCCL_', '--nodes=1', '--ntasks=1', '--overlap', '--cpu-bind=none', '-w', 'jpbo-001-42', 'bash', '-c', 'unset LD_PRELOAD PROXYCHAINS_CONF_FILE 2>/dev/null; ray stop --force']' timed out after 30 seconds Warning: Failed to stop Ray on jpbo-001-43: Command '['srun', '--export=ALL,WANDB_MODE=offline,GLOO_USE_IPV6=0,NCCL_SOCKET_FAMILY=AF_INET,VLLM_FORCE_IPV4=1,VLLM_SKIP_FLAG_DISCOVERY=1,SKYRL_ENABLE_NUMA_AFFINITY=1,DISABLE_AIOHTTP_TRANSPORT=True,VLLM_ALLREDUCE_USE_SYMM_MEM=0,TORCH_CUDNN_SDPA_ENABLED=0,PYTHONFAULTHANDLER=1,TORCH_NCCL_ASYNC_ERROR_HANDLING=1,TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC=1800,TORCH_NCCL_BLOCKING_WAIT_TIMEOUT_MS=1800000,VLLM_MQ_MAX_CHUNKS=240,OT_AGENT_RAY_LOG_DIR=/e/data1/datasets/playground/ot-baf/experiments/_ray_logs,TORCH_NCCL_TRACE_BUFFER_SIZE=10000,TORCH_FR_BUFFER_SIZE=10000,TORCH_NCCL_DESYNC_DEBUG=1,TORCH_NCCL_DEBUG_INFO_TEMP_FILE=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps/nccl_trace,VLLM_PYNCCL_TRACE_BUFFER_SIZE=10000,VLLM_PYNCCL_TRACE_DUMP_DIR=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps/nccl_trace,VLLM_RAY_EXTRA_ENV_VAR_PREFIXES_TO_COPY=TORCH_NCCL_,TORCH_FR_,VLLM_PYNCCL_', '--nodes=1', '--ntasks=1', '--overlap', '--cpu-bind=none', '-w', 'jpbo-001-43', 'bash', '-c', 'unset LD_PRELOAD PROXYCHAINS_CONF_FILE 2>/dev/null; ray stop --force']' timed out after 30 seconds Warning: Failed to stop Ray on jpbo-001-45: Command '['srun', '--export=ALL,WANDB_MODE=offline,GLOO_USE_IPV6=0,NCCL_SOCKET_FAMILY=AF_INET,VLLM_FORCE_IPV4=1,VLLM_SKIP_FLAG_DISCOVERY=1,SKYRL_ENABLE_NUMA_AFFINITY=1,DISABLE_AIOHTTP_TRANSPORT=True,VLLM_ALLREDUCE_USE_SYMM_MEM=0,TORCH_CUDNN_SDPA_ENABLED=0,PYTHONFAULTHANDLER=1,TORCH_NCCL_ASYNC_ERROR_HANDLING=1,TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC=1800,TORCH_NCCL_BLOCKING_WAIT_TIMEOUT_MS=1800000,VLLM_MQ_MAX_CHUNKS=240,OT_AGENT_RAY_LOG_DIR=/e/data1/datasets/playground/ot-baf/experiments/_ray_logs,TORCH_NCCL_TRACE_BUFFER_SIZE=10000,TORCH_FR_BUFFER_SIZE=10000,TORCH_NCCL_DESYNC_DEBUG=1,TORCH_NCCL_DEBUG_INFO_TEMP_FILE=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps/nccl_trace,VLLM_PYNCCL_TRACE_BUFFER_SIZE=10000,VLLM_PYNCCL_TRACE_DUMP_DIR=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps/nccl_trace,VLLM_RAY_EXTRA_ENV_VAR_PREFIXES_TO_COPY=TORCH_NCCL_,TORCH_FR_,VLLM_PYNCCL_', '--nodes=1', '--ntasks=1', '--overlap', '--cpu-bind=none', '-w', 'jpbo-001-45', 'bash', '-c', 'unset LD_PRELOAD PROXYCHAINS_CONF_FILE 2>/dev/null; ray stop --force']' timed out after 30 seconds Warning: Failed to stop Ray on jpbo-001-46: Command '['srun', '--export=ALL,WANDB_MODE=offline,GLOO_USE_IPV6=0,NCCL_SOCKET_FAMILY=AF_INET,VLLM_FORCE_IPV4=1,VLLM_SKIP_FLAG_DISCOVERY=1,SKYRL_ENABLE_NUMA_AFFINITY=1,DISABLE_AIOHTTP_TRANSPORT=True,VLLM_ALLREDUCE_USE_SYMM_MEM=0,TORCH_CUDNN_SDPA_ENABLED=0,PYTHONFAULTHANDLER=1,TORCH_NCCL_ASYNC_ERROR_HANDLING=1,TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC=1800,TORCH_NCCL_BLOCKING_WAIT_TIMEOUT_MS=1800000,VLLM_MQ_MAX_CHUNKS=240,OT_AGENT_RAY_LOG_DIR=/e/data1/datasets/playground/ot-baf/experiments/_ray_logs,TORCH_NCCL_TRACE_BUFFER_SIZE=10000,TORCH_FR_BUFFER_SIZE=10000,TORCH_NCCL_DESYNC_DEBUG=1,TORCH_NCCL_DEBUG_INFO_TEMP_FILE=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps/nccl_trace,VLLM_PYNCCL_TRACE_BUFFER_SIZE=10000,VLLM_PYNCCL_TRACE_DUMP_DIR=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps/nccl_trace,VLLM_RAY_EXTRA_ENV_VAR_PREFIXES_TO_COPY=TORCH_NCCL_,TORCH_FR_,VLLM_PYNCCL_', '--nodes=1', '--ntasks=1', '--overlap', '--cpu-bind=none', '-w', 'jpbo-001-46', 'bash', '-c', 'unset LD_PRELOAD PROXYCHAINS_CONF_FILE 2>/dev/null; ray stop --force']' timed out after 30 seconds Warning: Failed to stop Ray on jpbo-001-47: Command '['srun', '--export=ALL,WANDB_MODE=offline,GLOO_USE_IPV6=0,NCCL_SOCKET_FAMILY=AF_INET,VLLM_FORCE_IPV4=1,VLLM_SKIP_FLAG_DISCOVERY=1,SKYRL_ENABLE_NUMA_AFFINITY=1,DISABLE_AIOHTTP_TRANSPORT=True,VLLM_ALLREDUCE_USE_SYMM_MEM=0,TORCH_CUDNN_SDPA_ENABLED=0,PYTHONFAULTHANDLER=1,TORCH_NCCL_ASYNC_ERROR_HANDLING=1,TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC=1800,TORCH_NCCL_BLOCKING_WAIT_TIMEOUT_MS=1800000,VLLM_MQ_MAX_CHUNKS=240,OT_AGENT_RAY_LOG_DIR=/e/data1/datasets/playground/ot-baf/experiments/_ray_logs,TORCH_NCCL_TRACE_BUFFER_SIZE=10000,TORCH_FR_BUFFER_SIZE=10000,TORCH_NCCL_DESYNC_DEBUG=1,TORCH_NCCL_DEBUG_INFO_TEMP_FILE=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps/nccl_trace,VLLM_PYNCCL_TRACE_BUFFER_SIZE=10000,VLLM_PYNCCL_TRACE_DUMP_DIR=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps/nccl_trace,VLLM_RAY_EXTRA_ENV_VAR_PREFIXES_TO_COPY=TORCH_NCCL_,TORCH_FR_,VLLM_PYNCCL_', '--nodes=1', '--ntasks=1', '--overlap', '--cpu-bind=none', '-w', 'jpbo-001-47', 'bash', '-c', 'unset LD_PRELOAD PROXYCHAINS_CONF_FILE 2>/dev/null; ray stop --force']' timed out after 30 seconds Warning: Failed to stop Ray on jpbo-003-39: Command '['srun', '--export=ALL,WANDB_MODE=offline,GLOO_USE_IPV6=0,NCCL_SOCKET_FAMILY=AF_INET,VLLM_FORCE_IPV4=1,VLLM_SKIP_FLAG_DISCOVERY=1,SKYRL_ENABLE_NUMA_AFFINITY=1,DISABLE_AIOHTTP_TRANSPORT=True,VLLM_ALLREDUCE_USE_SYMM_MEM=0,TORCH_CUDNN_SDPA_ENABLED=0,PYTHONFAULTHANDLER=1,TORCH_NCCL_ASYNC_ERROR_HANDLING=1,TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC=1800,TORCH_NCCL_BLOCKING_WAIT_TIMEOUT_MS=1800000,VLLM_MQ_MAX_CHUNKS=240,OT_AGENT_RAY_LOG_DIR=/e/data1/datasets/playground/ot-baf/experiments/_ray_logs,TORCH_NCCL_TRACE_BUFFER_SIZE=10000,TORCH_FR_BUFFER_SIZE=10000,TORCH_NCCL_DESYNC_DEBUG=1,TORCH_NCCL_DEBUG_INFO_TEMP_FILE=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps/nccl_trace,VLLM_PYNCCL_TRACE_BUFFER_SIZE=10000,VLLM_PYNCCL_TRACE_DUMP_DIR=/e/data1/datasets/playground/ot-baf/experiments/_nccl_dumps/nccl_trace,VLLM_RAY_EXTRA_ENV_VAR_PREFIXES_TO_COPY=TORCH_NCCL_,TORCH_FR_,VLLM_PYNCCL_', '--nodes=1', '--ntasks=1', '--overlap', '--cpu-bind=none', '-w', 'jpbo-003-39', 'bash', '-c', 'unset LD_PRELOAD PROXYCHAINS_CONF_FILE 2>/dev/null; ray stop --force']' timed out after 30 seconds Ray cluster stopped [RLJobRunner] Crash detected (exit!=0) — preserving Ray logs to /e/data1/datasets/playground/ot-baf/ablation-pymethods2test-seqmean-arm0-tis/ablation-pymethods2test-seqmean-arm0-tis/ray_logs BEFORE trace upload (so a wall-clock kill can't lose crash evidence)... [RLJobRunner] Crash-time Ray log preservation timed out (600s); continuing. [RLJobRunner] No trace_jobs directory found at /e/data1/datasets/playground/ot-baf/ablation-pymethods2test-seqmean-arm0-tis/ablation-pymethods2test-seqmean-arm0-tis/trace_jobs, skipping upload. Preserving Ray logs to /e/data1/datasets/playground/ot-baf/ablation-pymethods2test-seqmean-arm0-tis/ray_logs/ Collecting Ray logs from worker jpbo-001-36... Collecting Ray logs from worker jpbo-001-37... Collecting Ray logs from worker jpbo-001-38... Collecting Ray logs from worker jpbo-001-41... Collecting Ray logs from worker jpbo-001-42... Collecting Ray logs from worker jpbo-001-43... Collecting Ray logs from worker jpbo-001-44... Collecting Ray logs from worker jpbo-001-45... Collecting Ray logs from worker jpbo-001-46... Collecting Ray logs from worker jpbo-001-47... Collecting Ray logs from worker jpbo-001-48... Collecting Ray logs from worker jpbo-003-37... Collecting Ray logs from worker jpbo-003-39... Ray log preservation complete