XPK Start: Tue Apr 21 12:06:23 UTC 2026 2026-04-21 12:06:40.712581: E external/local_xla/xla/stream_executor/cuda/cuda_platform.cc:51] failed call to cuInit: INTERNAL: CUDA error: Failed call to cuInit: UNKNOWN ERROR (303) I0421 12:06:44.317013 132528837793600 max_utils.py:273] Attempting to initialize the jax distributed system... INFO:2026-04-21 12:06:53,356:jax._src.distributed:149: Starting JAX distributed service on [::]:8482 I0421 12:06:53.356404 132528837793600 distributed.py:149] Starting JAX distributed service on [::]:8482 INFO:2026-04-21 12:06:53,358:jax._src.distributed:166: Connecting to JAX distributed service on mt-07-distill-smoke-clkfo-slice-job-0-0.mt-07-distill-smoke-clkfo:8482 I0421 12:06:53.358831 132528837793600 distributed.py:166] Connecting to JAX distributed service on mt-07-distill-smoke-clkfo-slice-job-0-0.mt-07-distill-smoke-clkfo:8482 I0421 12:06:54.419385 132528837793600 max_utils.py:284] Jax distributed system initialized! I0421 12:07:00.568473 132528837793600 max_utils.py:244] Jax distributed system is already initialized. I0421 12:07:01.042128 132528837793600 max_utils.py:244] Jax distributed system is already initialized. I0421 12:07:01.043320 132528837793600 pyconfig.py:432] Config param abort_on_inf_loss: True I0421 12:07:01.043370 132528837793600 pyconfig.py:432] Config param abort_on_nan_loss: True I0421 12:07:01.043398 132528837793600 pyconfig.py:432] Config param act_quantization_calibration_method: absmax I0421 12:07:01.043420 132528837793600 pyconfig.py:432] Config param activation_dropout_for_audio: 0.0 I0421 12:07:01.043441 132528837793600 pyconfig.py:432] Config param activation_function_for_audio: gelu I0421 12:07:01.043459 132528837793600 pyconfig.py:432] Config param activations_in_float32: False I0421 12:07:01.043479 132528837793600 pyconfig.py:432] Config param adam_b1: 0.9 I0421 12:07:01.043497 132528837793600 pyconfig.py:432] Config param adam_b2: 0.95 I0421 12:07:01.043514 132528837793600 pyconfig.py:432] Config param adam_eps: 1e-08 I0421 12:07:01.043537 132528837793600 pyconfig.py:432] Config param adam_eps_root: 0.0 I0421 12:07:01.043555 132528837793600 pyconfig.py:432] Config param adam_weight_decay: 0.1 I0421 12:07:01.043572 132528837793600 pyconfig.py:432] Config param adamw_mask: [] I0421 12:07:01.043589 132528837793600 pyconfig.py:432] Config param add_bos: True I0421 12:07:01.043606 132528837793600 pyconfig.py:432] Config param add_eos: True I0421 12:07:01.043623 132528837793600 pyconfig.py:432] Config param allow_split_physical_axes: False I0421 12:07:01.043640 132528837793600 pyconfig.py:432] Config param ar_cache_axis_order: 1,2,0,3 I0421 12:07:01.043657 132528837793600 pyconfig.py:432] Config param async_checkpointing: True I0421 12:07:01.043673 132528837793600 pyconfig.py:432] Config param async_scheduling: False I0421 12:07:01.043689 132528837793600 pyconfig.py:432] Config param attention: dot_product I0421 12:07:01.043706 132528837793600 pyconfig.py:432] Config param attention_bias: False I0421 12:07:01.043721 132528837793600 pyconfig.py:432] Config param attention_dropout_for_audio: 0.0 I0421 12:07:01.043752 132528837793600 pyconfig.py:432] Config param attention_out: RematLocation.REMAT I0421 12:07:01.043773 132528837793600 pyconfig.py:432] Config param attention_output_dim: -1 I0421 12:07:01.043794 132528837793600 pyconfig.py:432] Config param attention_sink: False I0421 12:07:01.043812 132528837793600 pyconfig.py:432] Config param attention_type: global I0421 12:07:01.043829 132528837793600 pyconfig.py:432] Config param attn_logits_soft_cap: None I0421 12:07:01.043846 132528837793600 pyconfig.py:432] Config param audio_path: I0421 12:07:01.043861 132528837793600 pyconfig.py:432] Config param audio_placeholder: <|audio|> I0421 12:07:01.043877 132528837793600 pyconfig.py:432] Config param autoregressive_decode_assert: I0421 12:07:01.043895 132528837793600 pyconfig.py:432] Config param base_config: base.yml I0421 12:07:01.043911 132528837793600 pyconfig.py:432] Config param base_emb_dim: 16 I0421 12:07:01.043928 132528837793600 pyconfig.py:432] Config param base_mlp_dim: 64 I0421 12:07:01.043944 132528837793600 pyconfig.py:432] Config param base_moe_mlp_dim: -1 I0421 12:07:01.043960 132528837793600 pyconfig.py:432] Config param base_num_decoder_layers: 1 I0421 12:07:01.043975 132528837793600 pyconfig.py:432] Config param base_num_kv_heads: 2 I0421 12:07:01.043991 132528837793600 pyconfig.py:432] Config param base_num_query_heads: 2 I0421 12:07:01.044008 132528837793600 pyconfig.py:432] Config param base_output_directory: I0421 12:07:01.044023 132528837793600 pyconfig.py:432] Config param batch_size: 1 I0421 12:07:01.044039 132528837793600 pyconfig.py:432] Config param batch_split_factor: 1 I0421 12:07:01.044055 132528837793600 pyconfig.py:432] Config param beta_fast: 32 I0421 12:07:01.044071 132528837793600 pyconfig.py:432] Config param beta_slow: 1 I0421 12:07:01.044087 132528837793600 pyconfig.py:432] Config param bwd_quantization_calibration_method: absmax I0421 12:07:01.044103 132528837793600 pyconfig.py:432] Config param capacity_factor: -1.0 I0421 12:07:01.044120 132528837793600 pyconfig.py:432] Config param cast_logits_to_fp32: True I0421 12:07:01.044136 132528837793600 pyconfig.py:432] Config param chat_template: I0421 12:07:01.044152 132528837793600 pyconfig.py:432] Config param chat_template_path: I0421 12:07:01.044170 132528837793600 pyconfig.py:432] Config param checkpoint_conversion_fn: None I0421 12:07:01.044187 132528837793600 pyconfig.py:432] Config param checkpoint_dir: None I0421 12:07:01.044205 132528837793600 pyconfig.py:432] Config param checkpoint_is_quantized: False I0421 12:07:01.044222 132528837793600 pyconfig.py:432] Config param checkpoint_period: 2000 I0421 12:07:01.044238 132528837793600 pyconfig.py:432] Config param checkpoint_storage_concurrent_gb: 96 I0421 12:07:01.044255 132528837793600 pyconfig.py:432] Config param checkpoint_storage_target_data_file_size_bytes: 2147483648 I0421 12:07:01.044272 132528837793600 pyconfig.py:432] Config param checkpoint_storage_use_ocdbt: True I0421 12:07:01.044289 132528837793600 pyconfig.py:432] Config param checkpoint_storage_use_zarr3: True I0421 12:07:01.044305 132528837793600 pyconfig.py:432] Config param checkpoint_todelete_full_path: None I0421 12:07:01.044320 132528837793600 pyconfig.py:432] Config param checkpoint_todelete_subdir: None I0421 12:07:01.044336 132528837793600 pyconfig.py:432] Config param chips_per_vm: 4 I0421 12:07:01.044352 132528837793600 pyconfig.py:432] Config param chunk_attn_window_size: 0 I0421 12:07:01.044367 132528837793600 pyconfig.py:432] Config param collect_stack_trace: False I0421 12:07:01.044383 132528837793600 pyconfig.py:432] Config param colocated_python_checkpointing: False I0421 12:07:01.044400 132528837793600 pyconfig.py:432] Config param colocated_python_data_input: False I0421 12:07:01.044418 132528837793600 pyconfig.py:432] Config param compile_topology: I0421 12:07:01.044434 132528837793600 pyconfig.py:432] Config param compile_topology_num_slices: -1 I0421 12:07:01.044450 132528837793600 pyconfig.py:432] Config param compile_xla_flags: I0421 12:07:01.044465 132528837793600 pyconfig.py:432] Config param compiled_trainstep_file: I0421 12:07:01.044480 132528837793600 pyconfig.py:432] Config param compute_axis_order: 0,1,2,3 I0421 12:07:01.044496 132528837793600 pyconfig.py:432] Config param constant_bound_config: [] I0421 12:07:01.044512 132528837793600 pyconfig.py:432] Config param context: RematLocation.REMAT I0421 12:07:01.044529 132528837793600 pyconfig.py:432] Config param context_parallel_load_balance: True I0421 12:07:01.044544 132528837793600 pyconfig.py:432] Config param context_parallel_size: 1 I0421 12:07:01.044559 132528837793600 pyconfig.py:432] Config param context_parallel_strategy: all_gather I0421 12:07:01.044575 132528837793600 pyconfig.py:432] Config param context_sharding: context I0421 12:07:01.044598 132528837793600 pyconfig.py:432] Config param conv_chunksize_for_audio: 500 I0421 12:07:01.044614 132528837793600 pyconfig.py:432] Config param conv_stride_for_vit: 14 I0421 12:07:01.044630 132528837793600 pyconfig.py:432] Config param cost_estimate_flops_bwd: -1 I0421 12:07:01.044646 132528837793600 pyconfig.py:432] Config param cost_estimate_flops_fwd: -1 I0421 12:07:01.044661 132528837793600 pyconfig.py:432] Config param custom_mesh: I0421 12:07:01.044677 132528837793600 pyconfig.py:432] Config param custom_mesh_and_rule: I0421 12:07:01.044692 132528837793600 pyconfig.py:432] Config param d_model_for_audio: 256 I0421 12:07:01.044708 132528837793600 pyconfig.py:432] Config param data_sharding: (('data', 'stage', 'fsdp', 'fsdp_transpose', 'sequence', 'context', 'context_autoregressive', 'tensor', 'tensor_transpose', 'tensor_sequence', 'expert', 'autoregressive'),) I0421 12:07:01.044728 132528837793600 pyconfig.py:432] Config param data_shuffle_seed: 0 I0421 12:07:01.044753 132528837793600 pyconfig.py:432] Config param dataset_name: c4/en:3.0.1 I0421 12:07:01.044769 132528837793600 pyconfig.py:432] Config param dataset_path: I0421 12:07:01.044790 132528837793600 pyconfig.py:432] Config param dataset_type: DatasetType.HF I0421 12:07:01.044807 132528837793600 pyconfig.py:432] Config param dcn_autoregressive_parallelism: 1 I0421 12:07:01.044823 132528837793600 pyconfig.py:432] Config param dcn_context_autoregressive_parallelism: 1 I0421 12:07:01.044839 132528837793600 pyconfig.py:432] Config param dcn_context_parallelism: 1 I0421 12:07:01.044854 132528837793600 pyconfig.py:432] Config param dcn_data_parallelism: -1 I0421 12:07:01.044870 132528837793600 pyconfig.py:432] Config param dcn_diloco_parallelism: 1 I0421 12:07:01.044886 132528837793600 pyconfig.py:432] Config param dcn_expert_parallelism: 1 I0421 12:07:01.044900 132528837793600 pyconfig.py:432] Config param dcn_fsdp_parallelism: 1 I0421 12:07:01.044915 132528837793600 pyconfig.py:432] Config param dcn_fsdp_transpose_parallelism: 1 I0421 12:07:01.044932 132528837793600 pyconfig.py:432] Config param dcn_parallelism: [1, -1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] I0421 12:07:01.044949 132528837793600 pyconfig.py:432] Config param dcn_pipeline_parallelism: 1 I0421 12:07:01.044965 132528837793600 pyconfig.py:432] Config param dcn_sequence_parallelism: 1 I0421 12:07:01.044981 132528837793600 pyconfig.py:432] Config param dcn_tensor_parallelism: 1 I0421 12:07:01.044995 132528837793600 pyconfig.py:432] Config param dcn_tensor_sequence_parallelism: 1 I0421 12:07:01.045011 132528837793600 pyconfig.py:432] Config param dcn_tensor_transpose_parallelism: 1 I0421 12:07:01.045026 132528837793600 pyconfig.py:432] Config param debug: {'rl': False} I0421 12:07:01.045043 132528837793600 pyconfig.py:432] Config param debug_sharding: False I0421 12:07:01.045060 132528837793600 pyconfig.py:432] Config param decode_sampling_nucleus_p: -1 I0421 12:07:01.045075 132528837793600 pyconfig.py:432] Config param decode_sampling_strategy: SamplingStrategy.GREEDY I0421 12:07:01.045092 132528837793600 pyconfig.py:432] Config param decode_sampling_temperature: 1.0 I0421 12:07:01.045108 132528837793600 pyconfig.py:432] Config param decode_sampling_top_k: 0 I0421 12:07:01.045123 132528837793600 pyconfig.py:432] Config param decoder_block: DecoderBlockType.GPT3 I0421 12:07:01.045140 132528837793600 pyconfig.py:432] Config param decoder_layer_input: RematLocation.DEVICE I0421 12:07:01.045157 132528837793600 pyconfig.py:432] Config param deepstack_visual_indexes_for_vit: [] I0421 12:07:01.045173 132528837793600 pyconfig.py:432] Config param degenerate_group_masking: True I0421 12:07:01.045187 132528837793600 pyconfig.py:432] Config param dense_init_scale: 1.0 I0421 12:07:01.045203 132528837793600 pyconfig.py:432] Config param diloco_outer_lr: 0.3 I0421 12:07:01.045220 132528837793600 pyconfig.py:432] Config param diloco_outer_momentum: 0.9 I0421 12:07:01.045235 132528837793600 pyconfig.py:432] Config param diloco_sync_period: 36 I0421 12:07:01.045250 132528837793600 pyconfig.py:432] Config param distill_alpha: 0.5 I0421 12:07:01.045266 132528837793600 pyconfig.py:432] Config param distill_alpha_end: None I0421 12:07:01.045282 132528837793600 pyconfig.py:432] Config param distill_alpha_schedule: constant I0421 12:07:01.045299 132528837793600 pyconfig.py:432] Config param distill_beta: 0.0 I0421 12:07:01.045314 132528837793600 pyconfig.py:432] Config param distill_beta_end: None I0421 12:07:01.045330 132528837793600 pyconfig.py:432] Config param distill_beta_schedule: constant I0421 12:07:01.045346 132528837793600 pyconfig.py:432] Config param distill_feature_loss_type: cosine I0421 12:07:01.045361 132528837793600 pyconfig.py:432] Config param distill_layer_indices: None I0421 12:07:01.045376 132528837793600 pyconfig.py:432] Config param distill_temperature: 1.0 I0421 12:07:01.045392 132528837793600 pyconfig.py:432] Config param distill_temperature_end: None I0421 12:07:01.045408 132528837793600 pyconfig.py:432] Config param distill_temperature_schedule: constant I0421 12:07:01.045424 132528837793600 pyconfig.py:432] Config param downsample_hidden_size_for_audio: 256 I0421 12:07:01.045440 132528837793600 pyconfig.py:432] Config param dpo_beta: 0.1 I0421 12:07:01.045456 132528837793600 pyconfig.py:432] Config param dpo_label_smoothing: 0.0 I0421 12:07:01.045472 132528837793600 pyconfig.py:432] Config param dq_reduction_steps: 0 I0421 12:07:01.045488 132528837793600 pyconfig.py:432] Config param dropout_rate: 0.0 I0421 12:07:01.045504 132528837793600 pyconfig.py:432] Config param dtype: bfloat16 I0421 12:07:01.045535 132528837793600 pyconfig.py:432] Config param dtype_mm: float32 I0421 12:07:01.045551 132528837793600 pyconfig.py:432] Config param dump_hlo: False I0421 12:07:01.045567 132528837793600 pyconfig.py:432] Config param dump_hlo_delete_local_after: True I0421 12:07:01.045582 132528837793600 pyconfig.py:432] Config param dump_hlo_gcs_dir: gpt3-52k_2026-04-21-12-07/xla_dump I0421 12:07:01.045598 132528837793600 pyconfig.py:432] Config param dump_hlo_local_dir: /tmp/xla_dump/ I0421 12:07:01.045613 132528837793600 pyconfig.py:432] Config param dump_hlo_local_module_name: jit_train_step I0421 12:07:01.045628 132528837793600 pyconfig.py:432] Config param dump_hlo_module_name: jit_train_step I0421 12:07:01.045644 132528837793600 pyconfig.py:432] Config param dump_hlo_upload_all: False I0421 12:07:01.045660 132528837793600 pyconfig.py:432] Config param dump_hlo_xla_flags: I0421 12:07:01.045675 132528837793600 pyconfig.py:432] Config param dump_jaxpr: False I0421 12:07:01.045691 132528837793600 pyconfig.py:432] Config param dump_jaxpr_delete_local_after: True I0421 12:07:01.045707 132528837793600 pyconfig.py:432] Config param dump_jaxpr_gcs_dir: gpt3-52k_2026-04-21-12-07/jaxpr_dump I0421 12:07:01.045723 132528837793600 pyconfig.py:432] Config param dump_jaxpr_local_dir: /tmp/jaxpr_dump/ I0421 12:07:01.045749 132528837793600 pyconfig.py:432] Config param dump_step: -1 I0421 12:07:01.045765 132528837793600 pyconfig.py:432] Config param elastic_enabled: False I0421 12:07:01.045785 132528837793600 pyconfig.py:432] Config param elastic_max_retries: 10 I0421 12:07:01.045801 132528837793600 pyconfig.py:432] Config param elastic_timeout_seconds: 300 I0421 12:07:01.045817 132528837793600 pyconfig.py:432] Config param emb_dim: 16 I0421 12:07:01.045832 132528837793600 pyconfig.py:432] Config param enable_autocheckpoint: False I0421 12:07:01.045848 132528837793600 pyconfig.py:432] Config param enable_checkpoint_cloud_logger: False I0421 12:07:01.045863 132528837793600 pyconfig.py:432] Config param enable_checkpointing: True I0421 12:07:01.045879 132528837793600 pyconfig.py:432] Config param enable_continuous_checkpointing: False I0421 12:07:01.045895 132528837793600 pyconfig.py:432] Config param enable_data_shuffling: True I0421 12:07:01.045910 132528837793600 pyconfig.py:432] Config param enable_diloco: False I0421 12:07:01.045926 132528837793600 pyconfig.py:432] Config param enable_dp_attention: False I0421 12:07:01.045941 132528837793600 pyconfig.py:432] Config param enable_dropout: False I0421 12:07:01.045956 132528837793600 pyconfig.py:432] Config param enable_emergency_checkpoint: False I0421 12:07:01.045971 132528837793600 pyconfig.py:432] Config param enable_expert_parallel: False I0421 12:07:01.045987 132528837793600 pyconfig.py:432] Config param enable_gcp_goodput_metrics: True I0421 12:07:01.046002 132528837793600 pyconfig.py:432] Config param enable_gcp_step_deviation_metrics: True I0421 12:07:01.046018 132528837793600 pyconfig.py:432] Config param enable_goodput_recording: False I0421 12:07:01.046033 132528837793600 pyconfig.py:432] Config param enable_jax_profiler: False I0421 12:07:01.046047 132528837793600 pyconfig.py:432] Config param enable_llm_inference_pool: False I0421 12:07:01.046063 132528837793600 pyconfig.py:432] Config param enable_model_warmup: False I0421 12:07:01.046077 132528837793600 pyconfig.py:432] Config param enable_multi_tier_checkpointing: False I0421 12:07:01.046092 132528837793600 pyconfig.py:432] Config param enable_nnx: False I0421 12:07:01.046108 132528837793600 pyconfig.py:432] Config param enable_orbax_v1: False I0421 12:07:01.046124 132528837793600 pyconfig.py:432] Config param enable_padding_causal_mask: True I0421 12:07:01.046140 132528837793600 pyconfig.py:432] Config param enable_pathways_goodput: False I0421 12:07:01.046154 132528837793600 pyconfig.py:432] Config param enable_prefix_caching: False I0421 12:07:01.046170 132528837793600 pyconfig.py:432] Config param enable_rampup_batch_size: False I0421 12:07:01.046185 132528837793600 pyconfig.py:432] Config param enable_single_controller: False I0421 12:07:01.046201 132528837793600 pyconfig.py:432] Config param enable_single_replica_ckpt_restoring: False I0421 12:07:01.046217 132528837793600 pyconfig.py:432] Config param enable_tensorboard: True I0421 12:07:01.046233 132528837793600 pyconfig.py:432] Config param enable_tunix_perf_metrics: False I0421 12:07:01.046248 132528837793600 pyconfig.py:432] Config param encoder_attention_heads_for_audio: 4 I0421 12:07:01.046264 132528837793600 pyconfig.py:432] Config param encoder_ffn_dim_for_audio: 512 I0421 12:07:01.046280 132528837793600 pyconfig.py:432] Config param encoder_layers_for_audio: 2 I0421 12:07:01.046297 132528837793600 pyconfig.py:432] Config param engram: RematLocation.REMAT I0421 12:07:01.046313 132528837793600 pyconfig.py:432] Config param engram_head_dim: 1280 I0421 12:07:01.046328 132528837793600 pyconfig.py:432] Config param engram_kernel_size: 4 I0421 12:07:01.046344 132528837793600 pyconfig.py:432] Config param engram_layers: [] I0421 12:07:01.046360 132528837793600 pyconfig.py:432] Config param engram_max_ngram_size: 3 I0421 12:07:01.046375 132528837793600 pyconfig.py:432] Config param engram_num_heads: 8 I0421 12:07:01.046391 132528837793600 pyconfig.py:432] Config param engram_seed: 0 I0421 12:07:01.046407 132528837793600 pyconfig.py:432] Config param engram_vocab_bases: [] I0421 12:07:01.046422 132528837793600 pyconfig.py:432] Config param epsilon_high: None I0421 12:07:01.046438 132528837793600 pyconfig.py:432] Config param eval_corr_lst: False I0421 12:07:01.046453 132528837793600 pyconfig.py:432] Config param eval_data_columns: ['text'] I0421 12:07:01.046469 132528837793600 pyconfig.py:432] Config param eval_dataset_name: c4/en:3.0.1 I0421 12:07:01.046485 132528837793600 pyconfig.py:432] Config param eval_image_column: image I0421 12:07:01.046501 132528837793600 pyconfig.py:432] Config param eval_interval: -1 I0421 12:07:01.046515 132528837793600 pyconfig.py:432] Config param eval_make_lst: False I0421 12:07:01.046531 132528837793600 pyconfig.py:432] Config param eval_per_device_batch_size: 2 I0421 12:07:01.046546 132528837793600 pyconfig.py:432] Config param eval_sampling_strategy: greedy I0421 12:07:01.046562 132528837793600 pyconfig.py:432] Config param eval_split: validation I0421 12:07:01.046577 132528837793600 pyconfig.py:432] Config param eval_steps: -1 I0421 12:07:01.046592 132528837793600 pyconfig.py:432] Config param expansion_factor_real_data: -1.0 I0421 12:07:01.046607 132528837793600 pyconfig.py:432] Config param final_logits_soft_cap: None I0421 12:07:01.046623 132528837793600 pyconfig.py:432] Config param first_num_dense_layers: 0 I0421 12:07:01.046638 132528837793600 pyconfig.py:432] Config param float32_gate_logits: False I0421 12:07:01.046653 132528837793600 pyconfig.py:432] Config param float32_logits: False I0421 12:07:01.046669 132528837793600 pyconfig.py:432] Config param float32_qk_product: False I0421 12:07:01.046684 132528837793600 pyconfig.py:432] Config param float32_weight_sum: True I0421 12:07:01.046700 132528837793600 pyconfig.py:432] Config param force_q_layout: False I0421 12:07:01.046715 132528837793600 pyconfig.py:432] Config param force_unroll: False I0421 12:07:01.046738 132528837793600 pyconfig.py:432] Config param freeze_audio_encoder_params: True I0421 12:07:01.046754 132528837793600 pyconfig.py:432] Config param freeze_vision_encoder_params: True I0421 12:07:01.046770 132528837793600 pyconfig.py:432] Config param fused_mlp: False I0421 12:07:01.046789 132528837793600 pyconfig.py:432] Config param fused_qkv: True I0421 12:07:01.046804 132528837793600 pyconfig.py:432] Config param gcs_metrics: False I0421 12:07:01.046819 132528837793600 pyconfig.py:432] Config param gdn_chunk_size: 64 I0421 12:07:01.046836 132528837793600 pyconfig.py:432] Config param gdn_conv_kernel_dim: 4 I0421 12:07:01.046852 132528837793600 pyconfig.py:432] Config param gdn_key_head_dim: 128 I0421 12:07:01.046868 132528837793600 pyconfig.py:432] Config param gdn_num_key_heads: 16 I0421 12:07:01.046884 132528837793600 pyconfig.py:432] Config param gdn_num_value_heads: 32 I0421 12:07:01.046898 132528837793600 pyconfig.py:432] Config param gdn_value_head_dim: 128 I0421 12:07:01.046914 132528837793600 pyconfig.py:432] Config param generate_padding_batch_eval: False I0421 12:07:01.046929 132528837793600 pyconfig.py:432] Config param generate_padding_batch_train: False I0421 12:07:01.046945 132528837793600 pyconfig.py:432] Config param generate_slice: v5e-16 I0421 12:07:01.046960 132528837793600 pyconfig.py:432] Config param generation_configs: {} I0421 12:07:01.046977 132528837793600 pyconfig.py:432] Config param global_batch_size_to_eval_on: 64 I0421 12:07:01.046991 132528837793600 pyconfig.py:432] Config param global_batch_size_to_load: 512 I0421 12:07:01.047007 132528837793600 pyconfig.py:432] Config param global_batch_size_to_load_eval: 64 I0421 12:07:01.047023 132528837793600 pyconfig.py:432] Config param global_batch_size_to_load_increment: None I0421 12:07:01.047037 132528837793600 pyconfig.py:432] Config param global_batch_size_to_load_start: None I0421 12:07:01.047054 132528837793600 pyconfig.py:432] Config param global_batch_size_to_train_on: 512 I0421 12:07:01.047070 132528837793600 pyconfig.py:432] Config param global_head_dim: 0 I0421 12:07:01.047084 132528837793600 pyconfig.py:432] Config param global_num_kv_heads: 0 I0421 12:07:01.047100 132528837793600 pyconfig.py:432] Config param global_parameter_scale: 1 I0421 12:07:01.047117 132528837793600 pyconfig.py:432] Config param global_rampup_samples: 500 I0421 12:07:01.047133 132528837793600 pyconfig.py:432] Config param global_rope_max_timescale: -1 I0421 12:07:01.047148 132528837793600 pyconfig.py:432] Config param global_rope_proportion: 0.25 I0421 12:07:01.047166 132528837793600 pyconfig.py:432] Config param goodput_upload_interval_seconds: 30 I0421 12:07:01.047182 132528837793600 pyconfig.py:432] Config param grad_dtype: float32 I0421 12:07:01.047217 132528837793600 pyconfig.py:432] Config param gradient_accumulation_steps: 8 I0421 12:07:01.047234 132528837793600 pyconfig.py:432] Config param gradient_clipping_threshold: 1.0 I0421 12:07:01.047250 132528837793600 pyconfig.py:432] Config param grain_data_source_max_workers: 16 I0421 12:07:01.047266 132528837793600 pyconfig.py:432] Config param grain_eval_files: I0421 12:07:01.047282 132528837793600 pyconfig.py:432] Config param grain_file_type: arrayrecord I0421 12:07:01.047299 132528837793600 pyconfig.py:432] Config param grain_num_threads: 16 I0421 12:07:01.047315 132528837793600 pyconfig.py:432] Config param grain_num_threads_eval: 16 I0421 12:07:01.047331 132528837793600 pyconfig.py:432] Config param grain_packing_type: first_fit I0421 12:07:01.047348 132528837793600 pyconfig.py:432] Config param grain_per_worker_buffer_size: 1 I0421 12:07:01.047364 132528837793600 pyconfig.py:432] Config param grain_per_worker_buffer_size_eval: 1 I0421 12:07:01.047379 132528837793600 pyconfig.py:432] Config param grain_prefetch_buffer_size: 500 I0421 12:07:01.047396 132528837793600 pyconfig.py:432] Config param grain_prefetch_buffer_size_eval: 500 I0421 12:07:01.047412 132528837793600 pyconfig.py:432] Config param grain_ram_budget_mb: 1024 I0421 12:07:01.047428 132528837793600 pyconfig.py:432] Config param grain_shuffle_buffer_size: 100 I0421 12:07:01.047443 132528837793600 pyconfig.py:432] Config param grain_train_files: I0421 12:07:01.047459 132528837793600 pyconfig.py:432] Config param grain_train_mixture_config_path: I0421 12:07:01.047474 132528837793600 pyconfig.py:432] Config param grain_worker_count: 1 I0421 12:07:01.047489 132528837793600 pyconfig.py:432] Config param grain_worker_count_eval: 1 I0421 12:07:01.047505 132528837793600 pyconfig.py:432] Config param grpo_beta: 0.08 I0421 12:07:01.047522 132528837793600 pyconfig.py:432] Config param grpo_epsilon: 0.2 I0421 12:07:01.047538 132528837793600 pyconfig.py:432] Config param hardware: tpu I0421 12:07:01.047554 132528837793600 pyconfig.py:432] Config param hbm_utilization_vllm: 0.72 I0421 12:07:01.047569 132528837793600 pyconfig.py:432] Config param head_dim: 8 I0421 12:07:01.047585 132528837793600 pyconfig.py:432] Config param heartbeat_reporting_interval_in_seconds: 5 I0421 12:07:01.047601 132528837793600 pyconfig.py:432] Config param hf_data_dir: None I0421 12:07:01.047617 132528837793600 pyconfig.py:432] Config param hf_eval_files: None I0421 12:07:01.047633 132528837793600 pyconfig.py:432] Config param hf_eval_split: None I0421 12:07:01.047649 132528837793600 pyconfig.py:432] Config param hf_name: None I0421 12:07:01.047664 132528837793600 pyconfig.py:432] Config param hf_path: OptimalScale/ClimbMix I0421 12:07:01.047679 132528837793600 pyconfig.py:432] Config param hf_train_files: None I0421 12:07:01.047695 132528837793600 pyconfig.py:432] Config param hidden_size_for_vit: 1408 I0421 12:07:01.047709 132528837793600 pyconfig.py:432] Config param hide_profiler_step_metric: False I0421 12:07:01.047724 132528837793600 pyconfig.py:432] Config param ici_autoregressive_parallelism: 1 I0421 12:07:01.047748 132528837793600 pyconfig.py:432] Config param ici_context_autoregressive_parallelism: 1 I0421 12:07:01.047762 132528837793600 pyconfig.py:432] Config param ici_context_parallelism: 1 I0421 12:07:01.047782 132528837793600 pyconfig.py:432] Config param ici_data_parallelism: 1 I0421 12:07:01.047820 132528837793600 pyconfig.py:432] Config param ici_diloco_parallelism: 1 I0421 12:07:01.047834 132528837793600 pyconfig.py:432] Config param ici_expert_parallelism: 1 I0421 12:07:01.047850 132528837793600 pyconfig.py:432] Config param ici_fsdp_parallelism: -1 I0421 12:07:01.047865 132528837793600 pyconfig.py:432] Config param ici_fsdp_transpose_parallelism: 1 I0421 12:07:01.047880 132528837793600 pyconfig.py:432] Config param ici_parallelism: [1, 1, 1, -1, 1, 1, 1, 1, 1, 1, 1, 1, 1] I0421 12:07:01.047898 132528837793600 pyconfig.py:432] Config param ici_pipeline_parallelism: 1 I0421 12:07:01.047913 132528837793600 pyconfig.py:432] Config param ici_sequence_parallelism: 1 I0421 12:07:01.047929 132528837793600 pyconfig.py:432] Config param ici_tensor_parallelism: 1 I0421 12:07:01.047943 132528837793600 pyconfig.py:432] Config param ici_tensor_sequence_parallelism: 1 I0421 12:07:01.047958 132528837793600 pyconfig.py:432] Config param ici_tensor_transpose_parallelism: 1 I0421 12:07:01.047973 132528837793600 pyconfig.py:432] Config param image_path: I0421 12:07:01.047988 132528837793600 pyconfig.py:432] Config param image_placeholder: <|image|> I0421 12:07:01.048004 132528837793600 pyconfig.py:432] Config param image_size_for_vit: 896 I0421 12:07:01.048019 132528837793600 pyconfig.py:432] Config param indexer_head_dim: 128 I0421 12:07:01.048035 132528837793600 pyconfig.py:432] Config param indexer_loss_scaling_factor: 0.0 I0421 12:07:01.048050 132528837793600 pyconfig.py:432] Config param indexer_n_heads: 64 I0421 12:07:01.048066 132528837793600 pyconfig.py:432] Config param indexer_sparse_training: False I0421 12:07:01.048080 132528837793600 pyconfig.py:432] Config param indexer_topk: 2048 I0421 12:07:01.048096 132528837793600 pyconfig.py:432] Config param inference_benchmark_test: False I0421 12:07:01.048110 132528837793600 pyconfig.py:432] Config param inference_metadata_file: I0421 12:07:01.048126 132528837793600 pyconfig.py:432] Config param inference_microbenchmark_log_file_path: I0421 12:07:01.048142 132528837793600 pyconfig.py:432] Config param inference_microbenchmark_loop_iters: 10 I0421 12:07:01.048157 132528837793600 pyconfig.py:432] Config param inference_microbenchmark_num_samples: [1, 2, 3, 4, 5] I0421 12:07:01.048173 132528837793600 pyconfig.py:432] Config param inference_microbenchmark_prefill_lengths: 64,128,256,512,1024 I0421 12:07:01.048189 132528837793600 pyconfig.py:432] Config param inference_microbenchmark_stages: prefill,generate I0421 12:07:01.048204 132528837793600 pyconfig.py:432] Config param inference_server: MaxtextInterleavedServer I0421 12:07:01.048219 132528837793600 pyconfig.py:432] Config param inhomogeneous_layer_cycle_interval: 1 I0421 12:07:01.048235 132528837793600 pyconfig.py:432] Config param init_weights_seed: 0 I0421 12:07:01.048251 132528837793600 pyconfig.py:432] Config param input_data_sharding_logical_axes: ['activation_embed_and_logits_batch', 'activation_norm_length'] I0421 12:07:01.048268 132528837793600 pyconfig.py:432] Config param interleave_moe_layer_step: 1 I0421 12:07:01.048283 132528837793600 pyconfig.py:432] Config param intermediate_size_for_vit: 5632 I0421 12:07:01.048299 132528837793600 pyconfig.py:432] Config param internal_compile: False I0421 12:07:01.048314 132528837793600 pyconfig.py:432] Config param internal_compile_num_devices: -1 I0421 12:07:01.048330 132528837793600 pyconfig.py:432] Config param jax_cache_dir: ~/jax_cache I0421 12:07:01.048344 132528837793600 pyconfig.py:432] Config param jax_debug_log_modules: I0421 12:07:01.048360 132528837793600 pyconfig.py:432] Config param jax_distributed_initialization_timeout: 300 I0421 12:07:01.048376 132528837793600 pyconfig.py:432] Config param jax_profiler_port: 9999 I0421 12:07:01.048391 132528837793600 pyconfig.py:432] Config param key_proj: RematLocation.REMAT I0421 12:07:01.048408 132528837793600 pyconfig.py:432] Config param kv_cache_buffer: 256 I0421 12:07:01.048424 132528837793600 pyconfig.py:432] Config param kv_lora_rank: 512 I0421 12:07:01.048440 132528837793600 pyconfig.py:432] Config param kv_quant_axis: KvQuantAxis.HEADS_AND_DKV I0421 12:07:01.048459 132528837793600 pyconfig.py:432] Config param kv_quant_dtype: int8 I0421 12:07:01.048475 132528837793600 pyconfig.py:432] Config param kv_wa_proj: RematLocation.REMAT I0421 12:07:01.048491 132528837793600 pyconfig.py:432] Config param learning_rate: 0.0002 I0421 12:07:01.048507 132528837793600 pyconfig.py:432] Config param learning_rate_final_fraction: 0.1 I0421 12:07:01.048522 132528837793600 pyconfig.py:432] Config param learning_rate_schedule_steps: 200000 I0421 12:07:01.048537 132528837793600 pyconfig.py:432] Config param load_balance_loss_weight: 0.0 I0421 12:07:01.048552 132528837793600 pyconfig.py:432] Config param load_checkpoint_only_once: False I0421 12:07:01.048568 132528837793600 pyconfig.py:432] Config param load_from_prefill_dir: False I0421 12:07:01.048583 132528837793600 pyconfig.py:432] Config param load_full_state_path: I0421 12:07:01.048598 132528837793600 pyconfig.py:432] Config param load_parameters_path: gs://lance-maxtext/pt_seed_ckpts/pt_seed_ckpts/pt_seed_ckpt_gpt352k_v32k_linen/checkpoints/4/items I0421 12:07:01.048613 132528837793600 pyconfig.py:432] Config param local_checkpoint_directory: I0421 12:07:01.048628 132528837793600 pyconfig.py:432] Config param local_checkpoint_period: 0 I0421 12:07:01.048644 132528837793600 pyconfig.py:432] Config param local_rope_max_timescale: -1 I0421 12:07:01.048660 132528837793600 pyconfig.py:432] Config param local_rope_proportion: 1.0 I0421 12:07:01.048675 132528837793600 pyconfig.py:432] Config param log_config: True I0421 12:07:01.048690 132528837793600 pyconfig.py:432] Config param log_period: 10 I0421 12:07:01.048705 132528837793600 pyconfig.py:432] Config param logical_axis_rules: (('activation_embed_and_logits_batch', ('data', 'stage', 'fsdp', 'fsdp_transpose', 'expert')), ('activation_embed_and_logits_batch_sequence', ('data', 'stage', 'fsdp', 'fsdp_transpose', 'sequence', 'context', 'expert')), ('activation_vocab', ('tensor', 'tensor_transpose', 'tensor_sequence')), ('activation_vocab', ('tensor', 'tensor_transpose')), ('activation_vocab', 'tensor_sequence'), ('activation_vocab', ('sequence', 'context')), ('vocab', ('tensor', 'tensor_transpose', 'tensor_sequence', 'autoregressive')), ('embed_vocab', ('fsdp', 'fsdp_transpose', 'sequence', 'context', 'expert')), ('activation_heads', ('tensor', 'tensor_transpose', 'sequence', 'tensor_sequence', 'autoregressive')), ('activation_kv_heads', ('tensor', 'tensor_transpose', 'sequence', 'tensor_sequence')), ('activation_attn_length', ('sequence', 'context')), ('activation_attn_length', ('context',)), ('activation_q_length', ('context',)), ('activation_kv_length', ()), ('activation_attn_embed', ('tensor', 'tensor_transpose')), ('activation_kv', ('tensor', 'tensor_transpose', 'tensor_sequence')), ('activation_kv_batch', ('data', 'fsdp', 'fsdp_transpose', 'expert')), ('activation_kv_head_dim', ('tensor', 'tensor_transpose', 'tensor_sequence')), ('heads', ('tensor', 'tensor_transpose', 'tensor_sequence', 'autoregressive')), ('q_heads', ('tensor', 'tensor_transpose', 'tensor_sequence', 'autoregressive')), ('kv_heads', ('tensor', 'tensor_transpose', 'tensor_sequence', 'autoregressive')), ('qkv', ()), ('kv', ()), ('kv_head_dim', ()), ('q_lora', ('fsdp', 'fsdp_transpose', 'sequence', 'context', 'tensor_transpose', 'expert')), ('q_lora', ('fsdp', 'sequence', 'context', 'tensor_transpose', 'expert')), ('q_lora', ('fsdp', 'fsdp_transpose', 'sequence', 'context', 'expert')), ('q_lora', ('fsdp', 'sequence', 'context', 'expert')), ('q_lora_up_proj', ()), ('kv_lora', ('fsdp', 'fsdp_transpose', 'sequence', 'context', 'tensor_transpose', 'expert')), ('kv_lora', ('fsdp', 'sequence', 'context', 'tensor_transpose', 'expert')), ('kv_lora', ('fsdp', 'fsdp_transpose', 'sequence', 'context', 'expert')), ('kv_lora', ('fsdp', 'sequence', 'context', 'expert')), ('kv_lora_up_proj', ()), ('activation_batch_moe', ('data', 'fsdp', 'fsdp_transpose')), ('activation_length_moe', ('sequence', 'context')), ('activation_length_moe', ('context',)), ('activation_norm_length_moe', ('tensor_sequence', 'context', 'sequence')), ('activation_embed_moe', ('tensor', 'tensor_transpose')), ('activation_mlp_moe', ('tensor', 'tensor_transpose', 'tensor_sequence')), ('activation_exp', ('expert',)), ('exp', 'expert'), ('mlp_moe', ('fsdp_transpose', 'tensor', 'tensor_sequence', 'autoregressive')), ('embed_moe', ('fsdp', 'fsdp_transpose', 'sequence', 'tensor_transpose', 'context')), ('embed_moe', ('fsdp', 'sequence', 'tensor_transpose', 'context')), ('embed_moe', ('fsdp', 'fsdp_transpose', 'sequence', 'context')), ('embed_moe', ('fsdp', 'sequence', 'context')), ('activation_mlp', ('tensor', 'tensor_transpose', 'tensor_sequence')), ('activation_batch', ('data', 'fsdp', 'fsdp_transpose', 'expert')), ('activation_length', ('sequence', 'context')), ('activation_length', ('context',)), ('activation_norm_length', ('tensor_sequence', 'context', 'sequence')), ('activation_embed', ('tensor', 'tensor_transpose')), ('activation_stage', 'stage'), ('mlp', ('fsdp_transpose', 'tensor', 'tensor_sequence', 'autoregressive')), ('embed', ('fsdp', 'fsdp_transpose', 'sequence', 'tensor_transpose', 'context', 'expert')), ('embed', ('fsdp', 'sequence', 'tensor_transpose', 'context', 'expert')), ('embed', ('fsdp', 'fsdp_transpose', 'sequence', 'context', 'expert')), ('embed', ('fsdp', 'sequence', 'context', 'expert')), ('norm', ('tensor', 'tensor_transpose')), ('layers', 'stage'), ('diloco', 'diloco'), ('engram_dim', ('tensor',)), ('dense_layers', ()), ('moe_layers', ()), ('mhc', ()), ('prefill_activation_length', ('sequence', 'context')), ('prefill_activation_norm_length', ('tensor_sequence', 'context', 'sequence')), ('activation_prefill_kv_batch', ('data', 'fsdp', 'fsdp_transpose', 'expert')), ('decode_batch', ('data', 'fsdp', 'fsdp_transpose', 'expert')), ('decode_length', ('sequence',)), ('cache_heads', ('autoregressive', 'tensor', 'tensor_transpose', 'tensor_sequence')), ('cache_heads', ('autoregressive', 'tensor', 'tensor_sequence')), ('paged_kv_heads', ('tensor',)), ('cache_batch_prefill', ()), ('cache_batch', ()), ('cache_heads_none', ()), ('cache_kv', ()), ('cache_sequence', ()), ('num_pages', ()), ('tokens_per_page', ()), ('paged_kv_head_dim_size', ()), ('mlp_no_fsdp', ('tensor', 'tensor_sequence', 'autoregressive')), ('embed_tensor_transpose', ('tensor_transpose',)), ('exp_with_fsdp', 'fsdp')) I0421 12:07:01.048795 132528837793600 pyconfig.py:432] Config param logits_dot_in_fp32: False I0421 12:07:01.048822 132528837793600 pyconfig.py:432] Config param logits_via_embedding: True I0421 12:07:01.048838 132528837793600 pyconfig.py:432] Config param lora_input_adapters_path: I0421 12:07:01.048854 132528837793600 pyconfig.py:432] Config param loss_algo: grpo I0421 12:07:01.048870 132528837793600 pyconfig.py:432] Config param lr_schedule_type: LearningRateScheduleType.COSINE I0421 12:07:01.048888 132528837793600 pyconfig.py:432] Config param managed_mldiagnostics: False I0421 12:07:01.048904 132528837793600 pyconfig.py:432] Config param managed_mldiagnostics_dir: None I0421 12:07:01.048920 132528837793600 pyconfig.py:432] Config param managed_mldiagnostics_run_group: I0421 12:07:01.048934 132528837793600 pyconfig.py:432] Config param matmul_precision: MatmulPrecision.DEFAULT I0421 12:07:01.048952 132528837793600 pyconfig.py:432] Config param max_checkify: False I0421 12:07:01.048968 132528837793600 pyconfig.py:432] Config param max_concurrency: 256 I0421 12:07:01.048984 132528837793600 pyconfig.py:432] Config param max_corpus_chars: 10000000 I0421 12:07:01.048999 132528837793600 pyconfig.py:432] Config param max_num_batched_tokens: None I0421 12:07:01.049014 132528837793600 pyconfig.py:432] Config param max_num_checkpoints_to_keep: None I0421 12:07:01.049029 132528837793600 pyconfig.py:432] Config param max_num_images_per_example: -1 I0421 12:07:01.049044 132528837793600 pyconfig.py:432] Config param max_num_seqs: None I0421 12:07:01.049060 132528837793600 pyconfig.py:432] Config param max_position_embeddings: 163840 I0421 12:07:01.049076 132528837793600 pyconfig.py:432] Config param max_prefill_predict_length: 64 I0421 12:07:01.049092 132528837793600 pyconfig.py:432] Config param max_sample_len_for_audio: 10000 I0421 12:07:01.049107 132528837793600 pyconfig.py:432] Config param max_segments_per_seq: -1 I0421 12:07:01.049122 132528837793600 pyconfig.py:432] Config param max_source_positions_for_audio: 1500 I0421 12:07:01.049138 132528837793600 pyconfig.py:432] Config param max_target_length: 2048 I0421 12:07:01.049154 132528837793600 pyconfig.py:432] Config param max_timescale_for_audio: 10000.0 I0421 12:07:01.049170 132528837793600 pyconfig.py:432] Config param megablox: True I0421 12:07:01.049184 132528837793600 pyconfig.py:432] Config param merge_gating_gmm: False I0421 12:07:01.049199 132528837793600 pyconfig.py:432] Config param mesh_axes: ['diloco', 'data', 'stage', 'fsdp', 'fsdp_transpose', 'sequence', 'context', 'context_autoregressive', 'tensor', 'tensor_transpose', 'tensor_sequence', 'expert', 'autoregressive'] I0421 12:07:01.049216 132528837793600 pyconfig.py:432] Config param metrics_dir: None I0421 12:07:01.049232 132528837793600 pyconfig.py:432] Config param metrics_file: I0421 12:07:01.049248 132528837793600 pyconfig.py:432] Config param mhc_expansion_rate: 1 I0421 12:07:01.049264 132528837793600 pyconfig.py:432] Config param micro_batch_size_to_eval_on: 64 I0421 12:07:01.049280 132528837793600 pyconfig.py:432] Config param micro_batch_size_to_train_on: 64 I0421 12:07:01.049296 132528837793600 pyconfig.py:432] Config param mla_kv: RematLocation.REMAT I0421 12:07:01.049313 132528837793600 pyconfig.py:432] Config param mla_naive_kvcache: True I0421 12:07:01.049329 132528837793600 pyconfig.py:432] Config param mla_q: RematLocation.REMAT I0421 12:07:01.049345 132528837793600 pyconfig.py:432] Config param mlp_activations: ['gelu'] I0421 12:07:01.049361 132528837793600 pyconfig.py:432] Config param mlp_activations_limit: -1.0 I0421 12:07:01.049377 132528837793600 pyconfig.py:432] Config param mlp_bias: False I0421 12:07:01.049392 132528837793600 pyconfig.py:432] Config param mlp_dim: 64 I0421 12:07:01.049408 132528837793600 pyconfig.py:432] Config param mlpwi: RematLocation.REMAT I0421 12:07:01.049424 132528837793600 pyconfig.py:432] Config param mlpwi_0: RematLocation.REMAT I0421 12:07:01.049439 132528837793600 pyconfig.py:432] Config param mlpwi_1: RematLocation.REMAT I0421 12:07:01.049455 132528837793600 pyconfig.py:432] Config param mlpwo: RematLocation.REMAT I0421 12:07:01.049470 132528837793600 pyconfig.py:432] Config param moba: False I0421 12:07:01.049484 132528837793600 pyconfig.py:432] Config param moba_chunk_size: 1024 I0421 12:07:01.049500 132528837793600 pyconfig.py:432] Config param moba_topk: 8 I0421 12:07:01.049515 132528837793600 pyconfig.py:432] Config param model_call_mode: I0421 12:07:01.049530 132528837793600 pyconfig.py:432] Config param model_name: gpt3-52k I0421 12:07:01.049546 132528837793600 pyconfig.py:432] Config param moe_expert_input_dim: -1 I0421 12:07:01.049560 132528837793600 pyconfig.py:432] Config param moe_fsdp_use_two_stage_all_gather: False I0421 12:07:01.049576 132528837793600 pyconfig.py:432] Config param moe_mlp_dim: -1 I0421 12:07:01.049590 132528837793600 pyconfig.py:432] Config param moe_mlpwi_0: RematLocation.REMAT I0421 12:07:01.049606 132528837793600 pyconfig.py:432] Config param moe_mlpwi_1: RematLocation.REMAT I0421 12:07:01.049621 132528837793600 pyconfig.py:432] Config param moe_mlpwo: RematLocation.REMAT I0421 12:07:01.049637 132528837793600 pyconfig.py:432] Config param monitor_goodput: False I0421 12:07:01.049651 132528837793600 pyconfig.py:432] Config param monitor_step_time_deviation: True I0421 12:07:01.049667 132528837793600 pyconfig.py:432] Config param mrope_section: [24, 20, 20] I0421 12:07:01.049682 132528837793600 pyconfig.py:432] Config param mscale: 1.0 I0421 12:07:01.049698 132528837793600 pyconfig.py:432] Config param mtc_data_parallelism: 0 I0421 12:07:01.049714 132528837793600 pyconfig.py:432] Config param mtp_eval_target_module: 0 I0421 12:07:01.049729 132528837793600 pyconfig.py:432] Config param mtp_loss_scaling_factor: 0.1 I0421 12:07:01.049757 132528837793600 pyconfig.py:432] Config param mtp_num_layers: 0 I0421 12:07:01.049771 132528837793600 pyconfig.py:432] Config param mu_dtype: float32 I0421 12:07:01.049799 132528837793600 pyconfig.py:432] Config param multi_sampling: False I0421 12:07:01.049816 132528837793600 pyconfig.py:432] Config param multi_tier_checkpointing_backup_interval_minutes: 0 I0421 12:07:01.049830 132528837793600 pyconfig.py:432] Config param muon_beta: 0.95 I0421 12:07:01.049846 132528837793600 pyconfig.py:432] Config param muon_consistent_rms: None I0421 12:07:01.049860 132528837793600 pyconfig.py:432] Config param muon_weight_decay: 0.0 I0421 12:07:01.049876 132528837793600 pyconfig.py:432] Config param n_routing_groups: -1 I0421 12:07:01.049890 132528837793600 pyconfig.py:432] Config param n_window_for_audio: 50 I0421 12:07:01.049906 132528837793600 pyconfig.py:432] Config param n_window_infer_for_audio: 800 I0421 12:07:01.049922 132528837793600 pyconfig.py:432] Config param nope_layer_interval: -1 I0421 12:07:01.049938 132528837793600 pyconfig.py:432] Config param norm_topk_prob: False I0421 12:07:01.049952 132528837793600 pyconfig.py:432] Config param normalization_layer_epsilon: 1e-05 I0421 12:07:01.049970 132528837793600 pyconfig.py:432] Config param normalize_embedding_logits: False I0421 12:07:01.049986 132528837793600 pyconfig.py:432] Config param num_attention_heads_for_vit: 16 I0421 12:07:01.050000 132528837793600 pyconfig.py:432] Config param num_batches: 4 I0421 12:07:01.050016 132528837793600 pyconfig.py:432] Config param num_channels_for_vit: 3 I0421 12:07:01.050030 132528837793600 pyconfig.py:432] Config param num_conv_layers_for_audio: 3 I0421 12:07:01.050045 132528837793600 pyconfig.py:432] Config param num_decoder_layers: 1 I0421 12:07:01.050060 132528837793600 pyconfig.py:432] Config param num_diloco_replicas: 1 I0421 12:07:01.050075 132528837793600 pyconfig.py:432] Config param num_epoch: 1 I0421 12:07:01.050090 132528837793600 pyconfig.py:432] Config param num_eval_passes: 1 I0421 12:07:01.050105 132528837793600 pyconfig.py:432] Config param num_experts: 1 I0421 12:07:01.050119 132528837793600 pyconfig.py:432] Config param num_experts_per_tok: 1 I0421 12:07:01.050135 132528837793600 pyconfig.py:432] Config param num_generations: 2 I0421 12:07:01.050151 132528837793600 pyconfig.py:432] Config param num_hidden_layers_for_vit: 34 I0421 12:07:01.050166 132528837793600 pyconfig.py:432] Config param num_iterations: 1 I0421 12:07:01.050182 132528837793600 pyconfig.py:432] Config param num_kv_heads: 2 I0421 12:07:01.050196 132528837793600 pyconfig.py:432] Config param num_layers_per_pipeline_stage: 1 I0421 12:07:01.050211 132528837793600 pyconfig.py:432] Config param num_mel_bins_for_audio: 128 I0421 12:07:01.050227 132528837793600 pyconfig.py:432] Config param num_pipeline_microbatches: -1 I0421 12:07:01.050243 132528837793600 pyconfig.py:432] Config param num_pipeline_repeats: -1 I0421 12:07:01.050258 132528837793600 pyconfig.py:432] Config param num_position_embeddings_for_vit: 1024 I0421 12:07:01.050274 132528837793600 pyconfig.py:432] Config param num_query_heads: 2 I0421 12:07:01.050288 132528837793600 pyconfig.py:432] Config param num_samplers_slices: -1 I0421 12:07:01.050304 132528837793600 pyconfig.py:432] Config param num_slices: 1 I0421 12:07:01.050320 132528837793600 pyconfig.py:432] Config param num_target_devices: 32 I0421 12:07:01.050335 132528837793600 pyconfig.py:432] Config param num_test_batches: 5 I0421 12:07:01.050350 132528837793600 pyconfig.py:432] Config param num_trainer_slices: -1 I0421 12:07:01.050366 132528837793600 pyconfig.py:432] Config param num_vocab_tiling: 1 I0421 12:07:01.050380 132528837793600 pyconfig.py:432] Config param off_policy_steps: 0 I0421 12:07:01.050396 132528837793600 pyconfig.py:432] Config param offline_data_dir: None I0421 12:07:01.050410 132528837793600 pyconfig.py:432] Config param opt_type: OptimizerType.ADAM_PAX I0421 12:07:01.050428 132528837793600 pyconfig.py:432] Config param optimize_mesh_for_tpu_v6e: False I0421 12:07:01.050444 132528837793600 pyconfig.py:432] Config param optimizer_memory_host_offload: False I0421 12:07:01.050459 132528837793600 pyconfig.py:432] Config param original_max_position_embeddings: 4096 I0421 12:07:01.050474 132528837793600 pyconfig.py:432] Config param out_hidden_size_for_vit: 512 I0421 12:07:01.050489 132528837793600 pyconfig.py:432] Config param out_proj: RematLocation.REMAT I0421 12:07:01.050504 132528837793600 pyconfig.py:432] Config param output_dim_for_audio: 512 I0421 12:07:01.050520 132528837793600 pyconfig.py:432] Config param override_logical_axis_rules: False I0421 12:07:01.050534 132528837793600 pyconfig.py:432] Config param override_model_config: True I0421 12:07:01.050550 132528837793600 pyconfig.py:432] Config param packing: True I0421 12:07:01.050565 132528837793600 pyconfig.py:432] Config param pagedattn_head_dim_alignment: 128 I0421 12:07:01.050580 132528837793600 pyconfig.py:432] Config param pagedattn_max_pages_per_group: -1 I0421 12:07:01.050596 132528837793600 pyconfig.py:432] Config param pagedattn_num_pages: 64 I0421 12:07:01.050611 132528837793600 pyconfig.py:432] Config param pagedattn_pages_per_compute_block: 4 I0421 12:07:01.050626 132528837793600 pyconfig.py:432] Config param pagedattn_tokens_per_page: 32 I0421 12:07:01.050641 132528837793600 pyconfig.py:432] Config param param_scan_axis: 1 I0421 12:07:01.050656 132528837793600 pyconfig.py:432] Config param parameter_memory_host_offload: False I0421 12:07:01.050671 132528837793600 pyconfig.py:432] Config param partial_rotary_factor: 1.0 I0421 12:07:01.050686 132528837793600 pyconfig.py:432] Config param patch_size_for_vit: 14 I0421 12:07:01.050702 132528837793600 pyconfig.py:432] Config param penalty_incorrect_answer: -1.0 I0421 12:07:01.050716 132528837793600 pyconfig.py:432] Config param penalty_incorrect_format: -0.5 I0421 12:07:01.050740 132528837793600 pyconfig.py:432] Config param per_device_batch_size: 2 I0421 12:07:01.050756 132528837793600 pyconfig.py:432] Config param per_device_batch_size_increment: 2.0 I0421 12:07:01.050771 132528837793600 pyconfig.py:432] Config param per_device_batch_size_start: 4.0 I0421 12:07:01.050790 132528837793600 pyconfig.py:432] Config param pipeline_delay_activation_forwarding: False I0421 12:07:01.050804 132528837793600 pyconfig.py:432] Config param pipeline_fsdp_ag_once: False I0421 12:07:01.050820 132528837793600 pyconfig.py:432] Config param pipeline_fsdp_ag_per_repeat: False I0421 12:07:01.050834 132528837793600 pyconfig.py:432] Config param pipeline_parallel_layers: 1 I0421 12:07:01.050850 132528837793600 pyconfig.py:432] Config param pixel_shuffle_ratio_for_vit: 0.5 I0421 12:07:01.050867 132528837793600 pyconfig.py:432] Config param posemb_type_for_vit: learn I0421 12:07:01.050883 132528837793600 pyconfig.py:432] Config param position_id_per_seconds: 25 I0421 12:07:01.050897 132528837793600 pyconfig.py:432] Config param prefill_cache_axis_order: 1,2,0,3 I0421 12:07:01.050913 132528837793600 pyconfig.py:432] Config param prefill_cache_dir: I0421 12:07:01.050927 132528837793600 pyconfig.py:432] Config param prefill_chunk_size: 256 I0421 12:07:01.050943 132528837793600 pyconfig.py:432] Config param prefill_slice: v5e-16 I0421 12:07:01.050958 132528837793600 pyconfig.py:432] Config param prefix_caching_dram_byte: 100000000000 I0421 12:07:01.050975 132528837793600 pyconfig.py:432] Config param prefix_caching_hbm_byte: 10000000000 I0421 12:07:01.050990 132528837793600 pyconfig.py:432] Config param profile_cleanly: True I0421 12:07:01.051006 132528837793600 pyconfig.py:432] Config param profile_periodically_period: -1 I0421 12:07:01.051020 132528837793600 pyconfig.py:432] Config param profile_power_events: False I0421 12:07:01.051036 132528837793600 pyconfig.py:432] Config param profiler: ProfilerType.NONE I0421 12:07:01.051053 132528837793600 pyconfig.py:432] Config param profiler_steps: 5 I0421 12:07:01.051069 132528837793600 pyconfig.py:432] Config param projector_dropout_for_vit: 0.0 I0421 12:07:01.051084 132528837793600 pyconfig.py:432] Config param projector_input_dim_for_vit: 4096 I0421 12:07:01.051100 132528837793600 pyconfig.py:432] Config param projector_output_dim_for_vit: 4096 I0421 12:07:01.051114 132528837793600 pyconfig.py:432] Config param prometheus_port: 0 I0421 12:07:01.051130 132528837793600 pyconfig.py:432] Config param prompt: I love to I0421 12:07:01.051144 132528837793600 pyconfig.py:432] Config param pure_nnx: False I0421 12:07:01.051160 132528837793600 pyconfig.py:432] Config param pure_nnx_decoder: False I0421 12:07:01.051174 132528837793600 pyconfig.py:432] Config param q_lora_rank: 0 I0421 12:07:01.051190 132528837793600 pyconfig.py:432] Config param qk_clip_threshold: 100.0 I0421 12:07:01.051204 132528837793600 pyconfig.py:432] Config param qk_nope_head_dim: 128 I0421 12:07:01.051220 132528837793600 pyconfig.py:432] Config param qk_norm_with_scale: True I0421 12:07:01.051236 132528837793600 pyconfig.py:432] Config param qk_rope_head_dim: 64 I0421 12:07:01.051250 132528837793600 pyconfig.py:432] Config param qkv_proj: RematLocation.REMAT I0421 12:07:01.051266 132528837793600 pyconfig.py:432] Config param quant_cfg_path: I0421 12:07:01.051282 132528837793600 pyconfig.py:432] Config param quantization: QuantizationType.NONE I0421 12:07:01.051298 132528837793600 pyconfig.py:432] Config param quantization_local_shard_count: 4 I0421 12:07:01.051314 132528837793600 pyconfig.py:432] Config param quantize_kvcache: False I0421 12:07:01.051330 132528837793600 pyconfig.py:432] Config param query_proj: RematLocation.REMAT I0421 12:07:01.051344 132528837793600 pyconfig.py:432] Config param query_wa_proj: RematLocation.REMAT I0421 12:07:01.051360 132528837793600 pyconfig.py:432] Config param ragged_block_size: 256 I0421 12:07:01.051375 132528837793600 pyconfig.py:432] Config param ragged_buffer_factor: -1.0 I0421 12:07:01.051389 132528837793600 pyconfig.py:432] Config param rampup_end_step: 0 I0421 12:07:01.051405 132528837793600 pyconfig.py:432] Config param rampup_samples_per_increment_to_load: None I0421 12:07:01.051422 132528837793600 pyconfig.py:432] Config param reasoning_end_token: </reasoning> I0421 12:07:01.051437 132528837793600 pyconfig.py:432] Config param reasoning_start_token: <reasoning> I0421 12:07:01.051453 132528837793600 pyconfig.py:432] Config param record_internal_nn_metrics: 0 I0421 12:07:01.051468 132528837793600 pyconfig.py:432] Config param remat_policy: full I0421 12:07:01.051483 132528837793600 pyconfig.py:432] Config param remat_policy_for_vit: minimal I0421 12:07:01.051498 132528837793600 pyconfig.py:432] Config param remove_size_one_mesh_axis_from_type: True I0421 12:07:01.051514 132528837793600 pyconfig.py:432] Config param replicate_quant_scale: False I0421 12:07:01.051529 132528837793600 pyconfig.py:432] Config param replicator_backup_interval_minutes: 0 I0421 12:07:01.051544 132528837793600 pyconfig.py:432] Config param report_heartbeat_metric_for_gcp_monitoring: False I0421 12:07:01.051560 132528837793600 pyconfig.py:432] Config param report_performance_metric_for_gcp_monitoring: False I0421 12:07:01.051576 132528837793600 pyconfig.py:432] Config param reshape_q: False I0421 12:07:01.051590 132528837793600 pyconfig.py:432] Config param return_log_prob: False I0421 12:07:01.051605 132528837793600 pyconfig.py:432] Config param reuse_example_batch: 0 I0421 12:07:01.051622 132528837793600 pyconfig.py:432] Config param reward_exact_answer: 5.0 I0421 12:07:01.051636 132528837793600 pyconfig.py:432] Config param reward_exact_format_match: 3.0 I0421 12:07:01.051652 132528837793600 pyconfig.py:432] Config param reward_partial_format_match: 0.5 I0421 12:07:01.051667 132528837793600 pyconfig.py:432] Config param reward_ratio_guess_to_answer_high: 0.5 I0421 12:07:01.051681 132528837793600 pyconfig.py:432] Config param reward_ratio_guess_to_answer_low: 0.25 I0421 12:07:01.051698 132528837793600 pyconfig.py:432] Config param reward_white_space_format_match: 1.5 I0421 12:07:01.051714 132528837793600 pyconfig.py:432] Config param rl: {'num_generations': 2, 'num_iterations': 1, 'grpo_beta': 0.08, 'grpo_epsilon': 0.2, 'loss_algo': 'grpo', 'use_agentic_rollout': False, 'max_concurrency': 256, 'off_policy_steps': 0, 'system_prompt': '', 'degenerate_group_masking': True, 'epsilon_high': None} I0421 12:07:01.051743 132528837793600 pyconfig.py:432] Config param rollout_data_parallelism: -1 I0421 12:07:01.051759 132528837793600 pyconfig.py:432] Config param rollout_expert_parallelism: 1 I0421 12:07:01.051774 132528837793600 pyconfig.py:432] Config param rollout_micro_batch_size: -1 I0421 12:07:01.051793 132528837793600 pyconfig.py:432] Config param rollout_tensor_parallelism: -1 I0421 12:07:01.051808 132528837793600 pyconfig.py:432] Config param rope_attention_scaling: False I0421 12:07:01.051823 132528837793600 pyconfig.py:432] Config param rope_factor: 40 I0421 12:07:01.051838 132528837793600 pyconfig.py:432] Config param rope_interleave: True I0421 12:07:01.051854 132528837793600 pyconfig.py:432] Config param rope_linear_scaling_factor: 1.0 I0421 12:07:01.051869 132528837793600 pyconfig.py:432] Config param rope_max_timescale: 10000 I0421 12:07:01.051884 132528837793600 pyconfig.py:432] Config param rope_min_timescale: 1 I0421 12:07:01.051899 132528837793600 pyconfig.py:432] Config param rope_theta_for_vit: 10000 I0421 12:07:01.051914 132528837793600 pyconfig.py:432] Config param rope_truncate: True I0421 12:07:01.051930 132528837793600 pyconfig.py:432] Config param rope_type: RopeType.DEFAULT I0421 12:07:01.051947 132528837793600 pyconfig.py:432] Config param rope_use_scale: True I0421 12:07:01.051963 132528837793600 pyconfig.py:432] Config param routed_bias: False I0421 12:07:01.051979 132528837793600 pyconfig.py:432] Config param routed_bias_update_rate: 0.0 I0421 12:07:01.051993 132528837793600 pyconfig.py:432] Config param routed_scaling_factor: 1.0 I0421 12:07:01.052009 132528837793600 pyconfig.py:432] Config param routed_score_func: I0421 12:07:01.052023 132528837793600 pyconfig.py:432] Config param run_name: gpt3-52k_2026-04-21-12-07 I0421 12:07:01.052039 132528837793600 pyconfig.py:432] Config param sa_block_kv: 512 I0421 12:07:01.052053 132528837793600 pyconfig.py:432] Config param sa_block_kv_compute: 512 I0421 12:07:01.052070 132528837793600 pyconfig.py:432] Config param sa_block_kv_dkv: 512 I0421 12:07:01.052084 132528837793600 pyconfig.py:432] Config param sa_block_kv_dkv_compute: 512 I0421 12:07:01.052098 132528837793600 pyconfig.py:432] Config param sa_block_kv_dq: 512 I0421 12:07:01.052114 132528837793600 pyconfig.py:432] Config param sa_block_q: 512 I0421 12:07:01.052130 132528837793600 pyconfig.py:432] Config param sa_block_q_dkv: 512 I0421 12:07:01.052145 132528837793600 pyconfig.py:432] Config param sa_block_q_dq: 512 I0421 12:07:01.052161 132528837793600 pyconfig.py:432] Config param sa_k_layout: HEAD_DIM_MINOR I0421 12:07:01.052177 132528837793600 pyconfig.py:432] Config param sa_q_layout: HEAD_DIM_MINOR I0421 12:07:01.052193 132528837793600 pyconfig.py:432] Config param sa_use_fused_bwd_kernel: False I0421 12:07:01.052207 132528837793600 pyconfig.py:432] Config param sa_v_layout: HEAD_DIM_MINOR I0421 12:07:01.052222 132528837793600 pyconfig.py:432] Config param sampler_devices_fraction: 0.5 I0421 12:07:01.052239 132528837793600 pyconfig.py:432] Config param save_checkpoint_on_completion: True I0421 12:07:01.052253 132528837793600 pyconfig.py:432] Config param save_config_to_gcs: False I0421 12:07:01.052269 132528837793600 pyconfig.py:432] Config param save_quantized_params_path: I0421 12:07:01.052285 132528837793600 pyconfig.py:432] Config param scale_embedding_for_audio: True I0421 12:07:01.052299 132528837793600 pyconfig.py:432] Config param scan_layers: True I0421 12:07:01.052315 132528837793600 pyconfig.py:432] Config param scan_layers_per_stage: False I0421 12:07:01.052330 132528837793600 pyconfig.py:432] Config param scan_pipeline_iterations: True I0421 12:07:01.052345 132528837793600 pyconfig.py:432] Config param scan_pipeline_repeats: False I0421 12:07:01.052359 132528837793600 pyconfig.py:432] Config param set_remat_policy_on_layers_per_stage: False I0421 12:07:01.052375 132528837793600 pyconfig.py:432] Config param set_remat_policy_on_pipeline_iterations: True I0421 12:07:01.052397 132528837793600 pyconfig.py:432] Config param sft_train_on_completion_only: False I0421 12:07:01.052412 132528837793600 pyconfig.py:432] Config param shard_exp_on_fsdp: False I0421 12:07:01.052428 132528837793600 pyconfig.py:432] Config param shard_mode: ShardMode.AUTO I0421 12:07:01.052443 132528837793600 pyconfig.py:432] Config param shard_optimizer_over_data: False I0421 12:07:01.052459 132528837793600 pyconfig.py:432] Config param sharding_strategy: None I0421 12:07:01.052474 132528837793600 pyconfig.py:432] Config param sharding_tolerance: 0.02 I0421 12:07:01.052491 132528837793600 pyconfig.py:432] Config param shardy: True I0421 12:07:01.052505 132528837793600 pyconfig.py:432] Config param share_kv_projections: False I0421 12:07:01.052520 132528837793600 pyconfig.py:432] Config param shared_experts: 0 I0421 12:07:01.052536 132528837793600 pyconfig.py:432] Config param sinkhorn_iterations: 20 I0421 12:07:01.052550 132528837793600 pyconfig.py:432] Config param skip_first_n_steps_for_profiler: 1 I0421 12:07:01.052566 132528837793600 pyconfig.py:432] Config param skip_jax_distributed_system: False I0421 12:07:01.052580 132528837793600 pyconfig.py:432] Config param skip_step_interval: 128 I0421 12:07:01.052596 132528837793600 pyconfig.py:432] Config param skip_step_on_spikes: False I0421 12:07:01.052611 132528837793600 pyconfig.py:432] Config param skip_step_scaling_factor: 6.0 I0421 12:07:01.052626 132528837793600 pyconfig.py:432] Config param sliding_window_size: 0 I0421 12:07:01.052643 132528837793600 pyconfig.py:432] Config param solution_end_token: </answer> I0421 12:07:01.052659 132528837793600 pyconfig.py:432] Config param solution_start_token: <answer> I0421 12:07:01.052676 132528837793600 pyconfig.py:432] Config param source_checkpoint_layout: orbax I0421 12:07:01.052691 132528837793600 pyconfig.py:432] Config param sparse_matmul: True I0421 12:07:01.052706 132528837793600 pyconfig.py:432] Config param spatial_merge_size_for_vit: 2 I0421 12:07:01.052721 132528837793600 pyconfig.py:432] Config param stack_prefill_result_cache: False I0421 12:07:01.052747 132528837793600 pyconfig.py:432] Config param stack_trace_interval_seconds: 600 I0421 12:07:01.052762 132528837793600 pyconfig.py:432] Config param stack_trace_to_cloud: False I0421 12:07:01.052781 132528837793600 pyconfig.py:432] Config param step_deviation_interval_seconds: 30 I0421 12:07:01.052796 132528837793600 pyconfig.py:432] Config param steps: 200000 I0421 12:07:01.052814 132528837793600 pyconfig.py:432] Config param stop_strings: None I0421 12:07:01.052830 132528837793600 pyconfig.py:432] Config param student_overrides: {'model_name': 'llama3.1-8b'} I0421 12:07:01.052846 132528837793600 pyconfig.py:432] Config param student_params_to_update: None I0421 12:07:01.052861 132528837793600 pyconfig.py:432] Config param subslice_shape: I0421 12:07:01.052878 132528837793600 pyconfig.py:432] Config param swap_space_vllm_gb: 2 I0421 12:07:01.052892 132528837793600 pyconfig.py:432] Config param system_prompt: I0421 12:07:01.052907 132528837793600 pyconfig.py:432] Config param target_eval_loss: 0.0 I0421 12:07:01.052923 132528837793600 pyconfig.py:432] Config param teacher_overrides: {'model_name': 'llama3.1-8b'} I0421 12:07:01.052938 132528837793600 pyconfig.py:432] Config param temperature_tuning: False I0421 12:07:01.052954 132528837793600 pyconfig.py:432] Config param temporal_patch_size_for_vit: 2 I0421 12:07:01.052968 132528837793600 pyconfig.py:432] Config param tensorboard_dir: None I0421 12:07:01.052984 132528837793600 pyconfig.py:432] Config param tensors_on_device: None I0421 12:07:01.052999 132528837793600 pyconfig.py:432] Config param tensors_to_offload: None I0421 12:07:01.053013 132528837793600 pyconfig.py:432] Config param test_batch_start_index: 0 I0421 12:07:01.053029 132528837793600 pyconfig.py:432] Config param tile_size_for_vit: 336 I0421 12:07:01.053044 132528837793600 pyconfig.py:432] Config param tokenize_eval_data: True I0421 12:07:01.053059 132528837793600 pyconfig.py:432] Config param tokenize_train_data: True I0421 12:07:01.053073 132528837793600 pyconfig.py:432] Config param tokenizer_path: meta-llama/Llama-3.1-8B I0421 12:07:01.053089 132528837793600 pyconfig.py:432] Config param tokenizer_type: TokenizerType.HUGGINGFACE I0421 12:07:01.053106 132528837793600 pyconfig.py:432] Config param topk_routing_group: -1 I0421 12:07:01.053122 132528837793600 pyconfig.py:432] Config param train_data_columns: ['text'] I0421 12:07:01.053139 132528837793600 pyconfig.py:432] Config param train_fraction: 1.0 I0421 12:07:01.053155 132528837793600 pyconfig.py:432] Config param train_image_column: image I0421 12:07:01.053169 132528837793600 pyconfig.py:432] Config param train_micro_batch_size: -1 I0421 12:07:01.053185 132528837793600 pyconfig.py:432] Config param train_split: train I0421 12:07:01.053199 132528837793600 pyconfig.py:432] Config param trainable_parameters_mask: [] I0421 12:07:01.053214 132528837793600 pyconfig.py:432] Config param trainable_position_size: 2048 I0421 12:07:01.053230 132528837793600 pyconfig.py:432] Config param trainer_devices_fraction: 0.5 I0421 12:07:01.053245 132528837793600 pyconfig.py:432] Config param upload_all_profiler_results: False I0421 12:07:01.053261 132528837793600 pyconfig.py:432] Config param use_2d_fsdp_sharding: False I0421 12:07:01.053275 132528837793600 pyconfig.py:432] Config param use_agentic_rollout: False I0421 12:07:01.053290 132528837793600 pyconfig.py:432] Config param use_audio: False I0421 12:07:01.053306 132528837793600 pyconfig.py:432] Config param use_audio_in_video: False I0421 12:07:01.053322 132528837793600 pyconfig.py:432] Config param use_batch_split_schedule: False I0421 12:07:01.053336 132528837793600 pyconfig.py:432] Config param use_chat_template: False I0421 12:07:01.053352 132528837793600 pyconfig.py:432] Config param use_chunked_prefill: False I0421 12:07:01.053367 132528837793600 pyconfig.py:432] Config param use_custom_sort_vjp: True I0421 12:07:01.053382 132528837793600 pyconfig.py:432] Config param use_dpo: False I0421 12:07:01.053396 132528837793600 pyconfig.py:432] Config param use_gather_mosaic_kernel: False I0421 12:07:01.053412 132528837793600 pyconfig.py:432] Config param use_grpo: True I0421 12:07:01.053428 132528837793600 pyconfig.py:432] Config param use_indexer: False I0421 12:07:01.053443 132528837793600 pyconfig.py:432] Config param use_iota_embed: True I0421 12:07:01.053457 132528837793600 pyconfig.py:432] Config param use_jax_splash: False I0421 12:07:01.053472 132528837793600 pyconfig.py:432] Config param use_max_logit_estimate: -1 I0421 12:07:01.053487 132528837793600 pyconfig.py:432] Config param use_mrope: False I0421 12:07:01.053503 132528837793600 pyconfig.py:432] Config param use_multimodal: False I0421 12:07:01.053518 132528837793600 pyconfig.py:432] Config param use_pathways: True I0421 12:07:01.053532 132528837793600 pyconfig.py:432] Config param use_post_attn_norm: False I0421 12:07:01.053548 132528837793600 pyconfig.py:432] Config param use_post_ffw_norm: False I0421 12:07:01.053564 132528837793600 pyconfig.py:432] Config param use_qk_clip: False I0421 12:07:01.053579 132528837793600 pyconfig.py:432] Config param use_qk_norm: False I0421 12:07:01.053593 132528837793600 pyconfig.py:432] Config param use_qk_norm_in_gdn: True I0421 12:07:01.053609 132528837793600 pyconfig.py:432] Config param use_qwix_quantization: False I0421 12:07:01.053623 132528837793600 pyconfig.py:432] Config param use_ragged_attention: False I0421 12:07:01.053639 132528837793600 pyconfig.py:432] Config param use_random_routing: False I0421 12:07:01.053654 132528837793600 pyconfig.py:432] Config param use_replicator_service: False I0421 12:07:01.053669 132528837793600 pyconfig.py:432] Config param use_ring_of_experts: False I0421 12:07:01.053684 132528837793600 pyconfig.py:432] Config param use_sft: False I0421 12:07:01.053698 132528837793600 pyconfig.py:432] Config param use_splash_scheduler: False I0421 12:07:01.053712 132528837793600 pyconfig.py:432] Config param use_tokamax_gmm: False I0421 12:07:01.053728 132528837793600 pyconfig.py:432] Config param use_tokamax_splash: False I0421 12:07:01.053750 132528837793600 pyconfig.py:432] Config param use_truncation: True I0421 12:07:01.053766 132528837793600 pyconfig.py:432] Config param use_tunix_gradient_accumulation: False I0421 12:07:01.053783 132528837793600 pyconfig.py:432] Config param use_untrainable_positional_embedding: False I0421 12:07:01.053799 132528837793600 pyconfig.py:432] Config param use_vertex_tensorboard: False I0421 12:07:01.053813 132528837793600 pyconfig.py:432] Config param using_pipeline_parallelism: False I0421 12:07:01.053829 132528837793600 pyconfig.py:432] Config param v_head_dim: 128 I0421 12:07:01.053845 132528837793600 pyconfig.py:432] Config param v_norm_with_scale: True I0421 12:07:01.053859 132528837793600 pyconfig.py:432] Config param value_proj: RematLocation.REMAT I0421 12:07:01.053875 132528837793600 pyconfig.py:432] Config param vertex_tensorboard_project: I0421 12:07:01.053889 132528837793600 pyconfig.py:432] Config param vertex_tensorboard_region: I0421 12:07:01.053905 132528837793600 pyconfig.py:432] Config param video_path: I0421 12:07:01.053920 132528837793600 pyconfig.py:432] Config param video_placeholder: <|video|> I0421 12:07:01.053935 132528837793600 pyconfig.py:432] Config param vision_output_dim_for_vit: 4096 I0421 12:07:01.053951 132528837793600 pyconfig.py:432] Config param vision_output_length: -1 I0421 12:07:01.053966 132528837793600 pyconfig.py:432] Config param vllm_additional_config: {} I0421 12:07:01.053982 132528837793600 pyconfig.py:432] Config param vllm_hf_config_path: I0421 12:07:01.053996 132528837793600 pyconfig.py:432] Config param vllm_hf_overrides: {} I0421 12:07:01.054011 132528837793600 pyconfig.py:432] Config param vocab_size: 32000 I0421 12:07:01.054026 132528837793600 pyconfig.py:432] Config param warmup_steps_fraction: 0.1 I0421 12:07:01.054043 132528837793600 pyconfig.py:432] Config param weight_dtype: float32 I0421 12:07:01.054064 132528837793600 pyconfig.py:432] Config param weight_quantization_calibration_method: absmax I0421 12:07:01.054079 132528837793600 pyconfig.py:432] Config param wi_tile_dlhs_batch_seq: 512 I0421 12:07:01.054094 132528837793600 pyconfig.py:432] Config param wi_tile_dlhs_embed_dim: 1024 I0421 12:07:01.054109 132528837793600 pyconfig.py:432] Config param wi_tile_dlhs_mlp_dim: 1024 I0421 12:07:01.054124 132528837793600 pyconfig.py:432] Config param wi_tile_drhs_batch_seq: 512 I0421 12:07:01.054139 132528837793600 pyconfig.py:432] Config param wi_tile_drhs_embed_dim: 1024 I0421 12:07:01.054155 132528837793600 pyconfig.py:432] Config param wi_tile_drhs_mlp_dim: 1024 I0421 12:07:01.054170 132528837793600 pyconfig.py:432] Config param wi_tile_fwd_batch_seq: 512 I0421 12:07:01.054185 132528837793600 pyconfig.py:432] Config param wi_tile_fwd_embed_dim: 1024 I0421 12:07:01.054202 132528837793600 pyconfig.py:432] Config param wi_tile_fwd_mlp_dim: 1024 I0421 12:07:01.054216 132528837793600 pyconfig.py:432] Config param wo_tile_dlhs_batch_seq: 512 I0421 12:07:01.054232 132528837793600 pyconfig.py:432] Config param wo_tile_dlhs_embed_dim: 1024 I0421 12:07:01.054248 132528837793600 pyconfig.py:432] Config param wo_tile_dlhs_mlp_dim: 1024 I0421 12:07:01.054263 132528837793600 pyconfig.py:432] Config param wo_tile_drhs_batch_seq: 512 I0421 12:07:01.054278 132528837793600 pyconfig.py:432] Config param wo_tile_drhs_embed_dim: 1024 I0421 12:07:01.054294 132528837793600 pyconfig.py:432] Config param wo_tile_drhs_mlp_dim: 1024 I0421 12:07:01.054308 132528837793600 pyconfig.py:432] Config param wo_tile_fwd_batch_seq: 512 I0421 12:07:01.054323 132528837793600 pyconfig.py:432] Config param wo_tile_fwd_embed_dim: 1024 I0421 12:07:01.054339 132528837793600 pyconfig.py:432] Config param wo_tile_fwd_mlp_dim: 1024 I0421 12:07:01.054353 132528837793600 pyconfig.py:432] Config param wsd_decay_steps_fraction: 0.1 I0421 12:07:01.054369 132528837793600 pyconfig.py:432] Config param wsd_decay_style: WsdDecayStyle.LINEAR I0421 12:07:01.054386 132528837793600 pyconfig.py:432] Config param xprof_e2e_enable_fw_power_level_event: False I0421 12:07:01.054401 132528837793600 pyconfig.py:432] Config param xprof_e2e_enable_fw_thermal_event: False I0421 12:07:01.054417 132528837793600 pyconfig.py:432] Config param xprof_e2e_enable_fw_throttle_event: False I0421 12:07:01.054432 132528837793600 pyconfig.py:432] Config param xprof_tpu_power_trace_level: 0 I0421 12:07:01.054448 132528837793600 pyconfig.py:432] Config param z_loss_multiplier: 0.0 I0421 12:07:01.054797 132528837793600 tokenizer.py:245] Tokenizer path: meta-llama/Llama-2-7b-chat-hf I0421 12:07:01.054834 132528837793600 tokenizer.py:224] Loading HF tokenizer: meta-llama/Llama-2-7b-chat-hf I0421 12:07:04.853672 132528837793600 _schedule.py:129] A polynomial schedule was set with a non-positive `transition_steps` value; this results in a constant schedule with value `init_value`. I0421 12:07:04.856702 132528837793600 maxtext_utils.py:1718] Num_devices: 32, shape (1, 4, 1, 8, 1, 1, 1, 1, 1, 1, 1, 1, 1) I0421 12:07:04.856836 132528837793600 train_distill.py:608] Applying logical axis rules for model initialization and training... I0421 12:07:04.856909 132528837793600 train_distill.py:612] Loading Student from ... I0421 12:07:04.856937 132528837793600 train_distill.py:169] --- Student Configuration --- I0421 12:07:04.856960 132528837793600 train_distill.py:170] Model Name: gpt3-52k I0421 12:07:04.856981 132528837793600 train_distill.py:171] Dimensions: 1 Layers, 16 Emb Dim, 8 Head Dim I0421 12:07:04.857004 132528837793600 train_distill.py:174] Attention Heads: 2 Query, 2 KV I0421 12:07:04.857023 132528837793600 train_distill.py:175] Vocab Size: 32000 I0421 12:07:04.857040 132528837793600 train_distill.py:176] Checkpoint: I0421 12:07:04.857058 132528837793600 train_distill.py:477] Initializing model: gpt3-52k... I0421 12:07:06.123166 132528837793600 train_distill.py:626] Loading Teacher from gs://lance-maxtext/pt_seed_ckpts/pt_seed_ckpts/pt_seed_ckpt_gpt352k_v32k_linen/checkpoints/4/items... I0421 12:07:06.123273 132528837793600 train_distill.py:169] --- Teacher Configuration --- I0421 12:07:06.123301 132528837793600 train_distill.py:170] Model Name: gpt3-52k I0421 12:07:06.123325 132528837793600 train_distill.py:171] Dimensions: 1 Layers, 16 Emb Dim, 8 Head Dim I0421 12:07:06.123345 132528837793600 train_distill.py:174] Attention Heads: 2 Query, 2 KV I0421 12:07:06.123365 132528837793600 train_distill.py:175] Vocab Size: 32000 I0421 12:07:06.123382 132528837793600 train_distill.py:176] Checkpoint: gs://lance-maxtext/pt_seed_ckpts/pt_seed_ckpts/pt_seed_ckpt_gpt352k_v32k_linen/checkpoints/4/items I0421 12:07:06.123401 132528837793600 train_distill.py:477] Initializing model: gpt3-52k... I0421 12:07:07.271403 132528837793600 pytree_checkpoint_handler.py:577] save_device_host_concurrent_bytes=None I0421 12:07:07.271847 132528837793600 base_pytree_checkpoint_handler.py:411] Created BasePyTreeCheckpointHandler: use_ocdbt=True, use_zarr3=True, pytree_metadata_options=PyTreeMetadataOptions(support_rich_types=False), array_metadata_store=<orbax.checkpoint._src.metadata.array_metadata_store.Store object at 0x78880ecd5910>, enable_pinned_host_transfer=False, save_concurrent_bytes: 96000000000 (89.4 GiB), restore_concurrent_bytes: 96000000000 (89.4 GiB) I0421 12:07:07.271906 132528837793600 abstract_checkpointer.py:35] orbax-checkpoint version: 0.11.28 W0421 12:07:07.799982 132528837793600 checkpoint.py:202] Metadata file does not exist: gs://lance-maxtext/pt_seed_ckpts/pt_seed_ckpts/pt_seed_ckpt_gpt352k_v32k_linen/checkpoints/4/items/_CHECKPOINT_METADATA I0421 12:07:08.347514 2190 google_auth_provider.cc:181] Running on GCE, using service account 562977990677-compute@developer.gserviceaccount.com I0421 12:07:09.915121 132528837793600 checkpointer.py:304] Restoring checkpoint from gs://lance-maxtext/pt_seed_ckpts/pt_seed_ckpts/pt_seed_ckpt_gpt352k_v32k_linen/checkpoints/4/items. W0421 12:07:12.087010 132528837793600 transform_utils.py:230] The transformations API will eventually be replaced by an upgraded design. The current API will not be removed until this point, but it will no longer be actively worked on. I0421 12:07:12.087378 132528837793600 transform_utils.py:288] The following keys are not loaded from the original tree after applying specified transforms: params/params/decoder/to_nnx__rngs/aqt/count, params/params/decoder/to_nnx__rngs/aqt/key, params/params/decoder/to_nnx__rngs/dropout/count, params/params/decoder/to_nnx__rngs/dropout/key, params/params/decoder/to_nnx__rngs/params/count, params/params/decoder/to_nnx__rngs/params/key I0421 12:07:12.679553 132528837793600 checkpointer.py:318] Finished restoring checkpoint in 3.56 seconds from gs://lance-maxtext/pt_seed_ckpts/pt_seed_ckpts/pt_seed_ckpt_gpt352k_v32k_linen/checkpoints/4/items. I0421 12:07:13.368643 132528837793600 train_distill.py:652] Initializing Data Iterators via MaxText pipeline... I0421 12:07:13.433447 132528837793600 config.py:112] TensorFlow version 2.20.0 available. I0421 12:07:13.433944 132528837793600 config.py:125] JAX version 0.8.3 available. E0421 12:07:15.534079 132528837793600 packing.py:209] PackAndBatchOperation is deprecated. Please use lazy_dataset.FirstFitPackIterDataset instead. I0421 12:07:15.534299 132528837793600 data_loader.py:408] Adding CopyNumPyArrayToSharedMemory MapTransform. I0421 12:07:15.537320 132528837793600 train_distill.py:422] Input Pipeline Checkpointing: DISABLED I0421 12:07:15.537385 132528837793600 train_distill.py:426] Reason: Iterator 'MultiHostDataLoadIterator' is not recognized as Grain (dataset_type='DatasetType.HF', has_save=False) I0421 12:07:15.537449 132528837793600 pytree_checkpoint_handler.py:577] save_device_host_concurrent_bytes=None I0421 12:07:15.537527 132528837793600 base_pytree_checkpoint_handler.py:411] Created BasePyTreeCheckpointHandler: use_ocdbt=True, use_zarr3=False, pytree_metadata_options=PyTreeMetadataOptions(support_rich_types=False), array_metadata_store=<orbax.checkpoint._src.metadata.array_metadata_store.Store object at 0x78880ecd5910>, enable_pinned_host_transfer=False, save_concurrent_bytes: 96000000000 (89.4 GiB), restore_concurrent_bytes: 96000000000 (89.4 GiB) I0421 12:07:15.537569 132528837793600 pytree_checkpoint_handler.py:577] save_device_host_concurrent_bytes=None I0421 12:07:15.537600 132528837793600 base_pytree_checkpoint_handler.py:411] Created BasePyTreeCheckpointHandler: use_ocdbt=True, use_zarr3=False, pytree_metadata_options=PyTreeMetadataOptions(support_rich_types=False), array_metadata_store=<orbax.checkpoint._src.metadata.array_metadata_store.Store object at 0x78880ecd5910>, enable_pinned_host_transfer=False, save_concurrent_bytes: 96000000000 (89.4 GiB), restore_concurrent_bytes: 96000000000 (89.4 GiB) I0421 12:07:15.537642 132528837793600 checkpoint_manager.py:702] [process=7][thread=MainThread] CheckpointManager init: checkpointers=None, item_names=None, item_handlers={'model_params': <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x787ccfcc1520>, 'optimizer_state': <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x78680c05f350>, 'custom_metadata': <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x78680c05f290>}, handler_registry=None I0421 12:07:15.537873 132528837793600 composite_checkpoint_handler.py:237] Deferred registration for item: "model_params". Adding handler `<orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x787ccfcc1520>` for item "model_params" and save args `<class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeSaveArgs'>` and restore args `<class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeRestoreArgs'>` to `_handler_registry`. I0421 12:07:15.537919 132528837793600 composite_checkpoint_handler.py:237] Deferred registration for item: "optimizer_state". Adding handler `<orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x78680c05f350>` for item "optimizer_state" and save args `<class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeSaveArgs'>` and restore args `<class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeRestoreArgs'>` to `_handler_registry`. I0421 12:07:15.537946 132528837793600 composite_checkpoint_handler.py:237] Deferred registration for item: "custom_metadata". Adding handler `<orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x78680c05f290>` for item "custom_metadata" and save args `<class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonSaveArgs'>` and restore args `<class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonRestoreArgs'>` to `_handler_registry`. I0421 12:07:15.537970 132528837793600 composite_checkpoint_handler.py:237] Deferred registration for item: "metrics". Adding handler `<orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x78680c05eed0>` for item "metrics" and save args `<class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonSaveArgs'>` and restore args `<class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonRestoreArgs'>` to `_handler_registry`. I0421 12:07:15.537998 132528837793600 composite_checkpoint_handler.py:505] Initialized registry DefaultCheckpointHandlerRegistry({('model_params', <class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeSaveArgs'>): <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x787ccfcc1520>, ('model_params', <class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeRestoreArgs'>): <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x787ccfcc1520>, ('optimizer_state', <class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeSaveArgs'>): <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x78680c05f350>, ('optimizer_state', <class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeRestoreArgs'>): <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x78680c05f350>, ('custom_metadata', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonSaveArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x78680c05f290>, ('custom_metadata', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonRestoreArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x78680c05f290>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonSaveArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x78680c05eed0>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonRestoreArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x78680c05eed0>}). I0421 12:07:15.538474 132528837793600 async_checkpointer.py:177] [process=7][thread=MainThread] Using barrier_sync_fn: <function get_barrier_sync_fn.<locals>._fn at 0x7867d0099620> timeout: 600 secs and primary_host=0 for async checkpoint writes I0421 12:07:18.008699 132528837793600 checkpoint_manager.py:1788] Found 0 checkpoint steps in gs://lance-maxtext/pt_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114106/pt_distill_nnx_xpk_feat_nnx_post_train_fixes_20260421_114106_07_distill_smoke/checkpoints I0421 12:07:18.463287 132528837793600 checkpoint_manager.py:921] [process=7][thread=MainThread] CheckpointManager created, primary_host=0, CheckpointManagerOptions=CheckpointManagerOptions(save_interval_steps=2000, max_to_keep=None, keep_time_interval=None, keep_period=None, should_keep_fn=None, best_fn=None, best_mode='max', keep_checkpoints_without_metrics=True, step_prefix=None, step_format_fixed_length=None, step_name_format=None, create=True, cleanup_tmp_directories=False, save_on_steps=frozenset(), single_host_load_and_broadcast=False, todelete_subdir=None, todelete_full_path=None, enable_hns=False, enable_background_delete=False, read_only=False, enable_async_checkpointing=True, async_options=None, multiprocessing_options=MultiprocessingOptions(primary_host=0, active_processes=None, barrier_sync_key_prefix=None), should_save_fn=None, file_options=FileOptions(path_permission_mode=None), save_root_metadata=True, temporary_path_class=None, save_decision_policy=None, preservation_policy=None, prevent_write_metrics=False, enable_should_save_is_saving_in_progress_check=True, enable_per_process_directory_creation=False), root_directory=gs://lance-maxtext/pt_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114106/pt_distill_nnx_xpk_feat_nnx_post_train_fixes_20260421_114106_07_distill_smoke/checkpoints: <orbax.checkpoint.checkpoint_manager.CheckpointManager object at 0x78680c05f260> I0421 12:07:18.463460 132528837793600 pytree_checkpoint_handler.py:577] save_device_host_concurrent_bytes=None I0421 12:07:18.463533 132528837793600 base_pytree_checkpoint_handler.py:411] Created BasePyTreeCheckpointHandler: use_ocdbt=True, use_zarr3=False, pytree_metadata_options=PyTreeMetadataOptions(support_rich_types=False), array_metadata_store=<orbax.checkpoint._src.metadata.array_metadata_store.Store object at 0x78880ecd5910>, enable_pinned_host_transfer=False, save_concurrent_bytes: 96000000000 (89.4 GiB), restore_concurrent_bytes: 96000000000 (89.4 GiB) I0421 12:07:18.463569 132528837793600 pytree_checkpoint_handler.py:577] save_device_host_concurrent_bytes=None I0421 12:07:18.463602 132528837793600 base_pytree_checkpoint_handler.py:411] Created BasePyTreeCheckpointHandler: use_ocdbt=True, use_zarr3=False, pytree_metadata_options=PyTreeMetadataOptions(support_rich_types=False), array_metadata_store=<orbax.checkpoint._src.metadata.array_metadata_store.Store object at 0x78880ecd5910>, enable_pinned_host_transfer=False, save_concurrent_bytes: 96000000000 (89.4 GiB), restore_concurrent_bytes: 96000000000 (89.4 GiB) I0421 12:07:18.463638 132528837793600 checkpoint_manager.py:1983] [process=7][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning. I0421 12:07:18.463693 132528837793600 checkpoint.py:459] Closing _NonBlockingMetadataStore(enable_write=True, _write_lock=<locked _thread.RLock object owner=132528837793600 count=1 at 0x7871a642ab40>, _store_impl=<orbax.checkpoint._src.metadata.checkpoint._MetadataStoreImpl object at 0x78680c05f050>, _single_thread_executor=<concurrent.futures.thread.ThreadPoolExecutor object at 0x78680c05f020>, _write_futures=[]) I0421 12:07:18.464090 132528837793600 checkpoint.py:459] Closing _NonBlockingMetadataStore(enable_write=True, _write_lock=<locked _thread.RLock object owner=132528837793600 count=1 at 0x7871a642ab40>, _store_impl=<orbax.checkpoint._src.metadata.checkpoint._MetadataStoreImpl object at 0x78680c05f050>, _single_thread_executor=<concurrent.futures.thread.ThreadPoolExecutor object at 0x78680c05f020>, _write_futures=[]) I0421 12:07:18.464118 132528837793600 checkpoint.py:459] Closing _NonBlockingMetadataStore(enable_write=True, _write_lock=<locked _thread.RLock object owner=132528837793600 count=1 at 0x7871a642ab40>, _store_impl=<orbax.checkpoint._src.metadata.checkpoint._MetadataStoreImpl object at 0x78680c05f050>, _single_thread_executor=<concurrent.futures.thread.ThreadPoolExecutor object at 0x78680c05f020>, _write_futures=[]) I0421 12:07:18.464151 132528837793600 checkpoint_manager.py:702] [process=7][thread=MainThread] CheckpointManager init: checkpointers=None, item_names=None, item_handlers={'model_params': <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x78680c05f230>, 'optimizer_state': <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x78680c05e630>, 'custom_metadata': <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x78680c05d7c0>, 'iter': <maxtext.common.checkpointing.GrainCheckpointHandler object at 0x78680c05d2b0>}, handler_registry=None I0421 12:07:18.464259 132528837793600 composite_checkpoint_handler.py:237] Deferred registration for item: "model_params". Adding handler `<orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x78680c05f230>` for item "model_params" and save args `<class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeSaveArgs'>` and restore args `<class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeRestoreArgs'>` to `_handler_registry`. I0421 12:07:18.464295 132528837793600 composite_checkpoint_handler.py:237] Deferred registration for item: "optimizer_state". Adding handler `<orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x78680c05e630>` for item "optimizer_state" and save args `<class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeSaveArgs'>` and restore args `<class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeRestoreArgs'>` to `_handler_registry`. I0421 12:07:18.464318 132528837793600 composite_checkpoint_handler.py:237] Deferred registration for item: "custom_metadata". Adding handler `<orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x78680c05d7c0>` for item "custom_metadata" and save args `<class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonSaveArgs'>` and restore args `<class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonRestoreArgs'>` to `_handler_registry`. I0421 12:07:18.464346 132528837793600 composite_checkpoint_handler.py:237] Deferred registration for item: "iter". Adding handler `<maxtext.common.checkpointing.GrainCheckpointHandler object at 0x78680c05d2b0>` for item "iter" and save args `<class 'maxtext.common.checkpointing.GrainCheckpointSave'>` and restore args `<class 'maxtext.common.checkpointing.GrainCheckpointRestore'>` to `_handler_registry`. I0421 12:07:18.464369 132528837793600 composite_checkpoint_handler.py:237] Deferred registration for item: "metrics". Adding handler `<orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x78680c05cce0>` for item "metrics" and save args `<class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonSaveArgs'>` and restore args `<class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonRestoreArgs'>` to `_handler_registry`. I0421 12:07:18.464394 132528837793600 composite_checkpoint_handler.py:505] Initialized registry DefaultCheckpointHandlerRegistry({('model_params', <class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeSaveArgs'>): <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x78680c05f230>, ('model_params', <class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeRestoreArgs'>): <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x78680c05f230>, ('optimizer_state', <class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeSaveArgs'>): <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x78680c05e630>, ('optimizer_state', <class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeRestoreArgs'>): <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x78680c05e630>, ('custom_metadata', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonSaveArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x78680c05d7c0>, ('custom_metadata', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonRestoreArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x78680c05d7c0>, ('iter', <class 'maxtext.common.checkpointing.GrainCheckpointSave'>): <maxtext.common.checkpointing.GrainCheckpointHandler object at 0x78680c05d2b0>, ('iter', <class 'maxtext.common.checkpointing.GrainCheckpointRestore'>): <maxtext.common.checkpointing.GrainCheckpointHandler object at 0x78680c05d2b0>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonSaveArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x78680c05cce0>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonRestoreArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x78680c05cce0>}). I0421 12:07:18.464464 132528837793600 async_checkpointer.py:177] [process=7][thread=MainThread] Using barrier_sync_fn: <function get_barrier_sync_fn.<locals>._fn at 0x7867d0099760> timeout: 600 secs and primary_host=0 for async checkpoint writes I0421 12:07:18.839772 132528837793600 checkpoint_manager.py:1788] Found 0 checkpoint steps in gs://lance-maxtext/pt_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114106/pt_distill_nnx_xpk_feat_nnx_post_train_fixes_20260421_114106_07_distill_smoke/checkpoints I0421 12:07:18.852114 132528837793600 checkpoint_manager.py:921] [process=7][thread=MainThread] CheckpointManager created, primary_host=0, CheckpointManagerOptions=CheckpointManagerOptions(save_interval_steps=2000, max_to_keep=None, keep_time_interval=None, keep_period=None, should_keep_fn=None, best_fn=None, best_mode='max', keep_checkpoints_without_metrics=True, step_prefix=None, step_format_fixed_length=None, step_name_format=None, create=True, cleanup_tmp_directories=False, save_on_steps=frozenset(), single_host_load_and_broadcast=False, todelete_subdir=None, todelete_full_path=None, enable_hns=False, enable_background_delete=False, read_only=False, enable_async_checkpointing=True, async_options=None, multiprocessing_options=MultiprocessingOptions(primary_host=0, active_processes=None, barrier_sync_key_prefix=None), should_save_fn=None, file_options=FileOptions(path_permission_mode=None), save_root_metadata=True, temporary_path_class=None, save_decision_policy=None, preservation_policy=None, prevent_write_metrics=False, enable_should_save_is_saving_in_progress_check=True, enable_per_process_directory_creation=False), root_directory=gs://lance-maxtext/pt_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114106/pt_distill_nnx_xpk_feat_nnx_post_train_fixes_20260421_114106_07_distill_smoke/checkpoints: <orbax.checkpoint.checkpoint_manager.CheckpointManager object at 0x78680c05cf50> I0421 12:07:18.852577 132528837793600 train_distill.py:703] Starting Distillation Training... I0421 12:07:18.852684 132528837793600 peft_trainer.py:584] Training with mesh: Mesh('diloco': 1, 'data': 4, 'stage': 1, 'fsdp': 8, 'fsdp_transpose': 1, 'sequence': 1, 'context': 1, 'context_autoregressive': 1, 'tensor': 1, 'tensor_transpose': 1, 'tensor_sequence': 1, 'expert': 1, 'autoregressive': 1, axis_types=(Auto, Auto, Auto, Auto, Auto, Auto, Auto, Auto, Auto, Auto, Auto, Auto, Auto)) I0421 12:07:18.973792 132528837793600 peft_trainer.py:594] Compiled train_step cache size: 0 I0421 12:07:18.975484 132385141139200 grain_pool.py:367] Grain pool will use 1 processes. I0421 12:07:19.002618 132385141139200 grain_pool.py:440] Grain pool will start child processes. I0421 12:07:19.007725 132385141139200 grain_pool.py:448] Grain pool started all child processes. 2026-04-21 12:07:25.058097: E external/local_xla/xla/stream_executor/cuda/cuda_platform.cc:51] failed call to cuInit: INTERNAL: CUDA error: Failed call to cuInit: UNKNOWN ERROR (303) /deps/src/maxtext/trainers/post_train/distillation/train_distill.py:281: DeprecationWarning: '.value' access is now deprecated. For Variable[Array] instances use: variable[...] For other Variable types use: variable.get_value() current_step = model.training_step.value I0421 12:07:31.332397 132528837793600 checkpoint_manager.py:1983] [process=7][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning. I0421 12:07:31.334380 132528837793600 checkpoint_manager.py:1501] [process=7] Saving checkpoint at step 1 I0421 12:07:31.337377 132528837793600 async_checkpointer.py:452] [process=7] Started async saving checkpoint to gs://lance-maxtext/pt_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114106/pt_distill_nnx_xpk_feat_nnx_post_train_fixes_20260421_114106_07_distill_smoke/checkpoints/1. I0421 12:07:31.870894 132528837793600 signaling_client.py:364] Using JaxDistributedSignalingClient I0421 12:07:31.871953 132528837793600 jax_array_handlers.py:347] Scheduling D2H of 37 prioritized jax.Array. I0421 12:07:31.872012 132528837793600 replica_slices.py:410] Transferring arrays to host memory with options: use_replica_parallel=True, min_slice_bytes_for_replica_parallel=None, max_replicas_for_replica_parallel=None, enable_pinned_host_transfer=False I0421 12:07:32.537761 132528837793600 base_pytree_checkpoint_handler.py:153] [process=7][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 0.666957s I0421 12:07:32.542562 132528837793600 base_pytree_checkpoint_handler.py:128] [process=7] /jax/checkpoint/write/blocking_gbytes_per_sec: 574.358 KiB/s (total gbytes: 535.1 KiB) (time elapsed: 931 milliseconds) (per-host) I0421 12:07:32.542628 132528837793600 base_pytree_checkpoint_handler.py:732] [process=7][thread=MainThread] Initiated Pytree async_save. Time taken: 0.931692s (batch_requests_ready=0.256245s, total_serialization_initiated=0.675333s, others=0.000114s) I0421 12:07:32.544438 132528837793600 jax_array_handlers.py:347] Scheduling D2H of 46 prioritized jax.Array. I0421 12:07:32.544494 132528837793600 replica_slices.py:410] Transferring arrays to host memory with options: use_replica_parallel=True, min_slice_bytes_for_replica_parallel=None, max_replicas_for_replica_parallel=None, enable_pinned_host_transfer=False I0421 12:07:32.549323 132528837793600 base_pytree_checkpoint_handler.py:153] [process=7][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 0.006618s I0421 12:07:32.549433 132528837793600 base_pytree_checkpoint_handler.py:128] [process=7] /jax/checkpoint/write/blocking_gbytes_per_sec: 284.015 KiB/s (total gbytes: 267.5 KiB) (time elapsed: 941 milliseconds) (per-host) I0421 12:07:32.549478 132528837793600 base_pytree_checkpoint_handler.py:732] [process=7][thread=MainThread] Initiated Pytree async_save. Time taken: 0.942028s (batch_requests_ready=0.930490s, total_serialization_initiated=0.011469s, others=0.000070s) I0421 12:07:32.549571 132528837793600 composite_checkpoint_handler.py:715] [process=7][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 0.946015s (all_items=0.000023s, per_item={'model_params': '0.00001860', 'optimizer_state': '0.00000405'}, temp_paths=0.945992) I0421 12:07:32.550561 132378172581632 async_checkpointer.py:79] [process=7][thread=async_save] Background save thread started. I0421 12:07:32.550748 132528837793600 async_checkpointer.py:561] Finished blocking save. Time taken: 1.216280s. Continuing background save to gs://lance-maxtext/pt_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114106/pt_distill_nnx_xpk_feat_nnx_post_train_fixes_20260421_114106_07_distill_smoke/checkpoints/1. I0421 12:07:32.576971 132528837793600 checkpoint_manager.py:1549] [process=7][thread=MainThread][step=1] Starting CheckpointManager Save Finalize thread=save_finalize I0421 12:07:32.577257 132378197759744 async_checkpointer.py:265] [process=7][thread=save_finalize] Waiting for background save thread=async_save. I0421 12:07:32.577411 132528837793600 standard_logger.py:34] {'step': 1, 'event_type': 'save', 'directory': 'gs://lance-maxtext/pt_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114106/pt_distill_nnx_xpk_feat_nnx_post_train_fixes_20260421_114106_07_distill_smoke/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776773251.3323634, 'wait_for_prev_duration_secs': 0.0001361370086669922, 'time_between_consecutive_saves_sec': None, 'checkpointer_blocking_start_time': 1776773251.33442, 'checkpointer_blocking_duration_secs': 1.2164409160614014, 'get_old_steps_start_time': 1776773252.5508857, 'get_old_steps_duration_secs': 7.987022399902344e-05, 'checkpoint_manager_blocking_start_time': 1776773251.0773523, 'checkpoint_manager_blocking_duration_secs': 1.5000245571136475} /deps/src/maxtext/trainers/post_train/distillation/train_distill.py:281: DeprecationWarning: '.value' access is now deprecated. For Variable[Array] instances use: variable[...] For other Variable types use: variable.get_value() current_step = model.training_step.value I0421 12:07:35.693797 132528837793600 peft_trainer.py:474] Train step 1 training loss: 15.990566 - training perplexity: 8802675.000000 I0421 12:07:35.714111 132528837793600 peft_trainer.py:474] Train step 2 training loss: 15.974588 - training perplexity: 8663145.000000 I0421 12:07:35.744526 132528837793600 peft_trainer.py:474] Train step 3 training loss: 16.008877 - training perplexity: 8965342.000000 I0421 12:07:35.763965 132528837793600 peft_trainer.py:474] Train step 4 training loss: 16.001873 - training perplexity: 8902770.000000 I0421 12:07:35.768807 132528837793600 peft_trainer.py:733] Train loop finished in: 16.7945 seconds I0421 12:07:35.769370 132528837793600 train_distill.py:712] Saving final checkpoint to gs://lance-maxtext/pt_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114106/pt_distill_nnx_xpk_feat_nnx_post_train_fixes_20260421_114106_07_distill_smoke/checkpoints/... I0421 12:07:37.781939 132385706333952 array_metadata_store.py:203] [process=7][thread=array_type_handler] Wrote 37 array_metadata.ArrayMetadata to gs://lance-maxtext/pt_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114106/pt_distill_nnx_xpk_feat_nnx_post_train_fixes_20260421_114106_07_distill_smoke/checkpoints/1/optimizer_state/array_metadatas/process_7 I0421 12:07:37.840428 132378180974336 array_metadata_store.py:203] [process=7][thread=array_type_handler] Wrote 46 array_metadata.ArrayMetadata to gs://lance-maxtext/pt_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114106/pt_distill_nnx_xpk_feat_nnx_post_train_fixes_20260421_114106_07_distill_smoke/checkpoints/1/model_params/array_metadatas/process_7 I0421 12:07:37.841520 132378172581632 base_pytree_checkpoint_handler.py:128] [process=7] /jax/checkpoint/write/gbytes_per_sec: 42.915 KiB/s (total gbytes: 267.5 KiB) (time elapsed: 6 seconds) (per-host) I0421 12:07:37.841630 132378172581632 base_pytree_checkpoint_handler.py:128] [process=7] /jax/checkpoint/write/gbytes_per_sec: 85.875 KiB/s (total gbytes: 535.1 KiB) (time elapsed: 6 seconds) (per-host) I0421 12:07:37.841668 132378172581632 async_checkpointer.py:90] [process=7][thread=async_save] 4 Handler Commit operations completed. Time taken: 5.290997s. I0421 12:07:48.109959 132528837793600 checkpoint_manager.py:1994] [process=7][thread=MainThread][step=1][wait_until_finished] Waiting for Save Finalize thread (save_finalize) to complete. I0421 12:07:48.979501 132378172581632 async_checkpointer.py:144] [process=7][thread=async_save] Background save thread done. Time taken: 16.428812s. I0421 12:07:48.979861 132378197759744 async_checkpointer.py:273] [process=7][thread=save_finalize] Done with waiting for background save thread=async_save. I0421 12:07:48.979979 132378197759744 async_checkpointer.py:283] [process=7][thread=save_finalize] No errors found in background save thread=async_save. I0421 12:07:48.980029 132378197759744 checkpoint_manager.py:2103] [process=7][thread=save_finalize][step=1] CheckpointManager Save Finalize is syncing with other hosts... I0421 12:07:48.981458 132378197759744 checkpoint_manager.py:2112] [process=7][thread=save_finalize][step=1] CheckpointManager Save Finalize is done on all hosts. I0421 12:07:48.981636 132528837793600 checkpoint_manager.py:2006] [process=7][thread=MainThread][step=1][wait_until_finished] Done waiting for Save Finalize thread (save_finalize) running at step=1. I0421 12:07:48.983293 132528837793600 checkpoint_manager.py:1501] [process=7] Saving checkpoint at step 5 I0421 12:07:48.986681 132528837793600 async_checkpointer.py:452] [process=7] Started async saving checkpoint to gs://lance-maxtext/pt_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114106/pt_distill_nnx_xpk_feat_nnx_post_train_fixes_20260421_114106_07_distill_smoke/checkpoints/5. I0421 12:07:49.529242 132528837793600 jax_array_handlers.py:347] Scheduling D2H of 37 prioritized jax.Array. I0421 12:07:49.529339 132528837793600 replica_slices.py:410] Transferring arrays to host memory with options: use_replica_parallel=True, min_slice_bytes_for_replica_parallel=None, max_replicas_for_replica_parallel=None, enable_pinned_host_transfer=False I0421 12:07:50.184609 132528837793600 base_pytree_checkpoint_handler.py:153] [process=7][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 0.656202s I0421 12:07:50.188256 132528837793600 base_pytree_checkpoint_handler.py:128] [process=7] /jax/checkpoint/write/blocking_gbytes_per_sec: 585.673 KiB/s (total gbytes: 535.1 KiB) (time elapsed: 913 milliseconds) (per-host) I0421 12:07:50.188319 132528837793600 base_pytree_checkpoint_handler.py:732] [process=7][thread=MainThread] Initiated Pytree async_save. Time taken: 0.913687s (batch_requests_ready=0.252215s, total_serialization_initiated=0.661368s, others=0.000103s) I0421 12:07:50.190164 132528837793600 jax_array_handlers.py:347] Scheduling D2H of 46 prioritized jax.Array. I0421 12:07:50.190222 132528837793600 replica_slices.py:410] Transferring arrays to host memory with options: use_replica_parallel=True, min_slice_bytes_for_replica_parallel=None, max_replicas_for_replica_parallel=None, enable_pinned_host_transfer=False I0421 12:07:50.195029 132528837793600 base_pytree_checkpoint_handler.py:153] [process=7][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 0.006599s I0421 12:07:50.195132 132528837793600 base_pytree_checkpoint_handler.py:128] [process=7] /jax/checkpoint/write/blocking_gbytes_per_sec: 289.707 KiB/s (total gbytes: 267.5 KiB) (time elapsed: 923 milliseconds) (per-host) I0421 12:07:50.195176 132528837793600 base_pytree_checkpoint_handler.py:732] [process=7][thread=MainThread] Initiated Pytree async_save. Time taken: 0.923516s (batch_requests_ready=0.913138s, total_serialization_initiated=0.010311s, others=0.000066s) I0421 12:07:50.195298 132528837793600 composite_checkpoint_handler.py:715] [process=7][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 0.927532s (all_items=0.000014s, per_item={'model_params': '0.00001121', 'optimizer_state': '0.00000286'}, temp_paths=0.927518) I0421 12:07:50.196250 132377099613952 async_checkpointer.py:79] [process=7][thread=async_save] Background save thread started. I0421 12:07:50.196407 132528837793600 async_checkpointer.py:561] Finished blocking save. Time taken: 1.213042s. Continuing background save to gs://lance-maxtext/pt_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114106/pt_distill_nnx_xpk_feat_nnx_post_train_fixes_20260421_114106_07_distill_smoke/checkpoints/5. I0421 12:07:50.222283 132528837793600 checkpoint_manager.py:1549] [process=7][thread=MainThread][step=5] Starting CheckpointManager Save Finalize thread=save_finalize I0421 12:07:50.222575 132378197759744 async_checkpointer.py:265] [process=7][thread=save_finalize] Waiting for background save thread=async_save. I0421 12:07:50.222758 132528837793600 standard_logger.py:34] {'step': 5, 'event_type': 'save', 'directory': 'gs://lance-maxtext/pt_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114106/pt_distill_nnx_xpk_feat_nnx_post_train_fixes_20260421_114106_07_distill_smoke/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776773268.1099184, 'wait_for_prev_duration_secs': 0.8718478679656982, 'time_between_consecutive_saves_sec': None, 'checkpointer_blocking_start_time': 1776773268.983333, 'checkpointer_blocking_duration_secs': 1.213179111480713, 'get_old_steps_start_time': 1776773270.1965382, 'get_old_steps_duration_secs': 7.891654968261719e-05, 'checkpoint_manager_blocking_start_time': 1776773255.7732153, 'checkpoint_manager_blocking_duration_secs': 14.449496269226074} I0421 12:07:50.222946 132528837793600 checkpoint_manager.py:1994] [process=7][thread=MainThread][step=5][wait_until_finished] Waiting for Save Finalize thread (save_finalize) to complete. I0421 12:07:54.718913 132378189367040 array_metadata_store.py:203] [process=7][thread=array_type_handler] Wrote 46 array_metadata.ArrayMetadata to gs://lance-maxtext/pt_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114106/pt_distill_nnx_xpk_feat_nnx_post_train_fixes_20260421_114106_07_distill_smoke/checkpoints/5/model_params/array_metadatas/process_7 I0421 12:07:54.720081 132377099613952 base_pytree_checkpoint_handler.py:128] [process=7] /jax/checkpoint/write/gbytes_per_sec: 49.103 KiB/s (total gbytes: 267.5 KiB) (time elapsed: 5 seconds) (per-host) I0421 12:07:54.755824 132385706333952 array_metadata_store.py:203] [process=7][thread=array_type_handler] Wrote 37 array_metadata.ArrayMetadata to gs://lance-maxtext/pt_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114106/pt_distill_nnx_xpk_feat_nnx_post_train_fixes_20260421_114106_07_distill_smoke/checkpoints/5/optimizer_state/array_metadatas/process_7 I0421 12:07:54.756923 132377099613952 base_pytree_checkpoint_handler.py:128] [process=7] /jax/checkpoint/write/gbytes_per_sec: 97.599 KiB/s (total gbytes: 535.1 KiB) (time elapsed: 5 seconds) (per-host) I0421 12:07:54.757021 132377099613952 async_checkpointer.py:90] [process=7][thread=async_save] 4 Handler Commit operations completed. Time taken: 4.560652s. I0421 12:08:06.003826 132377099613952 async_checkpointer.py:144] [process=7][thread=async_save] Background save thread done. Time taken: 15.807448s. I0421 12:08:06.004141 132378197759744 async_checkpointer.py:273] [process=7][thread=save_finalize] Done with waiting for background save thread=async_save. I0421 12:08:06.004258 132378197759744 async_checkpointer.py:283] [process=7][thread=save_finalize] No errors found in background save thread=async_save. I0421 12:08:06.004304 132378197759744 checkpoint_manager.py:2103] [process=7][thread=save_finalize][step=5] CheckpointManager Save Finalize is syncing with other hosts... I0421 12:08:06.005948 132378197759744 checkpoint_manager.py:2112] [process=7][thread=save_finalize][step=5] CheckpointManager Save Finalize is done on all hosts. I0421 12:08:06.006130 132528837793600 checkpoint_manager.py:2006] [process=7][thread=MainThread][step=5][wait_until_finished] Done waiting for Save Finalize thread (save_finalize) running at step=5. I0421 12:08:06.006251 132528837793600 train_distill.py:724] Final checkpoint saved. I0421 12:08:06.008454 132528837793600 peft_trainer.py:474] Train step 5 training loss: 15.987230 - training perplexity: 8773359.000000 I0421 12:08:06.008838 132528837793600 checkpoint_manager.py:1983] [process=7][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning. I0421 12:08:06.008913 132528837793600 checkpoint.py:459] Closing _NonBlockingMetadataStore(enable_write=True, _write_lock=<locked _thread.RLock object owner=132528837793600 count=1 at 0x78682c0748c0>, _store_impl=<orbax.checkpoint._src.metadata.checkpoint._MetadataStoreImpl object at 0x78680c05d0a0>, _single_thread_executor=<concurrent.futures.thread.ThreadPoolExecutor object at 0x78680c05cc80>, _write_futures=[]) I0421 12:08:06.008961 132528837793600 checkpoint.py:459] Closing _NonBlockingMetadataStore(enable_write=True, _write_lock=<locked _thread.RLock object owner=132528837793600 count=1 at 0x78682c0748c0>, _store_impl=<orbax.checkpoint._src.metadata.checkpoint._MetadataStoreImpl object at 0x78680c05d0a0>, _single_thread_executor=<concurrent.futures.thread.ThreadPoolExecutor object at 0x78680c05cc80>, _write_futures=[]) I0421 12:08:06.008991 132528837793600 checkpoint.py:459] Closing _NonBlockingMetadataStore(enable_write=True, _write_lock=<locked _thread.RLock object owner=132528837793600 count=1 at 0x78682c0748c0>, _store_impl=<orbax.checkpoint._src.metadata.checkpoint._MetadataStoreImpl object at 0x78680c05d0a0>, _single_thread_executor=<concurrent.futures.thread.ThreadPoolExecutor object at 0x78680c05cc80>, _write_futures=[]) I0421 12:08:06.009044 132528837793600 train_distill.py:734] Distillation Complete. I0421 12:08:06.273888 132385141139200 grain_pool.py:547] Shutting down multiprocessing system. I0421 12:08:07.711265 132385141139200 grain_pool.py:542] Grain pool is exiting. I0421 12:08:07.711373 132385141139200 grain_pool.py:547] Shutting down multiprocessing system. I0421 12:08:07.711437 132385141139200 grain_pool.py:547] Shutting down multiprocessing system. XPK End: Tue Apr 21 12:08:18 UTC 2026 EXIT_CODE=0