main (172d0f10a) vs fix/gate-fp8-nan-sanitization (3d212094d)| Metric | main 172d0f10a | fix/gate-fp8-nan-sanitization 3d212094d | Diff (fix/gate-fp8-nan-sanitization − main) |
|---|---|---|---|
| Parameters | 1.104 billion | 1.104 billion | — |
| Final loss | 8.7740 | 8.7740 | 0 |
| TFLOP/s | 111.415 | 32.633 | -78.782 |
| Tok/s | 16793.6 | 4918.9 | -11874.774 |
| Avg s/step | 2.761 | 2.772 | +0.011 |
| Memory % | 5.12 | 5.12 | 0 |
| JAX | 0.8.1 | 0.8.1 | — |
Diff = branch value − main value. Green = branch improved. Red = branch regressed.
2026-04-23 01:29:54.982893: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:467] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered WARNING: All log messages before absl::InitializeLog() is called are written to STDERR E0000 00:00:1776907794.995914 2265886 cuda_dnn.cc:8579] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered E0000 00:00:1776907794.999743 2265886 cuda_blas.cc:1407] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered W0000 00:00:1776907795.011254 2265886 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once. W0000 00:00:1776907795.011270 2265886 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once. W0000 00:00:1776907795.011272 2265886 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once. W0000 00:00:1776907795.011274 2265886 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once. Unrecognized keys in `rope_scaling` for 'rope_type'='yarn': {'rope_theta'} `rope_scaling`'s factor field must be a float >= 1, got 40 `rope_scaling`'s beta_fast field must be a float, got 32 `rope_scaling`'s beta_slow field must be a float, got 1 Unrecognized keys in `rope_scaling` for 'rope_type'='yarn': {'rope_theta'} Unrecognized keys in `rope_scaling` for 'rope_type'='yarn': {'rope_theta'} Unrecognized keys in `rope_scaling` for 'rope_type'='yarn': {'rope_theta'} 2026-04-23 01:29:58.686375: 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) I0423 01:29:58.822586 138271370295104 max_utils.py:238] Skipping jax distributed system due to skip_jax_distributed_system=True flag. I0423 01:30:30.938131 138271370295104 max_utils.py:800] System Information: Jax Version: 0.8.1 I0423 01:30:30.938414 138271370295104 max_utils.py:801] System Information: Jaxlib Version: 0.8.1 I0423 01:30:30.938466 138271370295104 max_utils.py:802] System Information: Jax Backend: PJRT C API TFRT TPU v6 lite Built on Nov 12 2025 14:16:36 (1762985796) cl/831091709 I0423 01:30:30.938498 138271370295104 train_utils.py:348] WARNING: 'dataset_path' might be pointing your local file system I0423 01:30:30.938684 138271370295104 train.py:703] [DECOUPLED NO-OP] skipping cloud diagnostics wrapper. I0423 01:30:31.600298 138271370295104 maxtext_utils.py:1604] Num_devices: 8, shape (1, 1, 1, 8, 1, 1, 1, 1, 1, 1, 1, 1, 1) I0423 01:30:31.600677 138271370295104 checkpointing.py:677] Setting up checkpoint logger... I0423 01:30:31.600724 138271370295104 checkpointing.py:233] Creating checkpoint manager with ocdbt=True and zarr3=True I0423 01:30:31.600761 138271370295104 pytree_checkpoint_handler.py:577] save_device_host_concurrent_bytes=None I0423 01:30:31.601217 138271370295104 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 0x7dc1865bb110>, enable_pinned_host_transfer=False, save_concurrent_bytes: 96000000000 (89.4 GiB), restore_concurrent_bytes: 96000000000 (89.4 GiB) I0423 01:30:34.001127 138271370295104 checkpointing.py:265] Enabling policy for fixed interval checkpointing. I0423 01:30:34.001506 138271370295104 checkpoint_manager.py:702] [process=0][thread=MainThread] CheckpointManager init: checkpointers=None, item_names=('items',), item_handlers={'items': <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7dbfb0101940>}, handler_registry=None I0423 01:30:34.001752 138271370295104 composite_checkpoint_handler.py:237] Deferred registration for item: "items". Adding handler `<orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7dbfb0101940>` for item "items" 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`. I0423 01:30:34.001795 138271370295104 composite_checkpoint_handler.py:237] Deferred registration for item: "metrics". Adding handler `<orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7dbfb0101f40>` 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`. I0423 01:30:34.001825 138271370295104 composite_checkpoint_handler.py:505] Initialized registry DefaultCheckpointHandlerRegistry({('items', <class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeSaveArgs'>): <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7dbfb0101940>, ('items', <class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeRestoreArgs'>): <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7dbfb0101940>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonSaveArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7dbfb0101f40>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonRestoreArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7dbfb0101f40>}). I0423 01:30:34.002125 138271370295104 abstract_checkpointer.py:35] orbax-checkpoint version: 0.11.28 I0423 01:30:34.002177 138271370295104 async_checkpointer.py:177] [process=0][thread=MainThread] Using barrier_sync_fn: <function get_barrier_sync_fn.<locals>.<lambda> at 0x7dbfb0110540> timeout: 600 secs and primary_host=0 for async checkpoint writes I0423 01:30:34.135480 138271370295104 checkpoint_manager.py:1788] Found 0 checkpoint steps in gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints I0423 01:30:34.135710 138271370295104 checkpoint_manager.py:921] [process=0][thread=MainThread] CheckpointManager created, primary_host=0, CheckpointManagerOptions=CheckpointManagerOptions(save_interval_steps=1, 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=FixedIntervalPolicy(interval=10), preservation_policy=LatestN(n=None), prevent_write_metrics=False, enable_should_save_is_saving_in_progress_check=True, enable_per_process_directory_creation=False), root_directory=gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints: <orbax.checkpoint.checkpoint_manager.CheckpointManager object at 0x7dc19ff4bf80> I0423 01:30:34.135992 138271370295104 checkpointing.py:301] Checkpoint manager created! I0423 01:30:34.214149 138271370295104 dataset_info.py:707] Load dataset info from tests/assets/local_datasets/c4_en_dataset_minimal/c4/en/3.1.0 I0423 01:30:34.217183 138271370295104 reader.py:262] Creating a tf.data.Dataset reading 8 files located in folders: tests/assets/local_datasets/c4_en_dataset_minimal/c4/en/3.1.0. I0423 01:30:34.276832 138271370295104 logging_logger.py:49] Constructing tf.data.Dataset __local_c4_builder for split train, from tests/assets/local_datasets/c4_en_dataset_minimal/c4/en/3.1.0 I0423 01:30:34.299800 138271370295104 tokenizer.py:245] Tokenizer path: src/maxtext/assets/tokenizers/tokenizer.llama2 I0423 01:30:34.299866 138271370295104 tokenizer.py:187] Loading sentencepiece tokenizer: src/maxtext/assets/tokenizers/tokenizer.llama2 I0423 01:30:35.160907 138271370295104 nnx_wrappers.py:437] Unknown Logical: bfloat16[8,2048,2048]....................................... ('activation_batch', 'activation_norm_length', 'activation_embed'). I0423 01:30:35.161019 138271370295104 nnx_wrappers.py:437] Unknown Physical: bfloat16[8,2048,2048]....................................... ('fsdp', None, None). I0423 01:30:35.464904 138271370295104 attentions.py:1088] attentions/inputs_q Logical: bfloat16[8,2048,2048]....................................... ('activation_batch', 'activation_attn_length', 'activation_attn_embed'). I0423 01:30:35.464999 138271370295104 attentions.py:1088] attentions/inputs_q Physical: bfloat16[8,2048,2048]....................................... ('fsdp', None, None). I0423 01:30:35.477925 138271370295104 attentions.py:1089] attentions/inputs_kv Logical: bfloat16[8,2048,2048]....................................... ('activation_batch', 'activation_attn_length', 'activation_attn_embed'). I0423 01:30:35.477982 138271370295104 attentions.py:1089] attentions/inputs_kv Physical: bfloat16[8,2048,2048]....................................... ('fsdp', None, None). I0423 01:30:35.544891 138271370295104 attentions.py:1154] attentions/query Logical: bfloat16[8,2048,16,128]..................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim'). I0423 01:30:35.544991 138271370295104 attentions.py:1154] attentions/query Physical: bfloat16[8,2048,16,128]..................................... ('fsdp', None, None, None). I0423 01:30:35.558003 138271370295104 attentions.py:1155] attentions/key Logical: bfloat16[8,2048,16,128]..................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim'). I0423 01:30:35.558070 138271370295104 attentions.py:1155] attentions/key Physical: bfloat16[8,2048,16,128]..................................... ('fsdp', None, None, None). I0423 01:30:35.570972 138271370295104 attentions.py:1156] attentions/value Logical: bfloat16[8,2048,16,128]..................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim'). I0423 01:30:35.571025 138271370295104 attentions.py:1156] attentions/value Physical: bfloat16[8,2048,16,128]..................................... ('fsdp', None, None, None). I0423 01:30:35.617683 138271370295104 attentions.py:1198] attentions/out Logical: bfloat16[8,2048,16,128]..................................... ('activation_batch', 'activation_attn_length', 'activation_heads', 'activation_kv'). I0423 01:30:35.617824 138271370295104 attentions.py:1198] attentions/out Physical: bfloat16[8,2048,16,128]..................................... ('fsdp', None, None, None). I0423 01:30:35.690109 138271370295104 linears.py:525] linears/x Logical: bfloat16[8,2048,7168]....................................... ('activation_batch', 'activation_length', 'activation_mlp'). I0423 01:30:35.690190 138271370295104 linears.py:525] linears/x Physical: bfloat16[8,2048,7168]....................................... ('fsdp', None, None). I0423 01:30:36.496059 138271370295104 checkpointing.py:577] checkpoint manager exists so trying to load this run's existing checkpoint I0423 01:30:36.496165 138271370295104 checkpointing.py:665] No existing checkpoints found, not restoring checkpoint. [DECOUPLED NO-OP] gcs_storage: using stubs. [DECOUPLED NO-OP] mldiagnostics: using stub. [DECOUPLED NO-OP] mldiagnostics: using stub. [DECOUPLED NO-OP] mldiagnostics: using stub. [DECOUPLED NO-OP] workload_monitor: using stub. [DECOUPLED NO-OP] vertex_tensorboard: using stub. fsdp: 8 I0423 01:30:37.781165 138271370295104 maxtext_utils.py:1707] params/params/decoder/decoder_norm/scale Shape: float32[2048] Logical: PartitionSpec('norm',) Physical: (None,) I0423 01:30:37.781289 138271370295104 maxtext_utils.py:1707] params/params/decoder/layers/mlp/wi_0/kernel Shape: float32[2048,16,7168] Logical: PartitionSpec('embed', 'layers', 'mlp') Physical: ('fsdp', None, None) I0423 01:30:37.781333 138271370295104 maxtext_utils.py:1707] params/params/decoder/layers/mlp/wi_1/kernel Shape: float32[2048,16,7168] Logical: PartitionSpec('embed', 'layers', 'mlp') Physical: ('fsdp', None, None) I0423 01:30:37.781379 138271370295104 maxtext_utils.py:1707] params/params/decoder/layers/mlp/wo/kernel Shape: float32[7168,16,2048] Logical: PartitionSpec('mlp', 'layers', 'embed') Physical: (None, None, 'fsdp') I0423 01:30:37.781422 138271370295104 maxtext_utils.py:1707] params/params/decoder/layers/post_self_attention_layer_norm/scale Shape: float32[2048,16] Logical: PartitionSpec('norm', 'layers') Physical: (None, None) I0423 01:30:37.781451 138271370295104 maxtext_utils.py:1707] params/params/decoder/layers/pre_self_attention_layer_norm/scale Shape: float32[2048,16] Logical: PartitionSpec('norm', 'layers') Physical: (None, None) I0423 01:30:37.781491 138271370295104 maxtext_utils.py:1707] params/params/decoder/layers/self_attention/key/kernel Shape: float32[2048,16,16,128] Logical: PartitionSpec('embed', 'layers', 'kv_heads', 'kv_head_dim') Physical: ('fsdp', None, None, None) I0423 01:30:37.781533 138271370295104 maxtext_utils.py:1707] params/params/decoder/layers/self_attention/out/kernel Shape: float32[16,16,128,2048] Logical: PartitionSpec('heads', 'layers', 'kv', 'embed') Physical: (None, None, None, 'fsdp') I0423 01:30:37.781563 138271370295104 maxtext_utils.py:1707] params/params/decoder/layers/self_attention/query/kernel Shape: float32[2048,16,16,128] Logical: PartitionSpec('embed', 'layers', 'q_heads', 'kv') Physical: ('fsdp', None, None, None) I0423 01:30:37.781591 138271370295104 maxtext_utils.py:1707] params/params/decoder/layers/self_attention/value/kernel Shape: float32[2048,16,16,128] Logical: PartitionSpec('embed', 'layers', 'kv_heads', 'kv_head_dim') Physical: ('fsdp', None, None, None) I0423 01:30:37.781628 138271370295104 maxtext_utils.py:1707] params/params/decoder/logits_dense/kernel Shape: float32[2048,32000] Logical: PartitionSpec('embed_vocab', 'vocab') Physical: ('fsdp', None) I0423 01:30:37.781665 138271370295104 maxtext_utils.py:1707] params/params/token_embedder/embedding Shape: float32[32000,2048] Logical: PartitionSpec('vocab', 'embed_vocab') Physical: (None, 'fsdp') I0423 01:30:39.555738 138271370295104 train.py:155] train/xent Logical: float32[8,2048]............................................. ('activation_embed_and_logits_batch', 'activation_length'). I0423 01:30:39.555830 138271370295104 train.py:155] train/xent Physical: float32[8,2048]............................................. ('fsdp', None). I0423 01:30:39.568155 138271370295104 train.py:162] train/z_loss Logical: float32[8,2048]............................................. ('activation_embed_and_logits_batch', 'activation_length'). I0423 01:30:39.568209 138271370295104 train.py:162] train/z_loss Physical: float32[8,2048]............................................. ('fsdp', None). I0423 01:30:41.605844 138271370295104 max_utils.py:791] Total memory size: 3.7 GB, Output size: 1.5 GB, Temp size: 2.2 GB, Argument size: 1.5 GB, Host temp size: 0.0 GB. I0423 01:30:41.606693 138271370295104 max_utils.py:194] tensorboardX not available; using no-op SummaryWriter. I0423 01:30:41.606859 138271370295104 metric_logger.py:301] number parameters: 1.104 billion I0423 01:30:48.291800 138271370295104 checkpointing.py:772] Waiting for step 0 to finish before checkpoint... I0423 01:30:48.420843 138271370295104 checkpointing.py:776] Waited 0.12910151481628418 seconds for step 0 to finish before starting checkpointing. I0423 01:30:48.421967 138271370295104 checkpoint_manager.py:1983] [process=0][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning. I0423 01:30:48.422266 138271370295104 checkpoint_manager.py:1501] [process=0] Saving checkpoint at step 0 I0423 01:30:48.423128 138271370295104 async_checkpointer.py:452] [process=0] Started async saving checkpoint to gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints/0. I0423 01:30:48.509599 138271370295104 signaling_client.py:373] Using ThreadSafeKeyValueSignalingClient I0423 01:30:48.602914 138271370295104 jax_array_handlers.py:347] Scheduling D2H of 39 prioritized jax.Array. I0423 01:30:48.603098 138271370295104 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 I0423 01:30:48.610949 138143728666176 atomicity.py:137] Creating tmp directory gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints/0 I0423 01:30:49.341102 138143718180416 atomicity.py:137] Creating tmp directory gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints/0/items I0423 01:30:49.480873 138143686723136 checkpoint.py:188] Wrote Metadata={'item_handlers': None, 'metrics': {}, 'performance_metrics': {}, 'init_timestamp_nsecs': 1776907849247808663, 'commit_timestamp_nsecs': None, 'custom_metadata': {}}, json={"item_handlers": null, "metrics": {}, "performance_metrics": {}, "init_timestamp_nsecs": 1776907849247808663, "commit_timestamp_nsecs": null, "custom_metadata": {}} to gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints/0/_CHECKPOINT_METADATA I0423 01:30:49.803727 2268521 google_auth_provider.cc:149] Using credentials at ~/.config/gcloud/application_default_credentials.json I0423 01:30:49.803808 2268521 google_auth_provider.cc:156] Using OAuth2 AuthProvider I0423 01:30:51.083353 138271370295104 base_pytree_checkpoint_handler.py:153] [process=0][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 2.481068s I0423 01:30:51.084091 138271370295104 base_pytree_checkpoint_handler.py:128] [process=0] /jax/checkpoint/write/blocking_gbytes_per_sec: 4.796 GiB/s (total gbytes: 12.3 GiB) (time elapsed: 2 seconds) (per-host) I0423 01:30:51.084164 138271370295104 base_pytree_checkpoint_handler.py:732] [process=0][thread=MainThread] Initiated Pytree async_save. Time taken: 2.573348s (batch_requests_ready=0.087513s, total_serialization_initiated=2.485202s, others=0.000633s) I0423 01:30:51.084245 138271370295104 composite_checkpoint_handler.py:715] [process=0][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 2.574034s (all_items=0.000043s, per_item={'items': '0.00004339'}, temp_paths=2.573991) I0423 01:30:51.089827 138144078890560 async_checkpointer.py:79] [process=0][thread=async_save] Background save thread started. I0423 01:30:51.089957 138271370295104 async_checkpointer.py:561] Finished blocking save. Time taken: 2.667637s. Continuing background save to gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints/0. I0423 01:30:51.090243 138271370295104 checkpoint_manager.py:1549] [process=0][thread=MainThread][step=0] Starting CheckpointManager Save Finalize thread=save_finalize I0423 01:30:51.090577 138142887708224 async_checkpointer.py:265] [process=0][thread=save_finalize] Waiting for background save thread=async_save. I0423 01:30:51.090705 138271370295104 standard_logger.py:34] {'step': 0, 'event_type': 'save', 'directory': 'gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776907848.421943, 'wait_for_prev_duration_secs': 5.626678466796875e-05, 'time_between_consecutive_saves_sec': None, 'checkpointer_blocking_start_time': 1776907848.4222898, 'checkpointer_blocking_duration_secs': 2.6678452491760254, 'get_old_steps_start_time': 1776907851.090157, 'get_old_steps_duration_secs': 4.267692565917969e-05, 'checkpoint_manager_blocking_start_time': 1776907848.4217026, 'checkpoint_manager_blocking_duration_secs': 2.6689741611480713} I0423 01:30:51.090834 138271370295104 checkpointing.py:408] Started an asynchronous checkpoint save for step 0 I0423 01:30:51.090900 138271370295104 max_utils.py:750] Memstats: After params initialized: I0423 01:30:51.090955 138271370295104 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_0(process=0,(0,0,0,0)) I0423 01:30:51.090988 138271370295104 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_1(process=0,(1,0,0,0)) I0423 01:30:51.091011 138271370295104 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_2(process=0,(0,1,0,0)) I0423 01:30:51.091028 138271370295104 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_3(process=0,(1,1,0,0)) I0423 01:30:51.091064 138271370295104 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_4(process=0,(0,2,0,0)) I0423 01:30:51.091083 138271370295104 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_5(process=0,(1,2,0,0)) I0423 01:30:51.091099 138271370295104 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_6(process=0,(0,3,0,0)) I0423 01:30:51.091117 138271370295104 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_7(process=0,(1,3,0,0)) I0423 01:30:51.466205 138271370295104 metric_logger.py:196] completed step: 0, seconds: 6.684, TFLOP/s/device: 2.033, Tokens/s/device: 306.391, total_weights: 13328, loss: 10.840, lm_loss: 10.840, perplexity: 51025.957 I0423 01:30:51.468384 138271370295104 metric_logger.py:281] To see full metrics 'tensorboard --logdir=gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/tensorboard/' I0423 01:30:51.584585 138271370295104 metric_logger.py:196] completed step: 1, seconds: 3.172, TFLOP/s/device: 4.284, Tokens/s/device: 645.729, total_weights: 12332, loss: 10.869, lm_loss: 10.869, perplexity: 52515.969 I0423 01:30:51.705591 138271370295104 metric_logger.py:196] completed step: 2, seconds: 0.012, TFLOP/s/device: 1098.216, Tokens/s/device: 165535.079, total_weights: 15161, loss: 9.874, lm_loss: 9.874, perplexity: 19421.770 I0423 01:30:51.823293 138143697208896 array_metadata_store.py:203] [process=0][thread=array_type_handler] Wrote 39 array_metadata.ArrayMetadata to gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints/0/items/array_metadatas/process_0 I0423 01:30:51.826754 138271370295104 metric_logger.py:196] completed step: 3, seconds: 0.119, TFLOP/s/device: 114.511, Tokens/s/device: 17260.415, total_weights: 13327, loss: 9.406, lm_loss: 9.406, perplexity: 12166.079 I0423 01:31:12.880112 138271370295104 metric_logger.py:196] completed step: 4, seconds: 0.121, TFLOP/s/device: 112.485, Tokens/s/device: 16954.906, total_weights: 11939, loss: 9.149, lm_loss: 9.149, perplexity: 9403.998 I0423 01:31:12.892339 138271370295104 metric_logger.py:196] completed step: 5, seconds: 0.121, TFLOP/s/device: 112.364, Tokens/s/device: 16936.678, total_weights: 15502, loss: 9.058, lm_loss: 9.058, perplexity: 8587.564 I0423 01:31:13.012173 138271370295104 metric_logger.py:196] completed step: 6, seconds: 21.055, TFLOP/s/device: 0.645, Tokens/s/device: 97.269, total_weights: 13864, loss: 8.892, lm_loss: 8.892, perplexity: 7274.692 I0423 01:31:13.133987 138271370295104 metric_logger.py:196] completed step: 7, seconds: 0.008, TFLOP/s/device: 1729.082, Tokens/s/device: 260626.114, total_weights: 12988, loss: 8.751, lm_loss: 8.751, perplexity: 6314.503 I0423 01:31:13.255376 138271370295104 metric_logger.py:196] completed step: 8, seconds: 0.121, TFLOP/s/device: 112.603, Tokens/s/device: 16972.751, total_weights: 13820, loss: 8.791, lm_loss: 8.791, perplexity: 6573.798 I0423 01:31:13.376469 138271370295104 checkpointing.py:772] Waiting for step 9 to finish before checkpoint... I0423 01:31:13.377474 138271370295104 checkpointing.py:776] Waited 0.0010411739349365234 seconds for step 9 to finish before starting checkpointing. I0423 01:31:13.377818 138271370295104 checkpoint_manager.py:1994] [process=0][thread=MainThread][step=0][wait_until_finished] Waiting for Save Finalize thread (save_finalize) to complete. I0423 01:31:19.615498 138143212766784 base_pytree_checkpoint_handler.py:1217] [process=0][thread=write_metadata_after_commits] Commit + Array metadata written. Time taken: 28.526173s (commit=27.998119s, array_metadata_write=0.528054s) I0423 01:31:19.616672 138144078890560 base_pytree_checkpoint_handler.py:128] [process=0] /jax/checkpoint/write/gbytes_per_sec: 406.252 MiB/s (total gbytes: 12.3 GiB) (time elapsed: 31 seconds) (per-host) I0423 01:31:19.616798 138144078890560 async_checkpointer.py:90] [process=0][thread=async_save] 3 Handler Commit operations completed. Time taken: 28.526787s. I0423 01:31:19.856575 138144078890560 checkpoint.py:228] Read Metadata={'item_handlers': None, 'metrics': {}, 'performance_metrics': {}, 'init_timestamp_nsecs': 1776907849247808663, 'commit_timestamp_nsecs': None, 'custom_metadata': {}} from gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints/0/_CHECKPOINT_METADATA I0423 01:31:20.104472 138144078890560 array_metadata_store.py:367] [process=0][thread=async_save] Skipped cross-host ArrayMetadata validation because only one process is found: process_index=0. I0423 01:31:20.300245 138143686723136 checkpoint.py:247] Updated Metadata={'item_handlers': {'items': 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler'}, 'metrics': {}, 'performance_metrics': {}, 'init_timestamp_nsecs': 1776907849247808663, 'commit_timestamp_nsecs': None, 'custom_metadata': {}} to gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints/0/_CHECKPOINT_METADATA I0423 01:31:20.457788 138144078890560 ocdbt_utils.py:56] Param validation support for Zarr3 will be added later (b/362328389). I0423 01:31:20.457984 138144078890560 base_pytree_checkpoint_handler.py:1342] [process=0][thread=async_save] Pytree save finalize (merge_ocdbt + ArrayMetadata validation) completed. Time taken: 0.560290s. use_zarr3=True, enable_post_merge_validation=True, directory=gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints/0/items I0423 01:31:20.458820 138144078890560 atomicity.py:608] Finalizing gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints/0/items I0423 01:31:20.698175 138144078890560 atomicity.py:608] Finalizing gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints/0 I0423 01:31:21.390166 138144078890560 atomicity.py:794] [process=0][thread=async_save] Finished saving checkpoint (finalized tmp dir) to `gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints/0`. I0423 01:31:21.390911 138144078890560 async_checkpointer.py:420] Finished async_save (blocking + background). Time taken: 32.968598s. directory=gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints/0 I0423 01:31:21.390999 138144078890560 async_checkpointer.py:144] [process=0][thread=async_save] Background save thread done. Time taken: 30.300992s. I0423 01:31:21.391262 138142887708224 async_checkpointer.py:273] [process=0][thread=save_finalize] Done with waiting for background save thread=async_save. I0423 01:31:21.391394 138142887708224 async_checkpointer.py:283] [process=0][thread=save_finalize] No errors found in background save thread=async_save. I0423 01:31:21.391464 138142887708224 checkpoint_manager.py:2103] [process=0][thread=save_finalize][step=0] CheckpointManager Save Finalize is syncing with other hosts... I0423 01:31:21.391518 138142887708224 checkpoint_manager.py:2112] [process=0][thread=save_finalize][step=0] CheckpointManager Save Finalize is done on all hosts. I0423 01:31:21.391634 138271370295104 checkpoint_manager.py:2006] [process=0][thread=MainThread][step=0][wait_until_finished] Done waiting for Save Finalize thread (save_finalize) running at step=0. W0423 01:31:21.391713 138271370295104 checkpoint_manager.py:1441] Waiting for previous save to complete took 8.013936 seconds. If this number is high, consider checkpointing less frequently. I0423 01:31:21.392844 138271370295104 checkpoint_manager.py:1501] [process=0] Saving checkpoint at step 9 I0423 01:31:21.393131 138271370295104 async_checkpointer.py:452] [process=0] Started async saving checkpoint to gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints/9. I0423 01:31:21.572162 138271370295104 jax_array_handlers.py:347] Scheduling D2H of 39 prioritized jax.Array. I0423 01:31:21.572268 138271370295104 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 I0423 01:31:21.579972 138144078890560 atomicity.py:137] Creating tmp directory gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints/9 I0423 01:31:22.271545 138143697208896 atomicity.py:137] Creating tmp directory gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints/9/items I0423 01:31:27.203591 138271370295104 base_pytree_checkpoint_handler.py:153] [process=0][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 5.631877s I0423 01:31:27.208406 138271370295104 base_pytree_checkpoint_handler.py:128] [process=0] /jax/checkpoint/write/blocking_gbytes_per_sec: 2.156 GiB/s (total gbytes: 12.3 GiB) (time elapsed: 5 seconds) (per-host) I0423 01:31:27.208710 138271370295104 base_pytree_checkpoint_handler.py:732] [process=0][thread=MainThread] Initiated Pytree async_save. Time taken: 5.723809s (batch_requests_ready=0.083699s, total_serialization_initiated=5.635201s, others=0.004909s) I0423 01:31:27.208825 138271370295104 composite_checkpoint_handler.py:715] [process=0][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 5.724429s (all_items=0.000035s, per_item={'items': '0.00003505'}, temp_paths=5.724394) I0423 01:31:27.210312 138143357470272 async_checkpointer.py:79] [process=0][thread=async_save] Background save thread started. I0423 01:31:27.210400 138271370295104 async_checkpointer.py:561] Finished blocking save. Time taken: 5.817509s. Continuing background save to gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints/9. I0423 01:31:27.210613 138271370295104 checkpoint_manager.py:1549] [process=0][thread=MainThread][step=9] Starting CheckpointManager Save Finalize thread=save_finalize I0423 01:31:27.210833 138142887708224 async_checkpointer.py:265] [process=0][thread=save_finalize] Waiting for background save thread=async_save. I0423 01:31:27.210971 138271370295104 standard_logger.py:34] {'step': 9, 'event_type': 'save', 'directory': 'gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776907873.3777525, 'wait_for_prev_duration_secs': 8.013935565948486, 'time_between_consecutive_saves_sec': None, 'checkpointer_blocking_start_time': 1776907881.3928702, 'checkpointer_blocking_duration_secs': 5.817642688751221, 'get_old_steps_start_time': 1776907887.2105312, 'get_old_steps_duration_secs': 4.100799560546875e-05, 'checkpoint_manager_blocking_start_time': 1776907873.377695, 'checkpoint_manager_blocking_duration_secs': 13.833244323730469} I0423 01:31:27.211133 138271370295104 checkpointing.py:408] Started an asynchronous checkpoint save for step 9 I0423 01:31:27.211172 138271370295104 checkpoint_manager.py:1994] [process=0][thread=MainThread][step=9][wait_until_finished] Waiting for Save Finalize thread (save_finalize) to complete. I0423 01:31:27.964764 138254722532928 array_metadata_store.py:203] [process=0][thread=array_type_handler] Wrote 39 array_metadata.ArrayMetadata to gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints/9/items/array_metadatas/process_0 I0423 01:32:07.045681 138143697208896 base_pytree_checkpoint_handler.py:1217] [process=0][thread=write_metadata_after_commits] Commit + Array metadata written. Time taken: 39.836457s (commit=39.336436s, array_metadata_write=0.500022s) I0423 01:32:07.046759 138143357470272 base_pytree_checkpoint_handler.py:128] [process=0] /jax/checkpoint/write/gbytes_per_sec: 277.355 MiB/s (total gbytes: 12.3 GiB) (time elapsed: 45 seconds) (per-host) I0423 01:32:07.046816 138143357470272 async_checkpointer.py:90] [process=0][thread=async_save] 3 Handler Commit operations completed. Time taken: 39.836371s. I0423 01:32:07.524817 138143357470272 array_metadata_store.py:367] [process=0][thread=async_save] Skipped cross-host ArrayMetadata validation because only one process is found: process_index=0. I0423 01:32:07.903451 138143357470272 ocdbt_utils.py:56] Param validation support for Zarr3 will be added later (b/362328389). I0423 01:32:07.903640 138143357470272 base_pytree_checkpoint_handler.py:1342] [process=0][thread=async_save] Pytree save finalize (merge_ocdbt + ArrayMetadata validation) completed. Time taken: 0.547645s. use_zarr3=True, enable_post_merge_validation=True, directory=gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints/9/items I0423 01:32:07.904376 138143357470272 atomicity.py:608] Finalizing gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints/9/items I0423 01:32:08.156528 138143357470272 atomicity.py:608] Finalizing gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints/9 I0423 01:32:08.895942 138143357470272 atomicity.py:794] [process=0][thread=async_save] Finished saving checkpoint (finalized tmp dir) to `gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints/9`. I0423 01:32:08.896702 138143357470272 async_checkpointer.py:420] Finished async_save (blocking + background). Time taken: 47.503817s. directory=gs://wanglance-maxtext/linen_ckpt_main_20260423_012248/linen_main_20260423_012248_04_int8/checkpoints/9 I0423 01:32:08.896780 138143357470272 async_checkpointer.py:144] [process=0][thread=async_save] Background save thread done. Time taken: 41.686336s. I0423 01:32:08.896935 138142887708224 async_checkpointer.py:273] [process=0][thread=save_finalize] Done with waiting for background save thread=async_save. I0423 01:32:08.897064 138142887708224 async_checkpointer.py:283] [process=0][thread=save_finalize] No errors found in background save thread=async_save. I0423 01:32:08.897131 138142887708224 checkpoint_manager.py:2103] [process=0][thread=save_finalize][step=9] CheckpointManager Save Finalize is syncing with other hosts... I0423 01:32:08.897172 138142887708224 checkpoint_manager.py:2112] [process=0][thread=save_finalize][step=9] CheckpointManager Save Finalize is done on all hosts. I0423 01:32:08.897340 138271370295104 checkpoint_manager.py:2006] [process=0][thread=MainThread][step=9][wait_until_finished] Done waiting for Save Finalize thread (save_finalize) running at step=9. I0423 01:32:08.897490 138271370295104 checkpoint_manager.py:1983] [process=0][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning. I0423 01:32:08.898451 138271370295104 metric_logger.py:196] completed step: 9, seconds: 0.122, TFLOP/s/device: 111.415, Tokens/s/device: 16793.630, total_weights: 12300, loss: 8.774, lm_loss: 8.774, perplexity: 6465.313 Per train step: Total TFLOPs: 13.59 split as 93.93% learnable weight flops and 6.07% attention flops
3d212094d
· fix_gate_fp8_nan_sanitization_20260423_024158
· full log
2026-04-23 02:51:34.227102: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:467] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered WARNING: All log messages before absl::InitializeLog() is called are written to STDERR E0000 00:00:1776912694.240977 2343980 cuda_dnn.cc:8579] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered E0000 00:00:1776912694.244960 2343980 cuda_blas.cc:1407] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered W0000 00:00:1776912694.257521 2343980 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once. W0000 00:00:1776912694.257537 2343980 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once. W0000 00:00:1776912694.257539 2343980 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once. W0000 00:00:1776912694.257541 2343980 computation_placer.cc:177] computation placer already registered. Please check linkage and avoid linking the same target more than once. Unrecognized keys in `rope_scaling` for 'rope_type'='yarn': {'rope_theta'} `rope_scaling`'s factor field must be a float >= 1, got 40 `rope_scaling`'s beta_fast field must be a float, got 32 `rope_scaling`'s beta_slow field must be a float, got 1 Unrecognized keys in `rope_scaling` for 'rope_type'='yarn': {'rope_theta'} Unrecognized keys in `rope_scaling` for 'rope_type'='yarn': {'rope_theta'} Unrecognized keys in `rope_scaling` for 'rope_type'='yarn': {'rope_theta'} 2026-04-23 02:51:37.927711: 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) I0423 02:51:38.061768 137821813905216 max_utils.py:238] Skipping jax distributed system due to skip_jax_distributed_system=True flag. ~/maxtext/maxtext_venv/lib/python3.12/site-packages/jax/_src/xla_bridge.py:204: UserWarning: TPU backend initialization is taking more than 60.0 seconds. Did you run your code on all TPU hosts? See https://docs.jax.dev/en/latest/multi_process.html for more information. warnings.warn( I0423 02:52:42.233637 137821813905216 max_utils.py:800] System Information: Jax Version: 0.8.1 I0423 02:52:42.233875 137821813905216 max_utils.py:801] System Information: Jaxlib Version: 0.8.1 I0423 02:52:42.233923 137821813905216 max_utils.py:802] System Information: Jax Backend: PJRT C API TFRT TPU v6 lite Built on Nov 12 2025 14:16:36 (1762985796) cl/831091709 I0423 02:52:42.233951 137821813905216 train_utils.py:348] WARNING: 'dataset_path' might be pointing your local file system I0423 02:52:42.234142 137821813905216 train.py:672] [DECOUPLED NO-OP] skipping cloud diagnostics wrapper. I0423 02:52:42.897392 137821813905216 maxtext_utils.py:1604] Num_devices: 8, shape (1, 1, 1, 8, 1, 1, 1, 1, 1, 1, 1, 1, 1) I0423 02:52:42.897765 137821813905216 checkpointing.py:677] Setting up checkpoint logger... I0423 02:52:42.897815 137821813905216 checkpointing.py:233] Creating checkpoint manager with ocdbt=True and zarr3=True I0423 02:52:42.897851 137821813905216 pytree_checkpoint_handler.py:577] save_device_host_concurrent_bytes=None I0423 02:52:42.898293 137821813905216 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 0x7d58dabaade0>, enable_pinned_host_transfer=False, save_concurrent_bytes: 96000000000 (89.4 GiB), restore_concurrent_bytes: 96000000000 (89.4 GiB) I0423 02:52:45.307390 137821813905216 checkpointing.py:265] Enabling policy for fixed interval checkpointing. I0423 02:52:45.307753 137821813905216 checkpoint_manager.py:702] [process=0][thread=MainThread] CheckpointManager init: checkpointers=None, item_names=('items',), item_handlers={'items': <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7d570c1f6630>}, handler_registry=None I0423 02:52:45.307983 137821813905216 composite_checkpoint_handler.py:237] Deferred registration for item: "items". Adding handler `<orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7d570c1f6630>` for item "items" 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`. I0423 02:52:45.308023 137821813905216 composite_checkpoint_handler.py:237] Deferred registration for item: "metrics". Adding handler `<orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7d570c1f5880>` 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`. I0423 02:52:45.308065 137821813905216 composite_checkpoint_handler.py:505] Initialized registry DefaultCheckpointHandlerRegistry({('items', <class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeSaveArgs'>): <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7d570c1f6630>, ('items', <class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeRestoreArgs'>): <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7d570c1f6630>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonSaveArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7d570c1f5880>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonRestoreArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7d570c1f5880>}). I0423 02:52:45.308344 137821813905216 abstract_checkpointer.py:35] orbax-checkpoint version: 0.11.28 I0423 02:52:45.308394 137821813905216 async_checkpointer.py:177] [process=0][thread=MainThread] Using barrier_sync_fn: <function get_barrier_sync_fn.<locals>.<lambda> at 0x7d570c224540> timeout: 600 secs and primary_host=0 for async checkpoint writes I0423 02:52:45.441937 137821813905216 checkpoint_manager.py:1788] Found 0 checkpoint steps in gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints I0423 02:52:45.442174 137821813905216 checkpoint_manager.py:921] [process=0][thread=MainThread] CheckpointManager created, primary_host=0, CheckpointManagerOptions=CheckpointManagerOptions(save_interval_steps=1, 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=FixedIntervalPolicy(interval=10), preservation_policy=LatestN(n=None), prevent_write_metrics=False, enable_should_save_is_saving_in_progress_check=True, enable_per_process_directory_creation=False), root_directory=gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints: <orbax.checkpoint.checkpoint_manager.CheckpointManager object at 0x7d570c1f7140> I0423 02:52:45.442445 137821813905216 checkpointing.py:301] Checkpoint manager created! I0423 02:52:45.519177 137821813905216 dataset_info.py:707] Load dataset info from tests/assets/local_datasets/c4_en_dataset_minimal/c4/en/3.1.0 I0423 02:52:45.522005 137821813905216 reader.py:262] Creating a tf.data.Dataset reading 8 files located in folders: tests/assets/local_datasets/c4_en_dataset_minimal/c4/en/3.1.0. I0423 02:52:45.581414 137821813905216 logging_logger.py:49] Constructing tf.data.Dataset __local_c4_builder for split train, from tests/assets/local_datasets/c4_en_dataset_minimal/c4/en/3.1.0 I0423 02:52:45.603947 137821813905216 tokenizer.py:245] Tokenizer path: src/maxtext/assets/tokenizers/tokenizer.llama2 I0423 02:52:45.604007 137821813905216 tokenizer.py:187] Loading sentencepiece tokenizer: src/maxtext/assets/tokenizers/tokenizer.llama2 I0423 02:52:46.445144 137821813905216 nnx_wrappers.py:437] Unknown Logical: bfloat16[8,2048,2048]....................................... ('activation_batch', 'activation_norm_length', 'activation_embed'). I0423 02:52:46.445236 137821813905216 nnx_wrappers.py:437] Unknown Physical: bfloat16[8,2048,2048]....................................... ('fsdp', None, None). I0423 02:52:46.745581 137821813905216 attentions.py:1088] attentions/inputs_q Logical: bfloat16[8,2048,2048]....................................... ('activation_batch', 'activation_attn_length', 'activation_attn_embed'). I0423 02:52:46.745667 137821813905216 attentions.py:1088] attentions/inputs_q Physical: bfloat16[8,2048,2048]....................................... ('fsdp', None, None). I0423 02:52:46.758384 137821813905216 attentions.py:1089] attentions/inputs_kv Logical: bfloat16[8,2048,2048]....................................... ('activation_batch', 'activation_attn_length', 'activation_attn_embed'). I0423 02:52:46.758434 137821813905216 attentions.py:1089] attentions/inputs_kv Physical: bfloat16[8,2048,2048]....................................... ('fsdp', None, None). I0423 02:52:46.823383 137821813905216 attentions.py:1154] attentions/query Logical: bfloat16[8,2048,16,128]..................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim'). I0423 02:52:46.823460 137821813905216 attentions.py:1154] attentions/query Physical: bfloat16[8,2048,16,128]..................................... ('fsdp', None, None, None). I0423 02:52:46.836454 137821813905216 attentions.py:1155] attentions/key Logical: bfloat16[8,2048,16,128]..................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim'). I0423 02:52:46.836506 137821813905216 attentions.py:1155] attentions/key Physical: bfloat16[8,2048,16,128]..................................... ('fsdp', None, None, None). I0423 02:52:46.849277 137821813905216 attentions.py:1156] attentions/value Logical: bfloat16[8,2048,16,128]..................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim'). I0423 02:52:46.849326 137821813905216 attentions.py:1156] attentions/value Physical: bfloat16[8,2048,16,128]..................................... ('fsdp', None, None, None). I0423 02:52:46.869176 137821813905216 attentions.py:1198] attentions/out Logical: bfloat16[8,2048,16,128]..................................... ('activation_batch', 'activation_attn_length', 'activation_heads', 'activation_kv'). I0423 02:52:46.869233 137821813905216 attentions.py:1198] attentions/out Physical: bfloat16[8,2048,16,128]..................................... ('fsdp', None, None, None). I0423 02:52:46.935840 137821813905216 linears.py:525] linears/x Logical: bfloat16[8,2048,7168]....................................... ('activation_batch', 'activation_length', 'activation_mlp'). I0423 02:52:46.935923 137821813905216 linears.py:525] linears/x Physical: bfloat16[8,2048,7168]....................................... ('fsdp', None, None). I0423 02:52:47.699073 137821813905216 checkpointing.py:577] checkpoint manager exists so trying to load this run's existing checkpoint I0423 02:52:47.699169 137821813905216 checkpointing.py:665] No existing checkpoints found, not restoring checkpoint. [DECOUPLED NO-OP] gcs_storage: using stubs. [DECOUPLED NO-OP] mldiagnostics: using stub. [DECOUPLED NO-OP] mldiagnostics: using stub. [DECOUPLED NO-OP] mldiagnostics: using stub. [DECOUPLED NO-OP] workload_monitor: using stub. [DECOUPLED NO-OP] vertex_tensorboard: using stub. fsdp: 8 I0423 02:52:48.921390 137821813905216 maxtext_utils.py:1707] params/params/decoder/decoder_norm/scale Shape: float32[2048] Logical: PartitionSpec('norm',) Physical: (None,) I0423 02:52:48.921501 137821813905216 maxtext_utils.py:1707] params/params/decoder/layers/mlp/wi_0/kernel Shape: float32[2048,16,7168] Logical: PartitionSpec('embed', 'layers', 'mlp') Physical: ('fsdp', None, None) I0423 02:52:48.921543 137821813905216 maxtext_utils.py:1707] params/params/decoder/layers/mlp/wi_1/kernel Shape: float32[2048,16,7168] Logical: PartitionSpec('embed', 'layers', 'mlp') Physical: ('fsdp', None, None) I0423 02:52:48.921589 137821813905216 maxtext_utils.py:1707] params/params/decoder/layers/mlp/wo/kernel Shape: float32[7168,16,2048] Logical: PartitionSpec('mlp', 'layers', 'embed') Physical: (None, None, 'fsdp') I0423 02:52:48.921636 137821813905216 maxtext_utils.py:1707] params/params/decoder/layers/post_self_attention_layer_norm/scale Shape: float32[2048,16] Logical: PartitionSpec('norm', 'layers') Physical: (None, None) I0423 02:52:48.921665 137821813905216 maxtext_utils.py:1707] params/params/decoder/layers/pre_self_attention_layer_norm/scale Shape: float32[2048,16] Logical: PartitionSpec('norm', 'layers') Physical: (None, None) I0423 02:52:48.921705 137821813905216 maxtext_utils.py:1707] params/params/decoder/layers/self_attention/key/kernel Shape: float32[2048,16,16,128] Logical: PartitionSpec('embed', 'layers', 'kv_heads', 'kv_head_dim') Physical: ('fsdp', None, None, None) I0423 02:52:48.921746 137821813905216 maxtext_utils.py:1707] params/params/decoder/layers/self_attention/out/kernel Shape: float32[16,16,128,2048] Logical: PartitionSpec('heads', 'layers', 'kv', 'embed') Physical: (None, None, None, 'fsdp') I0423 02:52:48.921776 137821813905216 maxtext_utils.py:1707] params/params/decoder/layers/self_attention/query/kernel Shape: float32[2048,16,16,128] Logical: PartitionSpec('embed', 'layers', 'q_heads', 'kv') Physical: ('fsdp', None, None, None) I0423 02:52:48.921803 137821813905216 maxtext_utils.py:1707] params/params/decoder/layers/self_attention/value/kernel Shape: float32[2048,16,16,128] Logical: PartitionSpec('embed', 'layers', 'kv_heads', 'kv_head_dim') Physical: ('fsdp', None, None, None) I0423 02:52:48.921838 137821813905216 maxtext_utils.py:1707] params/params/decoder/logits_dense/kernel Shape: float32[2048,32000] Logical: PartitionSpec('embed_vocab', 'vocab') Physical: ('fsdp', None) I0423 02:52:48.921872 137821813905216 maxtext_utils.py:1707] params/params/token_embedder/embedding Shape: float32[32000,2048] Logical: PartitionSpec('vocab', 'embed_vocab') Physical: (None, 'fsdp') I0423 02:52:50.666145 137821813905216 train.py:155] train/xent Logical: float32[8,2048]............................................. ('activation_embed_and_logits_batch', 'activation_length'). I0423 02:52:50.666254 137821813905216 train.py:155] train/xent Physical: float32[8,2048]............................................. ('fsdp', None). I0423 02:52:50.678474 137821813905216 train.py:162] train/z_loss Logical: float32[8,2048]............................................. ('activation_embed_and_logits_batch', 'activation_length'). I0423 02:52:50.678525 137821813905216 train.py:162] train/z_loss Physical: float32[8,2048]............................................. ('fsdp', None). I0423 02:52:52.707076 137821813905216 max_utils.py:791] Total memory size: 3.7 GB, Output size: 1.5 GB, Temp size: 2.2 GB, Argument size: 1.5 GB, Host temp size: 0.0 GB. I0423 02:52:52.707656 137821813905216 max_utils.py:194] tensorboardX not available; using no-op SummaryWriter. I0423 02:52:52.707820 137821813905216 metric_logger.py:301] number parameters: 1.104 billion I0423 02:52:58.768559 137821813905216 checkpointing.py:772] Waiting for step 0 to finish before checkpoint... I0423 02:52:58.895593 137821813905216 checkpointing.py:776] Waited 0.12702441215515137 seconds for step 0 to finish before starting checkpointing. I0423 02:52:58.896263 137821813905216 checkpoint_manager.py:1983] [process=0][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning. I0423 02:52:58.896421 137821813905216 checkpoint_manager.py:1501] [process=0] Saving checkpoint at step 0 I0423 02:52:58.896793 137821813905216 async_checkpointer.py:452] [process=0] Started async saving checkpoint to gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints/0. I0423 02:52:58.988540 137821813905216 signaling_client.py:373] Using ThreadSafeKeyValueSignalingClient I0423 02:52:59.505320 137694682285632 atomicity.py:137] Creating tmp directory gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints/0 I0423 02:52:59.512912 137821813905216 jax_array_handlers.py:347] Scheduling D2H of 39 prioritized jax.Array. I0423 02:52:59.513005 137821813905216 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 I0423 02:53:00.280321 137694671799872 atomicity.py:137] Creating tmp directory gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints/0/items I0423 02:53:00.442478 137694197843520 checkpoint.py:188] Wrote Metadata={'item_handlers': None, 'metrics': {}, 'performance_metrics': {}, 'init_timestamp_nsecs': 1776912780193129151, 'commit_timestamp_nsecs': None, 'custom_metadata': {}}, json={"item_handlers": null, "metrics": {}, "performance_metrics": {}, "init_timestamp_nsecs": 1776912780193129151, "commit_timestamp_nsecs": null, "custom_metadata": {}} to gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints/0/_CHECKPOINT_METADATA I0423 02:53:00.761521 2346570 google_auth_provider.cc:149] Using credentials at ~/.config/gcloud/application_default_credentials.json I0423 02:53:00.761576 2346570 google_auth_provider.cc:156] Using OAuth2 AuthProvider I0423 02:53:00.790650 137821813905216 base_pytree_checkpoint_handler.py:153] [process=0][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 1.278137s I0423 02:53:00.795627 137821813905216 base_pytree_checkpoint_handler.py:128] [process=0] /jax/checkpoint/write/blocking_gbytes_per_sec: 6.834 GiB/s (total gbytes: 12.3 GiB) (time elapsed: a second) (per-host) I0423 02:53:00.795711 137821813905216 base_pytree_checkpoint_handler.py:732] [process=0][thread=MainThread] Initiated Pytree async_save. Time taken: 1.805924s (batch_requests_ready=0.519163s, total_serialization_initiated=1.281928s, others=0.004833s) I0423 02:53:00.795807 137821813905216 composite_checkpoint_handler.py:715] [process=0][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 1.806794s (all_items=0.000020s, per_item={'items': '0.00001955'}, temp_paths=1.806774) I0423 02:53:00.797343 137694692771392 async_checkpointer.py:79] [process=0][thread=async_save] Background save thread started. I0423 02:53:00.797464 137821813905216 async_checkpointer.py:561] Finished blocking save. Time taken: 1.900995s. Continuing background save to gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints/0. I0423 02:53:00.797733 137821813905216 checkpoint_manager.py:1549] [process=0][thread=MainThread][step=0] Starting CheckpointManager Save Finalize thread=save_finalize I0423 02:53:00.797988 137693660972608 async_checkpointer.py:265] [process=0][thread=save_finalize] Waiting for background save thread=async_save. I0423 02:53:00.798140 137821813905216 standard_logger.py:34] {'step': 0, 'event_type': 'save', 'directory': 'gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776912778.89625, 'wait_for_prev_duration_secs': 4.506111145019531e-05, 'time_between_consecutive_saves_sec': None, 'checkpointer_blocking_start_time': 1776912778.8964434, 'checkpointer_blocking_duration_secs': 1.90118408203125, 'get_old_steps_start_time': 1776912780.7976534, 'get_old_steps_duration_secs': 3.218650817871094e-05, 'checkpoint_manager_blocking_start_time': 1776912778.8961701, 'checkpoint_manager_blocking_duration_secs': 1.9019315242767334} I0423 02:53:00.798249 137821813905216 checkpointing.py:408] Started an asynchronous checkpoint save for step 0 I0423 02:53:00.798299 137821813905216 max_utils.py:750] Memstats: After params initialized: I0423 02:53:00.798355 137821813905216 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_0(process=0,(0,0,0,0)) I0423 02:53:00.798383 137821813905216 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_1(process=0,(1,0,0,0)) I0423 02:53:00.798404 137821813905216 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_2(process=0,(0,1,0,0)) I0423 02:53:00.798424 137821813905216 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_3(process=0,(1,1,0,0)) I0423 02:53:00.798442 137821813905216 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_4(process=0,(0,2,0,0)) I0423 02:53:00.798460 137821813905216 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_5(process=0,(1,2,0,0)) I0423 02:53:00.798476 137821813905216 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_6(process=0,(0,3,0,0)) I0423 02:53:00.798494 137821813905216 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_7(process=0,(1,3,0,0)) I0423 02:53:01.171426 137821813905216 metric_logger.py:196] completed step: 0, seconds: 6.060, TFLOP/s/device: 2.242, Tokens/s/device: 337.942, total_weights: 13328, loss: 10.840, lm_loss: 10.840, perplexity: 51025.957 I0423 02:53:01.172346 137821813905216 metric_logger.py:281] To see full metrics 'tensorboard --logdir=gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/tensorboard/' I0423 02:53:01.290863 137821813905216 metric_logger.py:196] completed step: 1, seconds: 2.401, TFLOP/s/device: 5.660, Tokens/s/device: 853.135, total_weights: 12332, loss: 10.869, lm_loss: 10.869, perplexity: 52515.969 I0423 02:53:01.412251 137821813905216 metric_logger.py:196] completed step: 2, seconds: 0.011, TFLOP/s/device: 1258.650, Tokens/s/device: 189717.462, total_weights: 15161, loss: 9.874, lm_loss: 9.874, perplexity: 19421.770 I0423 02:53:01.533713 137821813905216 metric_logger.py:196] completed step: 3, seconds: 0.121, TFLOP/s/device: 112.103, Tokens/s/device: 16897.411, total_weights: 13327, loss: 9.406, lm_loss: 9.406, perplexity: 12166.079 I0423 02:53:01.540899 137693929408064 array_metadata_store.py:203] [process=0][thread=array_type_handler] Wrote 39 array_metadata.ArrayMetadata to gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints/0/items/array_metadatas/process_0 I0423 02:53:22.267761 137821813905216 metric_logger.py:196] completed step: 4, seconds: 0.121, TFLOP/s/device: 112.402, Tokens/s/device: 16942.422, total_weights: 11939, loss: 9.149, lm_loss: 9.149, perplexity: 9403.998 I0423 02:53:22.394293 137821813905216 metric_logger.py:196] completed step: 5, seconds: 0.122, TFLOP/s/device: 111.466, Tokens/s/device: 16801.345, total_weights: 15502, loss: 9.058, lm_loss: 9.058, perplexity: 8587.564 I0423 02:53:22.859387 137821813905216 metric_logger.py:196] completed step: 6, seconds: 20.826, TFLOP/s/device: 0.652, Tokens/s/device: 98.341, total_weights: 13864, loss: 8.892, lm_loss: 8.892, perplexity: 7274.692 I0423 02:53:23.299224 137821813905216 metric_logger.py:196] completed step: 7, seconds: 0.174, TFLOP/s/device: 78.298, Tokens/s/device: 11801.926, total_weights: 12988, loss: 8.751, lm_loss: 8.751, perplexity: 6314.503 I0423 02:53:23.715969 137821813905216 metric_logger.py:196] completed step: 8, seconds: 0.755, TFLOP/s/device: 18.002, Tokens/s/device: 2713.384, total_weights: 13820, loss: 8.791, lm_loss: 8.791, perplexity: 6573.798 I0423 02:53:23.849504 137821813905216 checkpointing.py:772] Waiting for step 9 to finish before checkpoint... I0423 02:53:23.947419 137821813905216 checkpointing.py:776] Waited 0.09790229797363281 seconds for step 9 to finish before starting checkpointing. I0423 02:53:24.016416 137821813905216 checkpoint_manager.py:1994] [process=0][thread=MainThread][step=0][wait_until_finished] Waiting for Save Finalize thread (save_finalize) to complete. I0423 02:54:16.260111 137693918922304 base_pytree_checkpoint_handler.py:1217] [process=0][thread=write_metadata_after_commits] Commit + Array metadata written. Time taken: 75.463621s (commit=74.980988s, array_metadata_write=0.482633s) I0423 02:54:16.261160 137694692771392 base_pytree_checkpoint_handler.py:128] [process=0] /jax/checkpoint/write/gbytes_per_sec: 163.538 MiB/s (total gbytes: 12.3 GiB) (time elapsed: a minute) (per-host) I0423 02:54:16.261273 137694692771392 async_checkpointer.py:90] [process=0][thread=async_save] 3 Handler Commit operations completed. Time taken: 75.463744s. I0423 02:54:16.574056 137694692771392 checkpoint.py:228] Read Metadata={'item_handlers': None, 'metrics': {}, 'performance_metrics': {}, 'init_timestamp_nsecs': 1776912780193129151, 'commit_timestamp_nsecs': None, 'custom_metadata': {}} from gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints/0/_CHECKPOINT_METADATA I0423 02:54:16.781145 137694692771392 array_metadata_store.py:367] [process=0][thread=async_save] Skipped cross-host ArrayMetadata validation because only one process is found: process_index=0. I0423 02:54:17.005439 137694197843520 checkpoint.py:247] Updated Metadata={'item_handlers': {'items': 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler'}, 'metrics': {}, 'performance_metrics': {}, 'init_timestamp_nsecs': 1776912780193129151, 'commit_timestamp_nsecs': None, 'custom_metadata': {}} to gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints/0/_CHECKPOINT_METADATA I0423 02:54:17.189851 137694692771392 ocdbt_utils.py:56] Param validation support for Zarr3 will be added later (b/362328389). I0423 02:54:17.190013 137694692771392 base_pytree_checkpoint_handler.py:1342] [process=0][thread=async_save] Pytree save finalize (merge_ocdbt + ArrayMetadata validation) completed. Time taken: 0.570441s. use_zarr3=True, enable_post_merge_validation=True, directory=gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints/0/items I0423 02:54:17.190772 137694692771392 atomicity.py:608] Finalizing gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints/0/items I0423 02:54:17.428842 137694692771392 atomicity.py:608] Finalizing gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints/0 I0423 02:54:18.149958 137694692771392 atomicity.py:794] [process=0][thread=async_save] Finished saving checkpoint (finalized tmp dir) to `gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints/0`. I0423 02:54:18.150590 137694692771392 async_checkpointer.py:420] Finished async_save (blocking + background). Time taken: 79.254132s. directory=gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints/0 I0423 02:54:18.150681 137694692771392 async_checkpointer.py:144] [process=0][thread=async_save] Background save thread done. Time taken: 77.353156s. I0423 02:54:18.150906 137693660972608 async_checkpointer.py:273] [process=0][thread=save_finalize] Done with waiting for background save thread=async_save. I0423 02:54:18.151039 137693660972608 async_checkpointer.py:283] [process=0][thread=save_finalize] No errors found in background save thread=async_save. I0423 02:54:18.151116 137693660972608 checkpoint_manager.py:2103] [process=0][thread=save_finalize][step=0] CheckpointManager Save Finalize is syncing with other hosts... I0423 02:54:18.151165 137693660972608 checkpoint_manager.py:2112] [process=0][thread=save_finalize][step=0] CheckpointManager Save Finalize is done on all hosts. I0423 02:54:18.151419 137821813905216 checkpoint_manager.py:2006] [process=0][thread=MainThread][step=0][wait_until_finished] Done waiting for Save Finalize thread (save_finalize) running at step=0. W0423 02:54:18.151496 137821813905216 checkpoint_manager.py:1441] Waiting for previous save to complete took 54.159565 seconds. If this number is high, consider checkpointing less frequently. I0423 02:54:18.152585 137821813905216 checkpoint_manager.py:1501] [process=0] Saving checkpoint at step 9 I0423 02:54:18.152857 137821813905216 async_checkpointer.py:452] [process=0] Started async saving checkpoint to gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints/9. I0423 02:54:18.325346 137821813905216 jax_array_handlers.py:347] Scheduling D2H of 39 prioritized jax.Array. I0423 02:54:18.325483 137694692771392 atomicity.py:137] Creating tmp directory gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints/9 I0423 02:54:18.325564 137821813905216 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 I0423 02:54:19.033083 137693650486848 atomicity.py:137] Creating tmp directory gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints/9/items I0423 02:54:22.323209 137821813905216 base_pytree_checkpoint_handler.py:153] [process=0][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 3.998388s I0423 02:54:22.328093 137821813905216 base_pytree_checkpoint_handler.py:128] [process=0] /jax/checkpoint/write/blocking_gbytes_per_sec: 3.018 GiB/s (total gbytes: 12.3 GiB) (time elapsed: 4 seconds) (per-host) I0423 02:54:22.328157 137821813905216 base_pytree_checkpoint_handler.py:732] [process=0][thread=MainThread] Initiated Pytree async_save. Time taken: 4.089374s (batch_requests_ready=0.083121s, total_serialization_initiated=4.001469s, others=0.004784s) I0423 02:54:22.328237 137821813905216 composite_checkpoint_handler.py:715] [process=0][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 4.089905s (all_items=0.000022s, per_item={'items': '0.00002241'}, temp_paths=4.089882) I0423 02:54:22.329867 137694692771392 async_checkpointer.py:79] [process=0][thread=async_save] Background save thread started. I0423 02:54:22.329986 137821813905216 async_checkpointer.py:561] Finished blocking save. Time taken: 4.177348s. Continuing background save to gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints/9. I0423 02:54:22.330163 137821813905216 checkpoint_manager.py:1549] [process=0][thread=MainThread][step=9] Starting CheckpointManager Save Finalize thread=save_finalize I0423 02:54:22.330414 137693660972608 async_checkpointer.py:265] [process=0][thread=save_finalize] Waiting for background save thread=async_save. I0423 02:54:22.330544 137821813905216 standard_logger.py:34] {'step': 9, 'event_type': 'save', 'directory': 'gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776912803.991907, 'wait_for_prev_duration_secs': 54.15956521034241, 'time_between_consecutive_saves_sec': None, 'checkpointer_blocking_start_time': 1776912858.1526163, 'checkpointer_blocking_duration_secs': 4.177448272705078, 'get_old_steps_start_time': 1776912862.3300838, 'get_old_steps_duration_secs': 3.6716461181640625e-05, 'checkpoint_manager_blocking_start_time': 1776912803.989501, 'checkpoint_manager_blocking_duration_secs': 58.3410165309906} I0423 02:54:22.330659 137821813905216 checkpointing.py:408] Started an asynchronous checkpoint save for step 9 I0423 02:54:22.330696 137821813905216 checkpoint_manager.py:1994] [process=0][thread=MainThread][step=9][wait_until_finished] Waiting for Save Finalize thread (save_finalize) to complete. I0423 02:54:23.103070 137694233495104 array_metadata_store.py:203] [process=0][thread=array_type_handler] Wrote 39 array_metadata.ArrayMetadata to gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints/9/items/array_metadatas/process_0 I0423 02:54:53.869263 137693650486848 base_pytree_checkpoint_handler.py:1217] [process=0][thread=write_metadata_after_commits] Commit + Array metadata written. Time taken: 31.540676s (commit=31.083406s, array_metadata_write=0.457270s) I0423 02:54:53.870474 137694692771392 base_pytree_checkpoint_handler.py:128] [process=0] /jax/checkpoint/write/gbytes_per_sec: 354.651 MiB/s (total gbytes: 12.3 GiB) (time elapsed: 35 seconds) (per-host) I0423 02:54:53.870585 137694692771392 async_checkpointer.py:90] [process=0][thread=async_save] 3 Handler Commit operations completed. Time taken: 31.540635s. I0423 02:54:54.323866 137694692771392 array_metadata_store.py:367] [process=0][thread=async_save] Skipped cross-host ArrayMetadata validation because only one process is found: process_index=0. I0423 02:54:54.726325 137694692771392 ocdbt_utils.py:56] Param validation support for Zarr3 will be added later (b/362328389). I0423 02:54:54.726484 137694692771392 base_pytree_checkpoint_handler.py:1342] [process=0][thread=async_save] Pytree save finalize (merge_ocdbt + ArrayMetadata validation) completed. Time taken: 0.564823s. use_zarr3=True, enable_post_merge_validation=True, directory=gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints/9/items I0423 02:54:54.727122 137694692771392 atomicity.py:608] Finalizing gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints/9/items I0423 02:54:54.968220 137694692771392 atomicity.py:608] Finalizing gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints/9 I0423 02:54:55.661465 137694692771392 atomicity.py:794] [process=0][thread=async_save] Finished saving checkpoint (finalized tmp dir) to `gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints/9`. I0423 02:54:55.662118 137694692771392 async_checkpointer.py:420] Finished async_save (blocking + background). Time taken: 37.509486s. directory=gs://wanglance-maxtext/linen_ckpt_fix_gate_fp8_nan_sanitization_20260423_024158/linen_fix_gate_fp8_nan_sanitization_20260423_024158_04_int8/checkpoints/9 I0423 02:54:55.662200 137694692771392 async_checkpointer.py:144] [process=0][thread=async_save] Background save thread done. Time taken: 33.332252s. I0423 02:54:55.662430 137693660972608 async_checkpointer.py:273] [process=0][thread=save_finalize] Done with waiting for background save thread=async_save. I0423 02:54:55.662528 137693660972608 async_checkpointer.py:283] [process=0][thread=save_finalize] No errors found in background save thread=async_save. I0423 02:54:55.662598 137693660972608 checkpoint_manager.py:2103] [process=0][thread=save_finalize][step=9] CheckpointManager Save Finalize is syncing with other hosts... I0423 02:54:55.662647 137693660972608 checkpoint_manager.py:2112] [process=0][thread=save_finalize][step=9] CheckpointManager Save Finalize is done on all hosts. I0423 02:54:55.662807 137821813905216 checkpoint_manager.py:2006] [process=0][thread=MainThread][step=9][wait_until_finished] Done waiting for Save Finalize thread (save_finalize) running at step=9. I0423 02:54:55.662947 137821813905216 checkpoint_manager.py:1983] [process=0][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning. I0423 02:54:55.664062 137821813905216 metric_logger.py:196] completed step: 9, seconds: 0.416, TFLOP/s/device: 32.633, Tokens/s/device: 4918.856, total_weights: 12300, loss: 8.774, lm_loss: 8.774, perplexity: 6465.313 Per train step: Total TFLOPs: 13.59 split as 93.93% learnable weight flops and 6.07% attention flops