feat/nnx-post-train-fixes| Metric | Linen d8cde296b | NNX d8cde296b | Diff (NNX − Linen) |
|---|---|---|---|
| Parameters | 1.104 billion | 1.104 billion | — |
| Final loss | 8.7740 | 8.7550 | -0.019 |
| TFLOP/s | 111.185 | 111.715 | +0.53 |
| Tok/s | 16759.0 | 16838.9 | +79.921 |
| Avg s/step | 3.351 | 3.176 | -0.175 |
| Memory % | 5.12 | 5.12 | 0 |
| JAX | 0.9.2 | 0.9.2 | — |
Diff = NNX value − Linen value. Green = NNX improved. Red = NNX regressed.
2026-04-16 18:23:40.379898: 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) I0416 18:23:40.496993 139593135533184 max_utils.py:238] Skipping jax distributed system due to skip_jax_distributed_system=True flag. I0416 18:24:31.339857 139593135533184 max_utils.py:800] System Information: Jax Version: 0.9.2 I0416 18:24:31.340078 139593135533184 max_utils.py:801] System Information: Jaxlib Version: 0.9.2 I0416 18:24:31.340122 139593135533184 max_utils.py:802] System Information: Jax Backend: PJRT C API TFRT TPU v6 lite Built on Apr 6 2026 20:48:10 (1775533690) cl/895581894 I0416 18:24:31.340149 139593135533184 train_utils.py:364] WARNING: 'dataset_path' might be pointing your local file system I0416 18:24:31.340287 139593135533184 train.py:811] [DECOUPLED NO-OP] skipping cloud diagnostics wrapper. W0416 18:24:31.434996 3976941 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. I0416 18:24:31.830518 139593135533184 maxtext_utils.py:1687] Num_devices: 8, shape (1, 1, 1, 8, 1, 1, 1, 1, 1, 1, 1, 1, 1) I0416 18:24:31.830693 139593135533184 maxtext_utils.py:1687] Num_devices: 8, shape (1, 1, 1, 8, 1, 1, 1, 1, 1, 1, 1, 1, 1) I0416 18:24:31.831064 139593135533184 checkpointing.py:688] Setting up checkpoint logger... I0416 18:24:31.831108 139593135533184 checkpointing.py:234] Creating checkpoint manager with ocdbt=True and zarr3=True I0416 18:24:31.831173 139593135533184 pytree_checkpoint_handler.py:577] save_device_host_concurrent_bytes=None I0416 18:24:31.831709 139593135533184 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 0x7ef4f85f4c50>, enable_pinned_host_transfer=False, save_concurrent_bytes: 96000000000 (89.4 GiB), restore_concurrent_bytes: 96000000000 (89.4 GiB) I0416 18:24:34.248652 139593135533184 checkpointing.py:266] Enabling policy for fixed interval checkpointing. I0416 18:24:34.248868 139593135533184 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 0x7eee7c33de80>}, handler_registry=None I0416 18:24:34.249189 139593135533184 composite_checkpoint_handler.py:237] Deferred registration for item: "items". Adding handler `<orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7eee7c33de80>` 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`. I0416 18:24:34.249232 139593135533184 composite_checkpoint_handler.py:237] Deferred registration for item: "metrics". Adding handler `<orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7eee7c33e030>` 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`. I0416 18:24:34.249260 139593135533184 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 0x7eee7c33de80>, ('items', <class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeRestoreArgs'>): <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7eee7c33de80>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonSaveArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7eee7c33e030>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonRestoreArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7eee7c33e030>}). I0416 18:24:34.249577 139593135533184 abstract_checkpointer.py:35] orbax-checkpoint version: 0.11.28 I0416 18:24:34.249631 139593135533184 async_checkpointer.py:177] [process=0][thread=MainThread] Using barrier_sync_fn: <function get_barrier_sync_fn.<locals>.<lambda> at 0x7eee7c329ee0> timeout: 600 secs and primary_host=0 for async checkpoint writes I0416 18:24:34.359250 139593135533184 checkpoint_manager.py:1788] Found 0 checkpoint steps in gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints I0416 18:24:34.359471 139593135533184 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_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints: <orbax.checkpoint.checkpoint_manager.CheckpointManager object at 0x7eee7c70af60> I0416 18:24:34.359557 139593135533184 checkpointing.py:302] Checkpoint manager created! I0416 18:24:34.445660 139593135533184 dataset_info.py:707] Load dataset info from tests/assets/local_datasets/c4_en_dataset_minimal/c4/en/3.1.0 I0416 18:24:34.449185 139593135533184 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. I0416 18:24:34.505671 139593135533184 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 I0416 18:24:34.535771 139593135533184 tokenizer.py:245] Tokenizer path: src/maxtext/assets/tokenizers/tokenizer.llama2 I0416 18:24:34.535834 139593135533184 tokenizer.py:187] Loading sentencepiece tokenizer: src/maxtext/assets/tokenizers/tokenizer.llama2 I0416 18:24:35.708932 139593135533184 nnx_wrappers.py:437] Unknown Logical: bfloat16[8,2048,2048]....................................... ('activation_batch', 'activation_norm_length', 'activation_embed'). I0416 18:24:35.709027 139593135533184 nnx_wrappers.py:437] Unknown Physical: bfloat16[8,2048,2048]....................................... ('fsdp', None, None). I0416 18:24:35.813749 139593135533184 attentions.py:1088] attentions/inputs_q Logical: bfloat16[8,2048,2048]....................................... ('activation_batch', 'activation_attn_length', 'activation_attn_embed'). I0416 18:24:35.813829 139593135533184 attentions.py:1088] attentions/inputs_q Physical: bfloat16[8,2048,2048]....................................... ('fsdp', None, None). I0416 18:24:35.829001 139593135533184 attentions.py:1089] attentions/inputs_kv Logical: bfloat16[8,2048,2048]....................................... ('activation_batch', 'activation_attn_length', 'activation_attn_embed'). I0416 18:24:35.829067 139593135533184 attentions.py:1089] attentions/inputs_kv Physical: bfloat16[8,2048,2048]....................................... ('fsdp', None, None). I0416 18:24:35.886563 139593135533184 attentions.py:1154] attentions/query Logical: bfloat16[8,2048,16,128]..................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim'). I0416 18:24:35.886635 139593135533184 attentions.py:1154] attentions/query Physical: bfloat16[8,2048,16,128]..................................... ('fsdp', None, None, None). I0416 18:24:35.901832 139593135533184 attentions.py:1155] attentions/key Logical: bfloat16[8,2048,16,128]..................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim'). I0416 18:24:35.901883 139593135533184 attentions.py:1155] attentions/key Physical: bfloat16[8,2048,16,128]..................................... ('fsdp', None, None, None). I0416 18:24:35.917021 139593135533184 attentions.py:1156] attentions/value Logical: bfloat16[8,2048,16,128]..................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim'). I0416 18:24:35.917085 139593135533184 attentions.py:1156] attentions/value Physical: bfloat16[8,2048,16,128]..................................... ('fsdp', None, None, None). I0416 18:24:35.939551 139593135533184 attentions.py:1197] attentions/out Logical: bfloat16[8,2048,16,128]..................................... ('activation_batch', 'activation_attn_length', 'activation_heads', 'activation_kv'). I0416 18:24:35.939610 139593135533184 attentions.py:1197] attentions/out Physical: bfloat16[8,2048,16,128]..................................... ('fsdp', None, None, None). I0416 18:24:35.996908 139593135533184 linears.py:525] linears/x Logical: bfloat16[8,2048,7168]....................................... ('activation_batch', 'activation_length', 'activation_mlp'). I0416 18:24:35.996977 139593135533184 linears.py:525] linears/x Physical: bfloat16[8,2048,7168]....................................... ('fsdp', None, None). I0416 18:24:36.438726 139593135533184 checkpointing.py:578] checkpoint manager exists so trying to load this run's existing checkpoint I0416 18:24:36.438822 139593135533184 checkpointing.py:676] 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 I0416 18:24:38.346309 139593135533184 maxtext_utils.py:1796] params/params/decoder/decoder_norm/scale Shape: float32[2048] Logical: P('norm',) Physical: (None,) I0416 18:24:38.346416 139593135533184 maxtext_utils.py:1796] params/params/decoder/layers/mlp/wi_0/kernel Shape: float32[2048,16,7168] Logical: P('embed', 'layers', 'mlp') Physical: ('fsdp', None, None) I0416 18:24:38.346458 139593135533184 maxtext_utils.py:1796] params/params/decoder/layers/mlp/wi_1/kernel Shape: float32[2048,16,7168] Logical: P('embed', 'layers', 'mlp') Physical: ('fsdp', None, None) I0416 18:24:38.346501 139593135533184 maxtext_utils.py:1796] params/params/decoder/layers/mlp/wo/kernel Shape: float32[7168,16,2048] Logical: P('mlp', 'layers', 'embed') Physical: (None, None, 'fsdp') I0416 18:24:38.346540 139593135533184 maxtext_utils.py:1796] params/params/decoder/layers/post_self_attention_layer_norm/scale Shape: float32[2048,16] Logical: P('norm', 'layers') Physical: (None, None) I0416 18:24:38.346569 139593135533184 maxtext_utils.py:1796] params/params/decoder/layers/pre_self_attention_layer_norm/scale Shape: float32[2048,16] Logical: P('norm', 'layers') Physical: (None, None) I0416 18:24:38.346611 139593135533184 maxtext_utils.py:1796] params/params/decoder/layers/self_attention/key/kernel Shape: float32[2048,16,16,128] Logical: P('embed', 'layers', 'kv_heads', 'kv_head_dim') Physical: ('fsdp', None, None, None) I0416 18:24:38.346653 139593135533184 maxtext_utils.py:1796] params/params/decoder/layers/self_attention/out/kernel Shape: float32[16,16,128,2048] Logical: P('heads', 'layers', 'kv', 'embed') Physical: (None, None, None, 'fsdp') I0416 18:24:38.346682 139593135533184 maxtext_utils.py:1796] params/params/decoder/layers/self_attention/query/kernel Shape: float32[2048,16,16,128] Logical: P('embed', 'layers', 'q_heads', 'kv') Physical: ('fsdp', None, None, None) I0416 18:24:38.346709 139593135533184 maxtext_utils.py:1796] params/params/decoder/layers/self_attention/value/kernel Shape: float32[2048,16,16,128] Logical: P('embed', 'layers', 'kv_heads', 'kv_head_dim') Physical: ('fsdp', None, None, None) I0416 18:24:38.346746 139593135533184 maxtext_utils.py:1796] params/params/decoder/logits_dense/kernel Shape: float32[2048,32000] Logical: P('embed_vocab', 'vocab') Physical: ('fsdp', None) I0416 18:24:38.346781 139593135533184 maxtext_utils.py:1796] params/params/token_embedder/embedding Shape: float32[32000,2048] Logical: P('vocab', 'embed_vocab') Physical: (None, 'fsdp') I0416 18:24:39.986885 139593135533184 train.py:157] train/xent Logical: float32[8,2048]............................................. ('activation_embed_and_logits_batch', 'activation_length'). I0416 18:24:39.987002 139593135533184 train.py:157] train/xent Physical: float32[8,2048]............................................. ('fsdp', None). I0416 18:24:40.001944 139593135533184 train.py:164] train/z_loss Logical: float32[8,2048]............................................. ('activation_embed_and_logits_batch', 'activation_length'). I0416 18:24:40.001999 139593135533184 train.py:164] train/z_loss Physical: float32[8,2048]............................................. ('fsdp', None). W0416 18:24:40.571165 3976941 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. I0416 18:24:40.755685 139593135533184 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. I0416 18:24:40.756330 139593135533184 max_utils.py:194] tensorboardX not available; using no-op SummaryWriter. I0416 18:24:40.756486 139593135533184 metric_logger.py:289] number parameters: 1.104 billion W0416 18:24:45.385007 3976941 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. I0416 18:24:45.609979 139593135533184 checkpointing.py:794] Waiting for step 0 to finish before checkpoint... I0416 18:24:45.736386 139593135533184 checkpointing.py:798] Waited 0.12639546394348145 seconds for step 0 to finish before starting checkpointing. I0416 18:24:45.736814 139593135533184 checkpoint_manager.py:1983] [process=0][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning. I0416 18:24:45.736988 139593135533184 checkpoint_manager.py:1501] [process=0] Saving checkpoint at step 0 I0416 18:24:45.737368 139593135533184 async_checkpointer.py:452] [process=0] Started async saving checkpoint to gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0. I0416 18:24:45.824622 139593135533184 signaling_client.py:373] Using ThreadSafeKeyValueSignalingClient I0416 18:24:45.911198 139593135533184 jax_array_handlers.py:347] Scheduling D2H of 39 prioritized jax.Array. I0416 18:24:45.911287 139593135533184 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 I0416 18:24:45.917221 139477840299584 atomicity.py:137] Creating tmp directory gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0 W0416 18:24:46.141358 3976941 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. W0416 18:24:46.152557 3976941 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. W0416 18:24:46.165680 3976941 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. W0416 18:24:46.176362 3976941 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. W0416 18:24:46.184310 3976941 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. W0416 18:24:46.189723 3976941 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. W0416 18:24:46.193981 3976941 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. W0416 18:24:46.199245 3976941 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. I0416 18:24:46.574914 139593135533184 base_pytree_checkpoint_handler.py:153] [process=0][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 0.664115s I0416 18:24:46.575357 139593135533184 base_pytree_checkpoint_handler.py:128] [process=0] /jax/checkpoint/write/blocking_gbytes_per_sec: 16.460 GiB/s (total gbytes: 12.3 GiB) (time elapsed: 749 milliseconds) (per-host) I0416 18:24:46.575407 139593135533184 base_pytree_checkpoint_handler.py:732] [process=0][thread=MainThread] Initiated Pytree async_save. Time taken: 0.749828s (batch_requests_ready=0.081782s, total_serialization_initiated=0.667720s, others=0.000327s) I0416 18:24:46.575467 139593135533184 composite_checkpoint_handler.py:715] [process=0][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 0.750355s (all_items=0.000019s, per_item={'items': '0.00001860'}, temp_paths=0.750337) I0416 18:24:46.576188 139477768996416 async_checkpointer.py:79] [process=0][thread=async_save] Background save thread started. I0416 18:24:46.576281 139593135533184 async_checkpointer.py:561] Finished blocking save. Time taken: 0.839255s. Continuing background save to gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0. I0416 18:24:46.576448 139593135533184 checkpoint_manager.py:1549] [process=0][thread=MainThread][step=0] Starting CheckpointManager Save Finalize thread=save_finalize I0416 18:24:46.576623 139477758510656 async_checkpointer.py:265] [process=0][thread=save_finalize] Waiting for background save thread=async_save. I0416 18:24:46.576732 139593135533184 standard_logger.py:34] {'step': 0, 'event_type': 'save', 'directory': 'gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776363885.7368026, 'wait_for_prev_duration_secs': 4.100799560546875e-05, 'time_between_consecutive_saves_sec': None, 'checkpointer_blocking_start_time': 1776363885.7370088, 'checkpointer_blocking_duration_secs': 0.839371919631958, 'get_old_steps_start_time': 1776363886.576396, 'get_old_steps_duration_secs': 2.0742416381835938e-05, 'checkpoint_manager_blocking_start_time': 1776363885.7367198, 'checkpoint_manager_blocking_duration_secs': 0.8399820327758789} I0416 18:24:46.576846 139593135533184 checkpointing.py:409] Started an asynchronous checkpoint save for step 0 I0416 18:24:46.576884 139593135533184 max_utils.py:750] Memstats: After params initialized: I0416 18:24:46.576967 139593135533184 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_0(process=0,(0,0,0,0)) I0416 18:24:46.576992 139593135533184 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_1(process=0,(1,0,0,0)) I0416 18:24:46.577011 139593135533184 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_2(process=0,(0,1,0,0)) I0416 18:24:46.577028 139593135533184 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_3(process=0,(1,1,0,0)) I0416 18:24:46.577058 139593135533184 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_4(process=0,(0,2,0,0)) I0416 18:24:46.577075 139593135533184 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_5(process=0,(1,2,0,0)) I0416 18:24:46.577092 139593135533184 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_6(process=0,(0,3,0,0)) I0416 18:24:46.577108 139593135533184 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_7(process=0,(1,3,0,0)) W0416 18:24:46.582905 3976941 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. W0416 18:24:46.588248 3976941 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. W0416 18:24:46.592797 3976941 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. I0416 18:24:46.689678 139477829813824 atomicity.py:137] Creating tmp directory gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0/items I0416 18:24:46.899258 139593135533184 metric_logger.py:185] completed step: 0, seconds: 4.853, TFLOP/s/device: 2.800, Tokens/s/device: 422.005, total_weights: 13328, loss: 10.840 I0416 18:24:46.900267 139593135533184 metric_logger.py:269] To see full metrics 'tensorboard --logdir=gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/tensorboard/' I0416 18:24:47.018843 139593135533184 metric_logger.py:185] completed step: 1, seconds: 1.287, TFLOP/s/device: 10.554, Tokens/s/device: 1590.825, total_weights: 12332, loss: 10.869 I0416 18:24:47.036090 139477850785344 checkpoint.py:188] Wrote Metadata={'item_handlers': None, 'metrics': {}, 'performance_metrics': {}, 'init_timestamp_nsecs': 1776363886593119488, 'commit_timestamp_nsecs': None, 'custom_metadata': {}}, json={"item_handlers": null, "metrics": {}, "performance_metrics": {}, "init_timestamp_nsecs": 1776363886593119488, "commit_timestamp_nsecs": null, "custom_metadata": {}} to gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0/_CHECKPOINT_METADATA I0416 18:24:47.140856 139593135533184 metric_logger.py:185] completed step: 2, seconds: 0.010, TFLOP/s/device: 1385.452, Tokens/s/device: 208830.427, total_weights: 15161, loss: 9.874 I0416 18:24:47.262702 139593135533184 metric_logger.py:185] completed step: 3, seconds: 0.121, TFLOP/s/device: 112.600, Tokens/s/device: 16972.329, total_weights: 13327, loss: 9.406 I0416 18:24:47.278914 3980170 google_auth_provider.cc:149] Using credentials at ~/.config/gcloud/application_default_credentials.json I0416 18:24:47.278971 3980170 google_auth_provider.cc:156] Using OAuth2 AuthProvider I0416 18:24:48.226354 139477789967936 array_metadata_store.py:203] [process=0][thread=array_type_handler] Wrote 39 array_metadata.ArrayMetadata to gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0/items/array_metadatas/process_0 I0416 18:25:15.506186 139593135533184 metric_logger.py:185] completed step: 4, seconds: 0.122, TFLOP/s/device: 111.660, Tokens/s/device: 16830.617, total_weights: 11939, loss: 9.149 I0416 18:25:15.516673 139593135533184 metric_logger.py:185] completed step: 5, seconds: 0.122, TFLOP/s/device: 111.285, Tokens/s/device: 16774.098, total_weights: 15502, loss: 9.058 I0416 18:25:15.637321 139593135533184 metric_logger.py:185] completed step: 6, seconds: 28.244, TFLOP/s/device: 0.481, Tokens/s/device: 72.511, total_weights: 13864, loss: 8.892 I0416 18:25:15.759576 139593135533184 metric_logger.py:185] completed step: 7, seconds: 0.006, TFLOP/s/device: 2108.166, Tokens/s/device: 317765.710, total_weights: 12988, loss: 8.751 I0416 18:25:15.881736 139593135533184 metric_logger.py:185] completed step: 8, seconds: 0.122, TFLOP/s/device: 111.497, Tokens/s/device: 16806.033, total_weights: 13820, loss: 8.791 I0416 18:25:16.004018 139593135533184 checkpointing.py:794] Waiting for step 9 to finish before checkpoint... I0416 18:25:16.004838 139593135533184 checkpointing.py:798] Waited 0.000850677490234375 seconds for step 9 to finish before starting checkpointing. I0416 18:25:16.005125 139593135533184 checkpoint_manager.py:1994] [process=0][thread=MainThread][step=0][wait_until_finished] Waiting for Save Finalize thread (save_finalize) to complete. I0416 18:25:19.290838 139477779482176 base_pytree_checkpoint_handler.py:1217] [process=0][thread=write_metadata_after_commits] Commit + Array metadata written. Time taken: 32.009210s (commit=31.578231s, array_metadata_write=0.430979s) I0416 18:25:19.291861 139477768996416 base_pytree_checkpoint_handler.py:128] [process=0] /jax/checkpoint/write/gbytes_per_sec: 377.598 MiB/s (total gbytes: 12.3 GiB) (time elapsed: 33 seconds) (per-host) I0416 18:25:19.291916 139477768996416 async_checkpointer.py:90] [process=0][thread=async_save] 3 Handler Commit operations completed. Time taken: 32.715589s. I0416 18:25:19.496364 139477768996416 checkpoint.py:228] Read Metadata={'item_handlers': None, 'metrics': {}, 'performance_metrics': {}, 'init_timestamp_nsecs': 1776363886593119488, 'commit_timestamp_nsecs': None, 'custom_metadata': {}} from gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0/_CHECKPOINT_METADATA I0416 18:25:19.669029 139477768996416 array_metadata_store.py:367] [process=0][thread=async_save] Skipped cross-host ArrayMetadata validation because only one process is found: process_index=0. I0416 18:25:19.876427 139477850785344 checkpoint.py:247] Updated Metadata={'item_handlers': {'items': 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler'}, 'metrics': {}, 'performance_metrics': {}, 'init_timestamp_nsecs': 1776363886593119488, 'commit_timestamp_nsecs': None, 'custom_metadata': {}} to gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0/_CHECKPOINT_METADATA I0416 18:25:20.108081 139477768996416 ocdbt_utils.py:56] Param validation support for Zarr3 will be added later (b/362328389). I0416 18:25:20.108676 139477768996416 base_pytree_checkpoint_handler.py:1342] [process=0][thread=async_save] Pytree save finalize (merge_ocdbt + ArrayMetadata validation) completed. Time taken: 0.573879s. use_zarr3=True, enable_post_merge_validation=True, directory=gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0/items I0416 18:25:20.109305 139477768996416 atomicity.py:608] Finalizing gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0/items I0416 18:25:20.370710 139477768996416 atomicity.py:608] Finalizing gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0 I0416 18:25:21.066282 139477768996416 atomicity.py:794] [process=0][thread=async_save] Finished saving checkpoint (finalized tmp dir) to `gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0`. I0416 18:25:21.066891 139477768996416 async_checkpointer.py:420] Finished async_save (blocking + background). Time taken: 35.329871s. directory=gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0 I0416 18:25:21.066966 139477768996416 async_checkpointer.py:144] [process=0][thread=async_save] Background save thread done. Time taken: 34.490641s. I0416 18:25:21.067158 139477758510656 async_checkpointer.py:273] [process=0][thread=save_finalize] Done with waiting for background save thread=async_save. I0416 18:25:21.067267 139477758510656 async_checkpointer.py:283] [process=0][thread=save_finalize] No errors found in background save thread=async_save. I0416 18:25:21.067329 139477758510656 checkpoint_manager.py:2103] [process=0][thread=save_finalize][step=0] CheckpointManager Save Finalize is syncing with other hosts... I0416 18:25:21.067374 139477758510656 checkpoint_manager.py:2112] [process=0][thread=save_finalize][step=0] CheckpointManager Save Finalize is done on all hosts. I0416 18:25:21.067514 139593135533184 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. W0416 18:25:21.067581 139593135533184 checkpoint_manager.py:1441] Waiting for previous save to complete took 5.062472 seconds. If this number is high, consider checkpointing less frequently. I0416 18:25:21.068380 139593135533184 checkpoint_manager.py:1501] [process=0] Saving checkpoint at step 9 I0416 18:25:21.068656 139593135533184 async_checkpointer.py:452] [process=0] Started async saving checkpoint to gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/9. I0416 18:25:21.258605 139477758510656 atomicity.py:137] Creating tmp directory gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/9 I0416 18:25:21.263825 139593135533184 jax_array_handlers.py:347] Scheduling D2H of 39 prioritized jax.Array. I0416 18:25:21.263927 139593135533184 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 I0416 18:25:21.961951 139477622195776 atomicity.py:137] Creating tmp directory gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/9/items I0416 18:25:26.143249 139593135533184 base_pytree_checkpoint_handler.py:153] [process=0][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 4.879803s I0416 18:25:26.147513 139593135533184 base_pytree_checkpoint_handler.py:128] [process=0] /jax/checkpoint/write/blocking_gbytes_per_sec: 2.475 GiB/s (total gbytes: 12.3 GiB) (time elapsed: 4 seconds) (per-host) I0416 18:25:26.147573 139593135533184 base_pytree_checkpoint_handler.py:732] [process=0][thread=MainThread] Initiated Pytree async_save. Time taken: 4.985748s (batch_requests_ready=0.098666s, total_serialization_initiated=4.882923s, others=0.004159s) I0416 18:25:26.147644 139593135533184 composite_checkpoint_handler.py:715] [process=0][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 4.986249s (all_items=0.000017s, per_item={'items': '0.00001717'}, temp_paths=4.986232) I0416 18:25:26.149103 139477789967936 async_checkpointer.py:79] [process=0][thread=async_save] Background save thread started. I0416 18:25:26.149194 139593135533184 async_checkpointer.py:561] Finished blocking save. Time taken: 5.080761s. Continuing background save to gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/9. I0416 18:25:26.149436 139593135533184 checkpoint_manager.py:1549] [process=0][thread=MainThread][step=9] Starting CheckpointManager Save Finalize thread=save_finalize I0416 18:25:26.149623 139477768996416 async_checkpointer.py:265] [process=0][thread=save_finalize] Waiting for background save thread=async_save. I0416 18:25:26.149727 139593135533184 standard_logger.py:34] {'step': 9, 'event_type': 'save', 'directory': 'gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776363916.0050836, 'wait_for_prev_duration_secs': 5.062471866607666, 'time_between_consecutive_saves_sec': None, 'checkpointer_blocking_start_time': 1776363921.0684044, 'checkpointer_blocking_duration_secs': 5.080932378768921, 'get_old_steps_start_time': 1776363926.1493552, 'get_old_steps_duration_secs': 3.314018249511719e-05, 'checkpoint_manager_blocking_start_time': 1776363916.00503, 'checkpoint_manager_blocking_duration_secs': 10.144667863845825} I0416 18:25:26.149859 139593135533184 checkpointing.py:409] Started an asynchronous checkpoint save for step 9 I0416 18:25:26.149893 139593135533184 checkpoint_manager.py:1994] [process=0][thread=MainThread][step=9][wait_until_finished] Waiting for Save Finalize thread (save_finalize) to complete. I0416 18:25:26.927970 139477840299584 array_metadata_store.py:203] [process=0][thread=array_type_handler] Wrote 39 array_metadata.ArrayMetadata to gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/9/items/array_metadatas/process_0 I0416 18:25:59.893119 139477758510656 base_pytree_checkpoint_handler.py:1217] [process=0][thread=write_metadata_after_commits] Commit + Array metadata written. Time taken: 33.744876s (commit=33.331750s, array_metadata_write=0.413126s) I0416 18:25:59.894420 139477789967936 base_pytree_checkpoint_handler.py:128] [process=0] /jax/checkpoint/write/gbytes_per_sec: 326.258 MiB/s (total gbytes: 12.3 GiB) (time elapsed: 38 seconds) (per-host) I0416 18:25:59.894539 139477789967936 async_checkpointer.py:90] [process=0][thread=async_save] 3 Handler Commit operations completed. Time taken: 33.745286s. I0416 18:26:00.375552 139477789967936 array_metadata_store.py:367] [process=0][thread=async_save] Skipped cross-host ArrayMetadata validation because only one process is found: process_index=0. I0416 18:26:00.826401 139477789967936 ocdbt_utils.py:56] Param validation support for Zarr3 will be added later (b/362328389). I0416 18:26:00.827004 139477789967936 base_pytree_checkpoint_handler.py:1342] [process=0][thread=async_save] Pytree save finalize (merge_ocdbt + ArrayMetadata validation) completed. Time taken: 0.603636s. use_zarr3=True, enable_post_merge_validation=True, directory=gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/9/items I0416 18:26:00.827633 139477789967936 atomicity.py:608] Finalizing gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/9/items I0416 18:26:01.071895 139477789967936 atomicity.py:608] Finalizing gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/9 I0416 18:26:01.753192 139477789967936 atomicity.py:794] [process=0][thread=async_save] Finished saving checkpoint (finalized tmp dir) to `gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/9`. I0416 18:26:01.753810 139477789967936 async_checkpointer.py:420] Finished async_save (blocking + background). Time taken: 40.685394s. directory=gs://wanglance-maxtext/linen_ckpt_feat_nnx_post_train_fixes_20260416_181521/linen_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/9 I0416 18:26:01.753894 139477789967936 async_checkpointer.py:144] [process=0][thread=async_save] Background save thread done. Time taken: 35.604644s. I0416 18:26:01.754102 139477768996416 async_checkpointer.py:273] [process=0][thread=save_finalize] Done with waiting for background save thread=async_save. I0416 18:26:01.754203 139477768996416 async_checkpointer.py:283] [process=0][thread=save_finalize] No errors found in background save thread=async_save. I0416 18:26:01.754275 139477768996416 checkpoint_manager.py:2103] [process=0][thread=save_finalize][step=9] CheckpointManager Save Finalize is syncing with other hosts... I0416 18:26:01.754324 139477768996416 checkpoint_manager.py:2112] [process=0][thread=save_finalize][step=9] CheckpointManager Save Finalize is done on all hosts. I0416 18:26:01.754814 139593135533184 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. I0416 18:26:01.755516 139593135533184 checkpoint_manager.py:1983] [process=0][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning. I0416 18:26:01.756357 139593135533184 metric_logger.py:185] completed step: 9, seconds: 0.122, TFLOP/s/device: 111.185, Tokens/s/device: 16758.999, total_weights: 12300, loss: 8.774 Per train step: Total TFLOPs: 13.59 split as 93.93% learnable weight flops and 6.07% attention flops
2026-04-16 19:48:05.982126: 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) I0416 19:48:06.099180 137864048299136 max_utils.py:238] Skipping jax distributed system due to skip_jax_distributed_system=True flag. I0416 19:49:05.985607 137864048299136 max_utils.py:800] System Information: Jax Version: 0.9.2 I0416 19:49:05.985717 137864048299136 max_utils.py:801] System Information: Jaxlib Version: 0.9.2 I0416 19:49:05.985751 137864048299136 max_utils.py:802] System Information: Jax Backend: PJRT C API TFRT TPU v6 lite Built on Apr 6 2026 20:48:10 (1775533690) cl/895581894 I0416 19:49:05.985775 137864048299136 train_utils.py:364] WARNING: 'dataset_path' might be pointing your local file system I0416 19:49:05.985857 137864048299136 train.py:811] [DECOUPLED NO-OP] skipping cloud diagnostics wrapper. W0416 19:49:06.078824 4059754 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. I0416 19:49:06.461632 137864048299136 maxtext_utils.py:1687] Num_devices: 8, shape (1, 1, 1, 8, 1, 1, 1, 1, 1, 1, 1, 1, 1) I0416 19:49:06.953729 137864048299136 checkpointing.py:688] Setting up checkpoint logger... I0416 19:49:06.953830 137864048299136 checkpointing.py:234] Creating checkpoint manager with ocdbt=True and zarr3=True I0416 19:49:06.953871 137864048299136 pytree_checkpoint_handler.py:577] save_device_host_concurrent_bytes=None I0416 19:49:06.954074 137864048299136 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 0x7d6262bf4c20>, enable_pinned_host_transfer=False, save_concurrent_bytes: 96000000000 (89.4 GiB), restore_concurrent_bytes: 96000000000 (89.4 GiB) I0416 19:49:09.349540 137864048299136 checkpointing.py:266] Enabling policy for fixed interval checkpointing. I0416 19:49:09.349732 137864048299136 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 0x7d5be59ee360>}, handler_registry=None I0416 19:49:09.350004 137864048299136 composite_checkpoint_handler.py:237] Deferred registration for item: "items". Adding handler `<orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7d5be59ee360>` 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`. I0416 19:49:09.350054 137864048299136 composite_checkpoint_handler.py:237] Deferred registration for item: "metrics". Adding handler `<orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7d5be59eef90>` 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`. I0416 19:49:09.350083 137864048299136 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 0x7d5be59ee360>, ('items', <class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeRestoreArgs'>): <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7d5be59ee360>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonSaveArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7d5be59eef90>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonRestoreArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7d5be59eef90>}). I0416 19:49:09.350606 137864048299136 abstract_checkpointer.py:35] orbax-checkpoint version: 0.11.28 I0416 19:49:09.350660 137864048299136 async_checkpointer.py:177] [process=0][thread=MainThread] Using barrier_sync_fn: <function get_barrier_sync_fn.<locals>.<lambda> at 0x7d5be551a660> timeout: 600 secs and primary_host=0 for async checkpoint writes I0416 19:49:09.477383 137864048299136 checkpoint_manager.py:1788] Found 0 checkpoint steps in gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints I0416 19:49:09.477617 137864048299136 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/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints: <orbax.checkpoint.checkpoint_manager.CheckpointManager object at 0x7d5be59ed3a0> I0416 19:49:09.477702 137864048299136 checkpointing.py:302] Checkpoint manager created! I0416 19:49:09.559836 137864048299136 dataset_info.py:707] Load dataset info from tests/assets/local_datasets/c4_en_dataset_minimal/c4/en/3.1.0 I0416 19:49:09.562842 137864048299136 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. I0416 19:49:09.617504 137864048299136 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 I0416 19:49:09.649197 137864048299136 tokenizer.py:245] Tokenizer path: src/maxtext/assets/tokenizers/tokenizer.llama2 I0416 19:49:09.649259 137864048299136 tokenizer.py:187] Loading sentencepiece tokenizer: src/maxtext/assets/tokenizers/tokenizer.llama2 I0416 19:49:11.032826 137864048299136 checkpointing.py:578] checkpoint manager exists so trying to load this run's existing checkpoint I0416 19:49:11.032951 137864048299136 checkpointing.py:676] No existing checkpoints found, not restoring checkpoint. W0416 19:49:11.489714 4059754 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. [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 I0416 19:49:11.953246 137864048299136 maxtext_utils.py:1805] decoder/decoder_norm/scale/value Shape: float32[2048] Physical: (None,) I0416 19:49:11.953346 137864048299136 maxtext_utils.py:1805] decoder/layers/mlp/wi_0/kernel/value Shape: float32[2048,16,7168] Physical: ('fsdp', None, None) I0416 19:49:11.953386 137864048299136 maxtext_utils.py:1805] decoder/layers/mlp/wi_1/kernel/value Shape: float32[2048,16,7168] Physical: ('fsdp', None, None) I0416 19:49:11.953418 137864048299136 maxtext_utils.py:1805] decoder/layers/mlp/wo/kernel/value Shape: float32[7168,16,2048] Physical: (None, None, 'fsdp') I0416 19:49:11.953445 137864048299136 maxtext_utils.py:1805] decoder/layers/post_self_attention_layer_norm/scale/value Shape: float32[2048,16] Physical: (None, None) I0416 19:49:11.953469 137864048299136 maxtext_utils.py:1805] decoder/layers/pre_self_attention_layer_norm/scale/value Shape: float32[2048,16] Physical: (None, None) I0416 19:49:11.953496 137864048299136 maxtext_utils.py:1805] decoder/layers/self_attention/key/kernel/value Shape: float32[2048,16,16,128] Physical: ('fsdp', None, None, None) I0416 19:49:11.953521 137864048299136 maxtext_utils.py:1805] decoder/layers/self_attention/out/kernel/value Shape: float32[16,16,128,2048] Physical: (None, None, None, 'fsdp') I0416 19:49:11.953545 137864048299136 maxtext_utils.py:1805] decoder/layers/self_attention/query/kernel/value Shape: float32[2048,16,16,128] Physical: ('fsdp', None, None, None) I0416 19:49:11.953568 137864048299136 maxtext_utils.py:1805] decoder/layers/self_attention/value/kernel/value Shape: float32[2048,16,16,128] Physical: ('fsdp', None, None, None) I0416 19:49:11.953591 137864048299136 maxtext_utils.py:1805] decoder/logits_dense/kernel/value Shape: float32[2048,32000] Physical: ('fsdp', None) I0416 19:49:11.953613 137864048299136 maxtext_utils.py:1805] token_embedder/embedding/value Shape: float32[32000,2048] Physical: (None, 'fsdp') I0416 19:49:12.089778 137864048299136 nnx_decoders.py:465] nnx_decoders/carry Logical: bfloat16[8,2048,2048]....................................... ('activation_batch', 'activation_norm_length', 'activation_embed'). I0416 19:49:12.089857 137864048299136 nnx_decoders.py:465] nnx_decoders/carry Physical: bfloat16[8,2048,2048]....................................... ('fsdp', None, None). I0416 19:49:12.094959 137864048299136 nnx_decoders.py:465] Unknown Logical: bfloat16[8,2048,2048]....................................... ('activation_batch', 'activation_norm_length', 'activation_embed'). I0416 19:49:12.095004 137864048299136 nnx_decoders.py:465] Unknown Physical: bfloat16[8,2048,2048]....................................... ('fsdp', None, None). I0416 19:49:12.110272 137864048299136 attentions.py:1088] attentions/inputs_q Logical: bfloat16[8,2048,2048]....................................... ('activation_batch', 'activation_attn_length', 'activation_attn_embed'). I0416 19:49:12.110323 137864048299136 attentions.py:1088] attentions/inputs_q Physical: bfloat16[8,2048,2048]....................................... ('fsdp', None, None). I0416 19:49:12.124968 137864048299136 attentions.py:1089] attentions/inputs_kv Logical: bfloat16[8,2048,2048]....................................... ('activation_batch', 'activation_attn_length', 'activation_attn_embed'). I0416 19:49:12.125017 137864048299136 attentions.py:1089] attentions/inputs_kv Physical: bfloat16[8,2048,2048]....................................... ('fsdp', None, None). I0416 19:49:12.195873 137864048299136 attentions.py:1154] attentions/query Logical: bfloat16[8,2048,16,128]..................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim'). I0416 19:49:12.195952 137864048299136 attentions.py:1154] attentions/query Physical: bfloat16[8,2048,16,128]..................................... ('fsdp', None, None, None). I0416 19:49:12.210627 137864048299136 attentions.py:1155] attentions/key Logical: bfloat16[8,2048,16,128]..................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim'). I0416 19:49:12.210677 137864048299136 attentions.py:1155] attentions/key Physical: bfloat16[8,2048,16,128]..................................... ('fsdp', None, None, None). I0416 19:49:12.225358 137864048299136 attentions.py:1156] attentions/value Logical: bfloat16[8,2048,16,128]..................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim'). I0416 19:49:12.225408 137864048299136 attentions.py:1156] attentions/value Physical: bfloat16[8,2048,16,128]..................................... ('fsdp', None, None, None). I0416 19:49:12.253231 137864048299136 attentions.py:1197] attentions/out Logical: bfloat16[8,2048,16,128]..................................... ('activation_batch', 'activation_attn_length', 'activation_heads', 'activation_kv'). I0416 19:49:12.253288 137864048299136 attentions.py:1197] attentions/out Physical: bfloat16[8,2048,16,128]..................................... ('fsdp', None, None, None). I0416 19:49:12.323709 137864048299136 linears.py:525] linears/x Logical: bfloat16[8,2048,7168]....................................... ('activation_batch', 'activation_length', 'activation_mlp'). I0416 19:49:12.323782 137864048299136 linears.py:525] linears/x Physical: bfloat16[8,2048,7168]....................................... ('fsdp', None, None). W0416 19:49:13.555309 4059754 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. I0416 19:49:13.769993 137864048299136 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. I0416 19:49:13.770630 137864048299136 max_utils.py:194] tensorboardX not available; using no-op SummaryWriter. I0416 19:49:13.772891 137864048299136 metric_logger.py:289] number parameters: 1.104 billion W0416 19:49:17.169590 4059754 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. I0416 19:49:17.409847 137864048299136 checkpointing.py:794] Waiting for step 0 to finish before checkpoint... I0416 19:49:17.533704 137864048299136 checkpointing.py:798] Waited 0.12385797500610352 seconds for step 0 to finish before starting checkpointing. I0416 19:49:17.534249 137864048299136 checkpoint_manager.py:1983] [process=0][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning. I0416 19:49:17.534433 137864048299136 checkpoint_manager.py:1501] [process=0] Saving checkpoint at step 0 I0416 19:49:17.534945 137864048299136 async_checkpointer.py:452] [process=0] Started async saving checkpoint to gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0. I0416 19:49:17.619491 137864048299136 signaling_client.py:373] Using ThreadSafeKeyValueSignalingClient I0416 19:49:17.706867 137750435071552 atomicity.py:137] Creating tmp directory gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0 I0416 19:49:17.751106 137864048299136 jax_array_handlers.py:347] Scheduling D2H of 111 prioritized jax.Array. I0416 19:49:17.751196 137864048299136 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 W0416 19:49:18.429017 4059754 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. W0416 19:49:18.446375 4059754 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. I0416 19:49:18.447175 137750424585792 atomicity.py:137] Creating tmp directory gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0/items W0416 19:49:18.458516 4059754 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. W0416 19:49:18.465497 4059754 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. W0416 19:49:18.470038 4059754 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. W0416 19:49:18.474151 4059754 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. W0416 19:49:18.478366 4059754 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. W0416 19:49:18.482435 4059754 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. I0416 19:49:18.603126 137750380545600 checkpoint.py:188] Wrote Metadata={'item_handlers': None, 'metrics': {}, 'performance_metrics': {}, 'init_timestamp_nsecs': 1776368958332265679, 'commit_timestamp_nsecs': None, 'custom_metadata': {}}, json={"item_handlers": null, "metrics": {}, "performance_metrics": {}, "init_timestamp_nsecs": 1776368958332265679, "commit_timestamp_nsecs": null, "custom_metadata": {}} to gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0/_CHECKPOINT_METADATA I0416 19:49:19.051524 4062555 google_auth_provider.cc:149] Using credentials at ~/.config/gcloud/application_default_credentials.json I0416 19:49:19.051578 4062555 google_auth_provider.cc:156] Using OAuth2 AuthProvider I0416 19:49:19.103745 137864048299136 base_pytree_checkpoint_handler.py:153] [process=0][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 1.354311s I0416 19:49:19.112649 137864048299136 base_pytree_checkpoint_handler.py:128] [process=0] /jax/checkpoint/write/blocking_gbytes_per_sec: 8.271 GiB/s (total gbytes: 12.3 GiB) (time elapsed: a second) (per-host) I0416 19:49:19.112720 137864048299136 base_pytree_checkpoint_handler.py:732] [process=0][thread=MainThread] Initiated Pytree async_save. Time taken: 1.492244s (batch_requests_ready=0.114531s, total_serialization_initiated=1.368901s, others=0.008813s) I0416 19:49:19.112808 137864048299136 composite_checkpoint_handler.py:715] [process=0][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 1.492832s (all_items=0.000027s, per_item={'items': '0.00002742'}, temp_paths=1.492804) I0416 19:49:19.114106 137750296659520 async_checkpointer.py:79] [process=0][thread=async_save] Background save thread started. I0416 19:49:19.114189 137864048299136 async_checkpointer.py:561] Finished blocking save. Time taken: 1.579711s. Continuing background save to gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0. I0416 19:49:19.114376 137864048299136 checkpoint_manager.py:1549] [process=0][thread=MainThread][step=0] Starting CheckpointManager Save Finalize thread=save_finalize I0416 19:49:19.114620 137750445557312 async_checkpointer.py:265] [process=0][thread=save_finalize] Waiting for background save thread=async_save. I0416 19:49:19.114740 137864048299136 standard_logger.py:34] {'step': 0, 'event_type': 'save', 'directory': 'gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776368957.5342345, 'wait_for_prev_duration_secs': 4.38690185546875e-05, 'time_between_consecutive_saves_sec': None, 'checkpointer_blocking_start_time': 1776368957.5344532, 'checkpointer_blocking_duration_secs': 1.579840898513794, 'get_old_steps_start_time': 1776368959.1143122, 'get_old_steps_duration_secs': 2.8133392333984375e-05, 'checkpoint_manager_blocking_start_time': 1776368957.5341334, 'checkpoint_manager_blocking_duration_secs': 1.5805799961090088} I0416 19:49:19.114824 137864048299136 checkpointing.py:409] Started an asynchronous checkpoint save for step 0 I0416 19:49:19.114879 137864048299136 max_utils.py:750] Memstats: After params initialized: I0416 19:49:19.114929 137864048299136 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_0(process=0,(0,0,0,0)) I0416 19:49:19.114955 137864048299136 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_1(process=0,(1,0,0,0)) I0416 19:49:19.114976 137864048299136 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_2(process=0,(0,1,0,0)) I0416 19:49:19.114996 137864048299136 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_3(process=0,(1,1,0,0)) I0416 19:49:19.115015 137864048299136 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_4(process=0,(0,2,0,0)) I0416 19:49:19.115032 137864048299136 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_5(process=0,(1,2,0,0)) I0416 19:49:19.115065 137864048299136 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_6(process=0,(0,3,0,0)) I0416 19:49:19.115084 137864048299136 max_utils.py:756] Using (GB) 1.6 / 31.25 (5.120000%) on TPU_7(process=0,(1,3,0,0)) W0416 19:49:19.120476 4059754 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. W0416 19:49:19.126074 4059754 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. W0416 19:49:19.130768 4059754 pjrt_executable.cc:642] Assume version compatibility. PjRt-IFRT does not track XLA executable versions. I0416 19:49:19.455177 137864048299136 metric_logger.py:185] completed step: 0, seconds: 3.635, TFLOP/s/device: 3.738, Tokens/s/device: 563.408, total_weights: 13328, loss: 10.880 I0416 19:49:19.456172 137864048299136 metric_logger.py:269] To see full metrics 'tensorboard --logdir=gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/tensorboard/' I0416 19:49:19.571270 137864048299136 metric_logger.py:185] completed step: 1, seconds: 2.040, TFLOP/s/device: 6.662, Tokens/s/device: 1004.100, total_weights: 12332, loss: 10.851 I0416 19:49:19.691782 137864048299136 metric_logger.py:185] completed step: 2, seconds: 0.033, TFLOP/s/device: 414.305, Tokens/s/device: 62448.544, total_weights: 15161, loss: 9.862 I0416 19:49:19.813416 137864048299136 metric_logger.py:185] completed step: 3, seconds: 0.102, TFLOP/s/device: 133.842, Tokens/s/device: 20174.160, total_weights: 13327, loss: 9.404 I0416 19:49:20.280575 137750328116800 array_metadata_store.py:203] [process=0][thread=array_type_handler] Wrote 111 array_metadata.ArrayMetadata to gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0/items/array_metadatas/process_0 I0416 19:49:40.106095 137750317631040 base_pytree_checkpoint_handler.py:1217] [process=0][thread=write_metadata_after_commits] Commit + Array metadata written. Time taken: 20.992763s (commit=20.524703s, array_metadata_write=0.468060s) I0416 19:49:40.107110 137750296659520 base_pytree_checkpoint_handler.py:128] [process=0] /jax/checkpoint/write/gbytes_per_sec: 561.970 MiB/s (total gbytes: 12.3 GiB) (time elapsed: 22 seconds) (per-host) I0416 19:49:40.107173 137750296659520 async_checkpointer.py:90] [process=0][thread=async_save] 3 Handler Commit operations completed. Time taken: 20.992936s. I0416 19:49:40.310816 137750296659520 checkpoint.py:228] Read Metadata={'item_handlers': None, 'metrics': {}, 'performance_metrics': {}, 'init_timestamp_nsecs': 1776368958332265679, 'commit_timestamp_nsecs': None, 'custom_metadata': {}} from gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0/_CHECKPOINT_METADATA I0416 19:49:40.500457 137750296659520 array_metadata_store.py:367] [process=0][thread=async_save] Skipped cross-host ArrayMetadata validation because only one process is found: process_index=0. I0416 19:49:40.690468 137750380545600 checkpoint.py:247] Updated Metadata={'item_handlers': {'items': 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler'}, 'metrics': {}, 'performance_metrics': {}, 'init_timestamp_nsecs': 1776368958332265679, 'commit_timestamp_nsecs': None, 'custom_metadata': {}} to gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0/_CHECKPOINT_METADATA I0416 19:49:40.923253 137750296659520 ocdbt_utils.py:56] Param validation support for Zarr3 will be added later (b/362328389). I0416 19:49:40.923895 137750296659520 base_pytree_checkpoint_handler.py:1342] [process=0][thread=async_save] Pytree save finalize (merge_ocdbt + ArrayMetadata validation) completed. Time taken: 0.568908s. use_zarr3=True, enable_post_merge_validation=True, directory=gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0/items I0416 19:49:40.924601 137750296659520 atomicity.py:608] Finalizing gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0/items I0416 19:49:41.177473 137750296659520 atomicity.py:608] Finalizing gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0 I0416 19:49:41.875277 137750296659520 atomicity.py:794] [process=0][thread=async_save] Finished saving checkpoint (finalized tmp dir) to `gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0`. I0416 19:49:41.875910 137750296659520 async_checkpointer.py:420] Finished async_save (blocking + background). Time taken: 24.341439s. directory=gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/0 I0416 19:49:41.875997 137750296659520 async_checkpointer.py:144] [process=0][thread=async_save] Background save thread done. Time taken: 22.761764s. I0416 19:49:41.876175 137750445557312 async_checkpointer.py:273] [process=0][thread=save_finalize] Done with waiting for background save thread=async_save. I0416 19:49:41.876284 137750445557312 async_checkpointer.py:283] [process=0][thread=save_finalize] No errors found in background save thread=async_save. I0416 19:49:41.876347 137750445557312 checkpoint_manager.py:2103] [process=0][thread=save_finalize][step=0] CheckpointManager Save Finalize is syncing with other hosts... I0416 19:49:41.876394 137750445557312 checkpoint_manager.py:2112] [process=0][thread=save_finalize][step=0] CheckpointManager Save Finalize is done on all hosts. I0416 19:49:45.736580 137864048299136 metric_logger.py:185] completed step: 4, seconds: 0.119, TFLOP/s/device: 114.144, Tokens/s/device: 17205.024, total_weights: 11939, loss: 9.133 I0416 19:49:45.748149 137864048299136 metric_logger.py:185] completed step: 5, seconds: 0.122, TFLOP/s/device: 111.698, Tokens/s/device: 16836.290, total_weights: 15502, loss: 9.029 I0416 19:49:45.865566 137864048299136 metric_logger.py:185] completed step: 6, seconds: 25.922, TFLOP/s/device: 0.524, Tokens/s/device: 79.007, total_weights: 13864, loss: 8.894 I0416 19:49:45.987192 137864048299136 metric_logger.py:185] completed step: 7, seconds: 0.009, TFLOP/s/device: 1448.985, Tokens/s/device: 218406.740, total_weights: 12988, loss: 8.806 I0416 19:49:46.109101 137864048299136 metric_logger.py:185] completed step: 8, seconds: 0.119, TFLOP/s/device: 114.625, Tokens/s/device: 17277.597, total_weights: 13820, loss: 8.803 I0416 19:49:46.232643 137864048299136 checkpointing.py:794] Waiting for step 9 to finish before checkpoint... I0416 19:49:46.234463 137864048299136 checkpointing.py:798] Waited 0.0018422603607177734 seconds for step 9 to finish before starting checkpointing. I0416 19:49:46.234910 137864048299136 checkpoint_manager.py:1983] [process=0][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning. I0416 19:49:46.235077 137864048299136 checkpoint_manager.py:1501] [process=0] Saving checkpoint at step 9 I0416 19:49:46.235351 137864048299136 async_checkpointer.py:452] [process=0] Started async saving checkpoint to gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/9. I0416 19:49:46.418251 137750349088320 atomicity.py:137] Creating tmp directory gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/9 I0416 19:49:46.418603 137864048299136 jax_array_handlers.py:347] Scheduling D2H of 111 prioritized jax.Array. I0416 19:49:46.418829 137864048299136 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 I0416 19:49:47.114248 137749094991424 atomicity.py:137] Creating tmp directory gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/9/items I0416 19:49:53.198852 137864048299136 base_pytree_checkpoint_handler.py:153] [process=0][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 6.782248s I0416 19:49:53.205115 137864048299136 base_pytree_checkpoint_handler.py:128] [process=0] /jax/checkpoint/write/blocking_gbytes_per_sec: 1.792 GiB/s (total gbytes: 12.3 GiB) (time elapsed: 6 seconds) (per-host) I0416 19:49:53.205189 137864048299136 base_pytree_checkpoint_handler.py:732] [process=0][thread=MainThread] Initiated Pytree async_save. Time taken: 6.888149s (batch_requests_ready=0.085847s, total_serialization_initiated=6.796152s, others=0.006149s) I0416 19:49:53.205277 137864048299136 composite_checkpoint_handler.py:715] [process=0][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 6.888696s (all_items=0.000016s, per_item={'items': '0.00001574'}, temp_paths=6.888681) I0416 19:49:53.209594 137750349088320 async_checkpointer.py:79] [process=0][thread=async_save] Background save thread started. I0416 19:49:53.209687 137864048299136 async_checkpointer.py:561] Finished blocking save. Time taken: 6.974565s. Continuing background save to gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/9. I0416 19:49:53.209875 137864048299136 checkpoint_manager.py:1549] [process=0][thread=MainThread][step=9] Starting CheckpointManager Save Finalize thread=save_finalize I0416 19:49:53.210063 137749881423424 async_checkpointer.py:265] [process=0][thread=save_finalize] Waiting for background save thread=async_save. I0416 19:49:53.210176 137864048299136 standard_logger.py:34] {'step': 9, 'event_type': 'save', 'directory': 'gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776368986.2348704, 'wait_for_prev_duration_secs': 6.747245788574219e-05, 'time_between_consecutive_saves_sec': 4.358460903167725, 'checkpointer_blocking_start_time': 1776368986.2351015, 'checkpointer_blocking_duration_secs': 6.974695920944214, 'get_old_steps_start_time': 1776368993.2098114, 'get_old_steps_duration_secs': 2.574920654296875e-05, 'checkpoint_manager_blocking_start_time': 1776368986.2348342, 'checkpoint_manager_blocking_duration_secs': 6.975313425064087} I0416 19:49:53.210296 137864048299136 checkpointing.py:409] Started an asynchronous checkpoint save for step 9 I0416 19:49:53.210332 137864048299136 checkpoint_manager.py:1994] [process=0][thread=MainThread][step=9][wait_until_finished] Waiting for Save Finalize thread (save_finalize) to complete. I0416 19:49:53.982576 137750445557312 array_metadata_store.py:203] [process=0][thread=array_type_handler] Wrote 111 array_metadata.ArrayMetadata to gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/9/items/array_metadatas/process_0 I0416 19:50:28.867222 137749094991424 base_pytree_checkpoint_handler.py:1217] [process=0][thread=write_metadata_after_commits] Commit + Array metadata written. Time taken: 35.658358s (commit=35.251570s, array_metadata_write=0.406789s) I0416 19:50:28.868429 137750349088320 base_pytree_checkpoint_handler.py:128] [process=0] /jax/checkpoint/write/gbytes_per_sec: 296.977 MiB/s (total gbytes: 12.3 GiB) (time elapsed: 42 seconds) (per-host) I0416 19:50:28.868497 137750349088320 async_checkpointer.py:90] [process=0][thread=async_save] 3 Handler Commit operations completed. Time taken: 35.658765s. I0416 19:50:29.327667 137750349088320 array_metadata_store.py:367] [process=0][thread=async_save] Skipped cross-host ArrayMetadata validation because only one process is found: process_index=0. I0416 19:50:29.779597 137750349088320 ocdbt_utils.py:56] Param validation support for Zarr3 will be added later (b/362328389). I0416 19:50:29.780575 137750349088320 base_pytree_checkpoint_handler.py:1342] [process=0][thread=async_save] Pytree save finalize (merge_ocdbt + ArrayMetadata validation) completed. Time taken: 0.595701s. use_zarr3=True, enable_post_merge_validation=True, directory=gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/9/items I0416 19:50:29.781187 137750349088320 atomicity.py:608] Finalizing gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/9/items I0416 19:50:30.017683 137750349088320 atomicity.py:608] Finalizing gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/9 I0416 19:50:30.725369 137750349088320 atomicity.py:794] [process=0][thread=async_save] Finished saving checkpoint (finalized tmp dir) to `gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/9`. I0416 19:50:30.725998 137750349088320 async_checkpointer.py:420] Finished async_save (blocking + background). Time taken: 44.490884s. directory=gs://wanglance-maxtext/nnx_ckpt_feat_nnx_post_train_fixes_20260416_181521/nnx_feat_nnx_post_train_fixes_20260416_181521_04_int8/checkpoints/9 I0416 19:50:30.726088 137750349088320 async_checkpointer.py:144] [process=0][thread=async_save] Background save thread done. Time taken: 37.516357s. I0416 19:50:30.726297 137749881423424 async_checkpointer.py:273] [process=0][thread=save_finalize] Done with waiting for background save thread=async_save. I0416 19:50:30.726402 137749881423424 async_checkpointer.py:283] [process=0][thread=save_finalize] No errors found in background save thread=async_save. I0416 19:50:30.726463 137749881423424 checkpoint_manager.py:2103] [process=0][thread=save_finalize][step=9] CheckpointManager Save Finalize is syncing with other hosts... I0416 19:50:30.726508 137749881423424 checkpoint_manager.py:2112] [process=0][thread=save_finalize][step=9] CheckpointManager Save Finalize is done on all hosts. I0416 19:50:30.727266 137864048299136 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. I0416 19:50:30.727466 137864048299136 checkpoint_manager.py:1983] [process=0][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning. I0416 19:50:30.728855 137864048299136 metric_logger.py:185] completed step: 9, seconds: 0.122, TFLOP/s/device: 111.715, Tokens/s/device: 16838.920, total_weights: 12300, loss: 8.755 Per train step: Total TFLOPs: 13.59 split as 93.93% learnable weight flops and 6.07% attention flops