main (574ad3fb9) vs feat/nnx-trainstate-and-training-loop (574ad3fb9)| Metric | main 574ad3fb9 | feat/nnx-trainstate-and-training-loop 574ad3fb9 | Diff (feat/nnx-trainstate-and-training-loop − main) |
|---|---|---|---|
| Parameters | 1.104 billion | 1.104 billion | — |
| Final loss | 9.4570 | 9.4570 | 0 |
| TFLOP/s | 93.968 | 93.971 | +0.003 |
| Tok/s | 14163.8 | 14164.3 | +0.49 |
| Avg s/step | 6.889 | 6.923 | +0.034 |
| Memory % | 1.38 | 1.38 | 0 |
| JAX | 0.9.2 | 0.9.2 | — |
Diff = branch value − main value. Green = branch improved. Red = branch regressed.
XPK Start: Sun Apr 19 18:25:28 UTC 2026 PyTorch was not found. Models won't be available and only tokenizers, configuration and file/data utilities can be used. 2026-04-19 18:25:52.615629: 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) I0419 18:25:52.797857 132980130875200 max_utils.py:273] Attempting to initialize the jax distributed system... I0419 18:26:01.839979 132980130875200 distributed.py:149] Starting JAX distributed service on [::]:8482 I0419 18:26:01.842478 132980130875200 distributed.py:172] Connecting to JAX distributed service on mt-06-grad-accum-u8ona-slice-job-0-0.mt-06-grad-accum-u8ona:8482 I0419 18:26:03.028610 132980130875200 max_utils.py:284] Jax distributed system initialized! I0419 18:26:09.131906 132980130875200 max_utils.py:800] System Information: Jax Version: 0.9.2 I0419 18:26:09.132016 132980130875200 max_utils.py:801] System Information: Jaxlib Version: 0.9.2 I0419 18:26:09.132056 132980130875200 max_utils.py:802] System Information: Jax Backend: PJRT C API TFRT TPU v6 lite Built on Mar 4 2026 11:32:08 (1772652728) cl/878335365 I0419 18:26:09.132090 132980130875200 train_utils.py:348] WARNING: Sequence packing is essentially ignored for synthetic data. Please use a real dataset to use sequence packing. I0419 18:26:09.827729 132980130875200 maxtext_utils.py:1551] Num_devices: 32, shape (1, 1, 1, 32, 1, 1, 1, 1, 1, 1, 1, 1, 1) I0419 18:26:09.828000 132980130875200 checkpointing.py:677] Setting up checkpoint logger... I0419 18:26:09.828054 132980130875200 checkpointing.py:233] Creating checkpoint manager with ocdbt=True and zarr3=True I0419 18:26:09.828098 132980130875200 pytree_checkpoint_handler.py:592] save_device_host_concurrent_bytes=None I0419 18:26:09.828442 132980130875200 base_pytree_checkpoint_handler.py:441] 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 0x78f13df18380>, enable_pinned_host_transfer=False, save_concurrent_bytes: 96000000000 (89.4 GiB), restore_concurrent_bytes: 96000000000 (89.4 GiB) I0419 18:26:13.636326 132980130875200 checkpointing.py:265] Enabling policy for fixed interval checkpointing. I0419 18:26:13.636695 132980130875200 checkpoint_manager.py:708] [process=6][thread=MainThread] CheckpointManager init: checkpointers=None, item_names=('items',), item_handlers={'items': <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x78f015e9c080>}, handler_registry=None I0419 18:26:13.636933 132980130875200 composite_checkpoint_handler.py:237] Deferred registration for item: "items". Adding handler `<orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x78f015e9c080>` 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`. I0419 18:26:13.636980 132980130875200 composite_checkpoint_handler.py:237] Deferred registration for item: "metrics". Adding handler `<orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x78dd003ceab0>` 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`. I0419 18:26:13.637015 132980130875200 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 0x78f015e9c080>, ('items', <class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeRestoreArgs'>): <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x78f015e9c080>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonSaveArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x78dd003ceab0>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonRestoreArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x78dd003ceab0>}). I0419 18:26:13.637331 132980130875200 abstract_checkpointer.py:35] orbax-checkpoint version: 0.11.34 I0419 18:26:13.637399 132980130875200 async_checkpointer.py:192] [process=6][thread=MainThread] Using barrier_sync_fn: <function get_barrier_sync_fn.<locals>._fn at 0x78dcd07d7c40> timeout: 1200 secs and primary_host=0 for async checkpoint writes I0419 18:26:14.345484 132980130875200 checkpoint_manager.py:1812] Found 0 checkpoint steps in gs://lance-maxtext/linen_ckpt_xpk_main_20260419_180002/linen_xpk_main_20260419_180002_06_grad_accum/checkpoints I0419 18:26:14.761777 132980130875200 checkpoint_manager.py:929] [process=6][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_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, lightweight_initialize=False), root_directory=gs://lance-maxtext/linen_ckpt_xpk_main_20260419_180002/linen_xpk_main_20260419_180002_06_grad_accum/checkpoints: <orbax.checkpoint.checkpoint_manager.CheckpointManager object at 0x78f0155faed0> I0419 18:26:14.761947 132980130875200 checkpointing.py:301] Checkpoint manager created! I0419 18:26:16.477314 132980130875200 nnx_wrappers.py:437] Unknown Logical: bfloat16[32,2048,2048]...................................... ('activation_batch', 'activation_norm_length', 'activation_embed'). I0419 18:26:16.477450 132980130875200 nnx_wrappers.py:437] Unknown Physical: bfloat16[32,2048,2048]...................................... ('fsdp', None, None). I0419 18:26:16.855048 132980130875200 attentions.py:1088] attentions/inputs_q Logical: bfloat16[32,2048,2048]...................................... ('activation_batch', 'activation_attn_length', 'activation_attn_embed'). I0419 18:26:16.855146 132980130875200 attentions.py:1088] attentions/inputs_q Physical: bfloat16[32,2048,2048]...................................... ('fsdp', None, None). I0419 18:26:16.871610 132980130875200 attentions.py:1089] attentions/inputs_kv Logical: bfloat16[32,2048,2048]...................................... ('activation_batch', 'activation_attn_length', 'activation_attn_embed'). I0419 18:26:16.871672 132980130875200 attentions.py:1089] attentions/inputs_kv Physical: bfloat16[32,2048,2048]...................................... ('fsdp', None, None). I0419 18:26:16.895489 132980130875200 attentions.py:1154] attentions/query Logical: bfloat16[32,2048,16,128].................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim'). I0419 18:26:16.895560 132980130875200 attentions.py:1154] attentions/query Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None). I0419 18:26:16.912130 132980130875200 attentions.py:1155] attentions/key Logical: bfloat16[32,2048,16,128].................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim'). I0419 18:26:16.912194 132980130875200 attentions.py:1155] attentions/key Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None). I0419 18:26:16.928604 132980130875200 attentions.py:1156] attentions/value Logical: bfloat16[32,2048,16,128].................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim'). I0419 18:26:16.928667 132980130875200 attentions.py:1156] attentions/value Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None). I0419 18:26:16.953328 132980130875200 attentions.py:1197] attentions/out Logical: bfloat16[32,2048,16,128].................................... ('activation_batch', 'activation_attn_length', 'activation_heads', 'activation_kv'). I0419 18:26:16.953399 132980130875200 attentions.py:1197] attentions/out Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None). I0419 18:26:16.974437 132980130875200 linears.py:525] linears/x Logical: bfloat16[32,2048,7168]...................................... ('activation_batch', 'activation_length', 'activation_mlp'). I0419 18:26:16.974497 132980130875200 linears.py:525] linears/x Physical: bfloat16[32,2048,7168]...................................... ('fsdp', None, None). I0419 18:26:17.197075 132980130875200 checkpointing.py:577] checkpoint manager exists so trying to load this run's existing checkpoint I0419 18:26:17.197187 132980130875200 checkpointing.py:665] No existing checkpoints found, not restoring checkpoint. fsdp: 32 I0419 18:26:18.620092 132980130875200 maxtext_utils.py:1654] params/params/decoder/decoder_norm/scale Shape: float32[2048] Logical: P('norm',) Physical: (None,) I0419 18:26:18.620259 132980130875200 maxtext_utils.py:1654] params/params/decoder/layers/mlp/wi_0/kernel Shape: float32[2048,16,7168] Logical: P('embed', 'layers', 'mlp') Physical: ('fsdp', None, None) I0419 18:26:18.620324 132980130875200 maxtext_utils.py:1654] params/params/decoder/layers/mlp/wi_1/kernel Shape: float32[2048,16,7168] Logical: P('embed', 'layers', 'mlp') Physical: ('fsdp', None, None) I0419 18:26:18.620426 132980130875200 maxtext_utils.py:1654] params/params/decoder/layers/mlp/wo/kernel Shape: float32[7168,16,2048] Logical: P('mlp', 'layers', 'embed') Physical: (None, None, 'fsdp') I0419 18:26:18.620504 132980130875200 maxtext_utils.py:1654] params/params/decoder/layers/post_self_attention_layer_norm/scale Shape: float32[2048,16] Logical: P('norm', 'layers') Physical: (None, None) I0419 18:26:18.620573 132980130875200 maxtext_utils.py:1654] params/params/decoder/layers/pre_self_attention_layer_norm/scale Shape: float32[2048,16] Logical: P('norm', 'layers') Physical: (None, None) I0419 18:26:18.620664 132980130875200 maxtext_utils.py:1654] 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) I0419 18:26:18.620751 132980130875200 maxtext_utils.py:1654] 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') I0419 18:26:18.620809 132980130875200 maxtext_utils.py:1654] 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) I0419 18:26:18.620849 132980130875200 maxtext_utils.py:1654] 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) I0419 18:26:18.620903 132980130875200 maxtext_utils.py:1654] params/params/decoder/logits_dense/kernel Shape: float32[2048,32000] Logical: P('embed_vocab', 'vocab') Physical: ('fsdp', None) I0419 18:26:18.620983 132980130875200 maxtext_utils.py:1654] params/params/token_embedder/embedding Shape: float32[32000,2048] Logical: P('vocab', 'embed_vocab') Physical: (None, 'fsdp') I0419 18:26:18.644057 132980130875200 gradient_accumulation.py:68] gradient_accumulation/inputs Logical: float32[2048]............................................... Unknown. I0419 18:26:18.644119 132980130875200 gradient_accumulation.py:68] gradient_accumulation/inputs Physical: float32[2048]............................................... (None,). I0419 18:26:18.659997 132980130875200 gradient_accumulation.py:68] gradient_accumulation/inputs Logical: float32[2048,16,7168]....................................... Unknown. I0419 18:26:18.660053 132980130875200 gradient_accumulation.py:68] gradient_accumulation/inputs Physical: float32[2048,16,7168]....................................... ('fsdp', None, None). I0419 18:26:18.693085 132980130875200 gradient_accumulation.py:68] gradient_accumulation/inputs Logical: float32[7168,16,2048]....................................... Unknown. I0419 18:26:18.693202 132980130875200 gradient_accumulation.py:68] gradient_accumulation/inputs Physical: float32[7168,16,2048]....................................... (None, None, 'fsdp'). I0419 18:26:18.710482 132980130875200 gradient_accumulation.py:68] gradient_accumulation/inputs Logical: float32[2048,16]............................................ Unknown. I0419 18:26:18.710540 132980130875200 gradient_accumulation.py:68] gradient_accumulation/inputs Physical: float32[2048,16]............................................ (None, None). I0419 18:26:18.739964 132980130875200 gradient_accumulation.py:68] gradient_accumulation/inputs Logical: float32[2048,16,16,128]..................................... Unknown. I0419 18:26:18.740026 132980130875200 gradient_accumulation.py:68] gradient_accumulation/inputs Physical: float32[2048,16,16,128]..................................... ('fsdp', None, None, None). I0419 18:26:18.754849 132980130875200 gradient_accumulation.py:68] gradient_accumulation/inputs Logical: float32[16,16,128,2048]..................................... Unknown. I0419 18:26:18.754904 132980130875200 gradient_accumulation.py:68] gradient_accumulation/inputs Physical: float32[16,16,128,2048]..................................... (None, None, None, 'fsdp'). I0419 18:26:18.798887 132980130875200 gradient_accumulation.py:68] gradient_accumulation/inputs Logical: float32[2048,32000]......................................... Unknown. I0419 18:26:18.798961 132980130875200 gradient_accumulation.py:68] gradient_accumulation/inputs Physical: float32[2048,32000]......................................... ('fsdp', None). I0419 18:26:18.813732 132980130875200 gradient_accumulation.py:68] gradient_accumulation/inputs Logical: float32[32000,2048]......................................... Unknown. I0419 18:26:18.813789 132980130875200 gradient_accumulation.py:68] gradient_accumulation/inputs Physical: float32[32000,2048]......................................... (None, 'fsdp'). I0419 18:26:19.485064 132980130875200 train.py:155] train/xent Logical: float32[32,2048]............................................ ('activation_embed_and_logits_batch', 'activation_length'). I0419 18:26:19.485165 132980130875200 train.py:155] train/xent Physical: float32[32,2048]............................................ ('fsdp', None). I0419 18:26:19.501015 132980130875200 train.py:162] train/z_loss Logical: float32[32,2048]............................................ ('activation_embed_and_logits_batch', 'activation_length'). I0419 18:26:19.501074 132980130875200 train.py:162] train/z_loss Physical: float32[32,2048]............................................ ('fsdp', None). I0419 18:26:30.911311 132980130875200 max_utils.py:791] Total memory size: 1.7 GB, Output size: 0.4 GB, Temp size: 1.3 GB, Argument size: 0.4 GB, Host temp size: 0.0 GB. I0419 18:26:30.912202 132980130875200 metric_logger.py:301] number parameters: 1.104 billion I0419 18:26:32.769883 132980130875200 checkpointing.py:772] Waiting for step 0 to finish before checkpoint... I0419 18:26:43.396878 132980130875200 checkpointing.py:776] Waited 10.626973390579224 seconds for step 0 to finish before starting checkpointing. I0419 18:26:43.399330 132980130875200 checkpoint_manager.py:2009] [process=6][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning. I0419 18:26:43.401144 132980130875200 checkpoint_manager.py:1512] [process=6] Saving checkpoint at step 0 I0419 18:26:43.402567 132980130875200 event_tracking.py:70] [process=6] [async] Started save checkpoint @ gs://lance-maxtext/linen_ckpt_xpk_main_20260419_180002/linen_xpk_main_20260419_180002_06_grad_accum/checkpoints/0. I0419 18:26:44.144714 132980130875200 signaling_client.py:364] Using JaxDistributedSignalingClient I0419 18:26:44.145643 132980130875200 jax_array_handlers.py:360] Scheduling D2H of 39 prioritized jax.Array. I0419 18:26:44.145697 132980130875200 replica_slices.py:424] 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 I0419 18:26:44.419607 132980130875200 base_pytree_checkpoint_handler.py:154] [process=6][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 0.274983s I0419 18:26:44.419784 132980130875200 base_pytree_checkpoint_handler.py:130] [process=6] /jax/orbax/write/blocking_gbytes_per_sec: 5.492 GiB/s (total gbytes: 1.5 GiB) (time elapsed: 0.2808678150177002 s) (per-host) I0419 18:26:44.419836 132980130875200 base_pytree_checkpoint_handler.py:768] [process=6][thread=MainThread] Initiated Pytree async_save. Time taken: 0.280932s (batch_requests_ready=0.002651s, total_serialization_initiated=0.278208s, others=0.000073s) I0419 18:26:44.419926 132980130875200 composite_checkpoint_handler.py:715] [process=6][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 0.285139s (all_items=0.000019s, per_item={'items': '0.00001860'}, temp_paths=0.285121) I0419 18:26:44.420747 132980130875200 event_tracking.py:125] [process=6] [async] Finished blocking save in 1.02 seconds. Continuing save @ gs://lance-maxtext/linen_ckpt_xpk_main_20260419_180002/linen_xpk_main_20260419_180002_06_grad_accum/checkpoints/0. I0419 18:26:44.421082 132851928397568 async_checkpointer.py:76] [process=6][thread=async_save] Background save thread started. Deadline for this save operation is 2026-04-19 18:46:44.421044 I0419 18:26:44.430727 132980130875200 checkpoint_manager.py:1560] [process=6][thread=MainThread][step=0] Starting CheckpointManager Save Finalize thread=save_finalize I0419 18:26:44.431041 132851873855232 async_checkpointer.py:280] [process=6][thread=save_finalize] Waiting for background save thread=async_save. I0419 18:26:44.431200 132980130875200 standard_logger.py:34] {'step': 0, 'event_type': 'save', 'directory': 'gs://lance-maxtext/linen_ckpt_xpk_main_20260419_180002/linen_xpk_main_20260419_180002_06_grad_accum/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776623203.3993113, 'wait_for_prev_duration_secs': 6.079673767089844e-05, 'time_between_consecutive_saves_sec': None, 'checkpointer_blocking_start_time': 1776623203.401182, 'checkpointer_blocking_duration_secs': 1.020052194595337, 'get_old_steps_start_time': 1776623204.4212599, 'get_old_steps_duration_secs': 3.0517578125e-05, 'checkpoint_manager_blocking_start_time': 1776623203.3974037, 'checkpoint_manager_blocking_duration_secs': 1.0337538719177246} I0419 18:26:44.431302 132980130875200 checkpointing.py:408] Started an asynchronous checkpoint save for step 0 I0419 18:26:44.431351 132980130875200 max_utils.py:750] Memstats: After params initialized: I0419 18:26:44.431400 132980130875200 max_utils.py:756] Using (GB) 0.43 / 31.25 (1.376000%) on TPU_24(process=6,(0,6,0,0)) I0419 18:26:44.431463 132980130875200 max_utils.py:756] Using (GB) 0.43 / 31.25 (1.376000%) on TPU_25(process=6,(1,6,0,0)) I0419 18:26:44.431494 132980130875200 max_utils.py:756] Using (GB) 0.43 / 31.25 (1.376000%) on TPU_28(process=6,(0,7,0,0)) I0419 18:26:44.431522 132980130875200 max_utils.py:756] Using (GB) 0.43 / 31.25 (1.376000%) on TPU_29(process=6,(1,7,0,0)) I0419 18:26:44.743344 132980130875200 metric_logger.py:196] completed step: 0, seconds: 1.858, TFLOP/s/device: 29.258, Tokens/s/device: 4410.091, total_weights: 262144, loss: 10.877, lm_loss: 10.877, perplexity: 52959.059 I0419 18:26:45.343747 132980130875200 metric_logger.py:196] completed step: 1, seconds: 11.972, TFLOP/s/device: 4.540, Tokens/s/device: 684.265, total_weights: 262144, loss: 10.877, lm_loss: 10.877, perplexity: 52959.059 I0419 18:26:46.172034 132980130875200 metric_logger.py:196] completed step: 2, seconds: 0.022, TFLOP/s/device: 2438.575, Tokens/s/device: 367568.538, total_weights: 262144, loss: 10.563, lm_loss: 10.563, perplexity: 38662.707 I0419 18:26:46.750270 132980130875200 metric_logger.py:196] completed step: 3, seconds: 0.828, TFLOP/s/device: 65.601, Tokens/s/device: 9888.155, total_weights: 262144, loss: 10.272, lm_loss: 10.272, perplexity: 28909.668 I0419 18:26:47.907034 132980130875200 metric_logger.py:196] completed step: 4, seconds: 0.584, TFLOP/s/device: 93.083, Tokens/s/device: 14030.545, total_weights: 262144, loss: 10.022, lm_loss: 10.022, perplexity: 22524.992 I0419 18:26:47.912811 132980130875200 metric_logger.py:196] completed step: 5, seconds: 0.578, TFLOP/s/device: 93.993, Tokens/s/device: 14167.667, total_weights: 262144, loss: 9.820, lm_loss: 9.820, perplexity: 18401.865 I0419 18:26:48.260593 2536 google_auth_provider.cc:181] Running on GCE, using service account 562977990677-compute@developer.gserviceaccount.com I0419 18:26:50.784546 132851882247936 array_metadata_store.py:203] [process=6][thread=array_type_handler] Wrote 39 array_metadata.ArrayMetadata to gs://lance-maxtext/linen_ckpt_xpk_main_20260419_180002/linen_xpk_main_20260419_180002_06_grad_accum/checkpoints/0/items/array_metadatas/process_6 I0419 18:27:22.849209 132851928397568 base_pytree_checkpoint_handler.py:130] [process=6] /jax/orbax/write/gbytes_per_sec: 40.806 MiB/s (total gbytes: 1.5 GiB) (time elapsed: 38.710243463516235 s) (per-host) I0419 18:27:22.849341 132851928397568 async_checkpointer.py:90] [process=6][thread=async_save] 3 Handler Commit operations completed. Time taken: 38.428145s. I0419 18:27:31.494488 132851928397568 async_checkpointer.py:160] [process=6][thread=async_save] Background save thread done. Time taken: 47.073274s. I0419 18:27:31.494769 132851873855232 async_checkpointer.py:288] [process=6][thread=save_finalize] Done with waiting for background save thread=async_save. I0419 18:27:31.494895 132851873855232 async_checkpointer.py:298] [process=6][thread=save_finalize] No errors found in background save thread=async_save. I0419 18:27:31.494943 132851873855232 checkpoint_manager.py:2137] [process=6][thread=save_finalize][step=0] CheckpointManager Save Finalize is syncing with other hosts... I0419 18:27:31.863566 132851873855232 checkpoint_manager.py:2146] [process=6][thread=save_finalize][step=0] CheckpointManager Save Finalize is done on all hosts. I0419 18:27:34.187725 132980130875200 metric_logger.py:196] completed step: 6, seconds: 1.157, TFLOP/s/device: 46.969, Tokens/s/device: 7079.683, total_weights: 262144, loss: 9.667, lm_loss: 9.667, perplexity: 15787.604 I0419 18:27:34.766147 132980130875200 metric_logger.py:196] completed step: 7, seconds: 45.697, TFLOP/s/device: 1.189, Tokens/s/device: 179.268, total_weights: 262144, loss: 9.561, lm_loss: 9.561, perplexity: 14203.827 I0419 18:27:35.344442 132980130875200 metric_logger.py:196] completed step: 8, seconds: 0.584, TFLOP/s/device: 93.133, Tokens/s/device: 14037.950, total_weights: 262144, loss: 9.496, lm_loss: 9.496, perplexity: 13302.920 I0419 18:27:35.922034 132980130875200 checkpointing.py:772] Waiting for step 9 to finish before checkpoint... I0419 18:27:35.922736 132980130875200 checkpointing.py:776] Waited 0.0007228851318359375 seconds for step 9 to finish before starting checkpointing. I0419 18:27:35.924768 132980130875200 checkpoint_manager.py:2009] [process=6][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning. I0419 18:27:35.926425 132980130875200 checkpoint_manager.py:1512] [process=6] Saving checkpoint at step 9 I0419 18:27:35.927845 132980130875200 event_tracking.py:70] [process=6] [async] Started save checkpoint @ gs://lance-maxtext/linen_ckpt_xpk_main_20260419_180002/linen_xpk_main_20260419_180002_06_grad_accum/checkpoints/9. I0419 18:27:36.639468 132980130875200 jax_array_handlers.py:360] Scheduling D2H of 39 prioritized jax.Array. I0419 18:27:36.639564 132980130875200 replica_slices.py:424] 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 I0419 18:27:36.673887 132980130875200 base_pytree_checkpoint_handler.py:154] [process=6][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 0.035497s I0419 18:27:36.674056 132980130875200 base_pytree_checkpoint_handler.py:130] [process=6] /jax/orbax/write/blocking_gbytes_per_sec: 39.453 GiB/s (total gbytes: 1.5 GiB) (time elapsed: 0.03909873962402344 s) (per-host) I0419 18:27:36.674112 132980130875200 base_pytree_checkpoint_handler.py:768] [process=6][thread=MainThread] Initiated Pytree async_save. Time taken: 0.039161s (batch_requests_ready=0.001805s, total_serialization_initiated=0.037287s, others=0.000069s) I0419 18:27:36.674227 132980130875200 composite_checkpoint_handler.py:715] [process=6][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 0.043589s (all_items=0.000017s, per_item={'items': '0.00001693'}, temp_paths=0.043572) I0419 18:27:36.674926 132980130875200 event_tracking.py:125] [process=6] [async] Finished blocking save in 0.75 seconds. Continuing save @ gs://lance-maxtext/linen_ckpt_xpk_main_20260419_180002/linen_xpk_main_20260419_180002_06_grad_accum/checkpoints/9. I0419 18:27:36.675193 132846014424832 async_checkpointer.py:76] [process=6][thread=async_save] Background save thread started. Deadline for this save operation is 2026-04-19 18:47:36.675169 I0419 18:27:36.680979 132980130875200 checkpoint_manager.py:1560] [process=6][thread=MainThread][step=9] Starting CheckpointManager Save Finalize thread=save_finalize I0419 18:27:36.681243 132851873855232 async_checkpointer.py:280] [process=6][thread=save_finalize] Waiting for background save thread=async_save. I0419 18:27:36.681401 132980130875200 standard_logger.py:34] {'step': 9, 'event_type': 'save', 'directory': 'gs://lance-maxtext/linen_ckpt_xpk_main_20260419_180002/linen_xpk_main_20260419_180002_06_grad_accum/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776623255.9247344, 'wait_for_prev_duration_secs': 7.486343383789062e-05, 'time_between_consecutive_saves_sec': 4.0610737800598145, 'checkpointer_blocking_start_time': 1776623255.9264627, 'checkpointer_blocking_duration_secs': 0.7488372325897217, 'get_old_steps_start_time': 1776623256.675322, 'get_old_steps_duration_secs': 3.075599670410156e-05, 'checkpoint_manager_blocking_start_time': 1776623255.9229615, 'checkpoint_manager_blocking_duration_secs': 0.7584059238433838} I0419 18:27:36.681520 132980130875200 checkpointing.py:408] Started an asynchronous checkpoint save for step 9 I0419 18:27:36.681566 132980130875200 checkpoint_manager.py:2020] [process=6][thread=MainThread][step=9][wait_until_finished] Waiting for Save Finalize thread (save_finalize) to complete. I0419 18:27:42.119360 132851928397568 array_metadata_store.py:203] [process=6][thread=array_type_handler] Wrote 39 array_metadata.ArrayMetadata to gs://lance-maxtext/linen_ckpt_xpk_main_20260419_180002/linen_xpk_main_20260419_180002_06_grad_accum/checkpoints/9/items/array_metadatas/process_6 I0419 18:28:18.916639 132846014424832 base_pytree_checkpoint_handler.py:130] [process=6] /jax/orbax/write/gbytes_per_sec: 37.359 MiB/s (total gbytes: 1.5 GiB) (time elapsed: 42.28164052963257 s) (per-host) I0419 18:28:18.916764 132846014424832 async_checkpointer.py:90] [process=6][thread=async_save] 3 Handler Commit operations completed. Time taken: 42.241490s. I0419 18:28:26.599153 132846014424832 async_checkpointer.py:160] [process=6][thread=async_save] Background save thread done. Time taken: 49.923864s. I0419 18:28:26.599463 132851873855232 async_checkpointer.py:288] [process=6][thread=save_finalize] Done with waiting for background save thread=async_save. I0419 18:28:26.599608 132851873855232 async_checkpointer.py:298] [process=6][thread=save_finalize] No errors found in background save thread=async_save. I0419 18:28:26.599668 132851873855232 checkpoint_manager.py:2137] [process=6][thread=save_finalize][step=9] CheckpointManager Save Finalize is syncing with other hosts... I0419 18:28:26.601341 132851873855232 checkpoint_manager.py:2146] [process=6][thread=save_finalize][step=9] CheckpointManager Save Finalize is done on all hosts. I0419 18:28:26.605472 132980130875200 checkpoint_manager.py:2032] [process=6][thread=MainThread][step=9][wait_until_finished] Done waiting for Save Finalize thread (save_finalize) running at step=9. I0419 18:28:26.605562 132980130875200 checkpoint_manager.py:2009] [process=6][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning. I0419 18:28:26.606436 132980130875200 metric_logger.py:196] completed step: 9, seconds: 0.578, TFLOP/s/device: 93.968, Tokens/s/device: 14163.821, total_weights: 262144, loss: 9.457, lm_loss: 9.457, perplexity: 12802.546 Per train step: Total TFLOPs: 54.35 split as 93.93% learnable weight flops and 6.07% attention flops XPK End: Sun Apr 19 18:28:38 UTC 2026 EXIT_CODE=0
574ad3fb9
· feat_nnx_trainstate_and_training_loop_20260419_214525
· full log
XPK Start: Sun Apr 19 22:10:10 UTC 2026 PyTorch was not found. Models won't be available and only tokenizers, configuration and file/data utilities can be used. 2026-04-19 22:10:34.953676: 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) I0419 22:10:35.131678 138166740404032 max_utils.py:273] Attempting to initialize the jax distributed system... I0419 22:10:44.171529 138166740404032 distributed.py:149] Starting JAX distributed service on [::]:8482 I0419 22:10:44.173829 138166740404032 distributed.py:172] Connecting to JAX distributed service on mt-06-grad-accum-nyp2l-slice-job-0-0.mt-06-grad-accum-nyp2l:8482 I0419 22:10:46.394285 138166740404032 max_utils.py:284] Jax distributed system initialized! I0419 22:10:51.582746 138166740404032 max_utils.py:800] System Information: Jax Version: 0.9.2 I0419 22:10:51.582847 138166740404032 max_utils.py:801] System Information: Jaxlib Version: 0.9.2 I0419 22:10:51.582887 138166740404032 max_utils.py:802] System Information: Jax Backend: PJRT C API TFRT TPU v6 lite Built on Mar 4 2026 11:32:08 (1772652728) cl/878335365 I0419 22:10:51.582919 138166740404032 train_utils.py:378] WARNING: Sequence packing is essentially ignored for synthetic data. Please use a real dataset to use sequence packing. I0419 22:10:52.278792 138166740404032 maxtext_utils.py:1718] Num_devices: 32, shape (1, 1, 1, 32, 1, 1, 1, 1, 1, 1, 1, 1, 1) I0419 22:10:52.279425 138166740404032 maxtext_utils.py:1718] Num_devices: 32, shape (1, 1, 1, 32, 1, 1, 1, 1, 1, 1, 1, 1, 1) I0419 22:10:52.279617 138166740404032 checkpointing.py:688] Setting up checkpoint logger... I0419 22:10:52.279670 138166740404032 checkpointing.py:234] Creating checkpoint manager with ocdbt=True and zarr3=True I0419 22:10:52.279727 138166740404032 pytree_checkpoint_handler.py:592] save_device_host_concurrent_bytes=None I0419 22:10:52.280089 138166740404032 base_pytree_checkpoint_handler.py:441] 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 0x7da8d7f140e0>, enable_pinned_host_transfer=False, save_concurrent_bytes: 96000000000 (89.4 GiB), restore_concurrent_bytes: 96000000000 (89.4 GiB) I0419 22:10:58.179375 138166740404032 checkpointing.py:266] Enabling policy for fixed interval checkpointing. I0419 22:10:58.179554 138166740404032 checkpoint_manager.py:708] [process=4][thread=MainThread] CheckpointManager init: checkpointers=None, item_names=('items',), item_handlers={'items': <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7d954c0c2b70>}, handler_registry=None I0419 22:10:58.179791 138166740404032 composite_checkpoint_handler.py:237] Deferred registration for item: "items". Adding handler `<orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7d954c0c2b70>` 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`. I0419 22:10:58.179839 138166740404032 composite_checkpoint_handler.py:237] Deferred registration for item: "metrics". Adding handler `<orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7d952c501760>` 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`. I0419 22:10:58.179874 138166740404032 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 0x7d954c0c2b70>, ('items', <class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeRestoreArgs'>): <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7d954c0c2b70>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonSaveArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7d952c501760>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonRestoreArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7d952c501760>}). I0419 22:10:58.180189 138166740404032 abstract_checkpointer.py:35] orbax-checkpoint version: 0.11.34 I0419 22:10:58.180258 138166740404032 async_checkpointer.py:192] [process=4][thread=MainThread] Using barrier_sync_fn: <function get_barrier_sync_fn.<locals>._fn at 0x7d952c5642c0> timeout: 1200 secs and primary_host=0 for async checkpoint writes I0419 22:10:59.746704 138166740404032 checkpoint_manager.py:1812] Found 0 checkpoint steps in gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_trainstate_and_training_loop_20260419_214525/linen_xpk_feat_nnx_trainstate_and_training_loop_20260419_214525_06_grad_accum/checkpoints I0419 22:11:00.716830 138166740404032 checkpoint_manager.py:929] [process=4][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_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, lightweight_initialize=False), root_directory=gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_trainstate_and_training_loop_20260419_214525/linen_xpk_feat_nnx_trainstate_and_training_loop_20260419_214525_06_grad_accum/checkpoints: <orbax.checkpoint.checkpoint_manager.CheckpointManager object at 0x7d954c0c60c0> I0419 22:11:00.716996 138166740404032 checkpointing.py:302] Checkpoint manager created! I0419 22:11:02.429880 138166740404032 nnx_wrappers.py:437] Unknown Logical: bfloat16[32,2048,2048]...................................... ('activation_batch', 'activation_norm_length', 'activation_embed'). I0419 22:11:02.429980 138166740404032 nnx_wrappers.py:437] Unknown Physical: bfloat16[32,2048,2048]...................................... ('fsdp', None, None). I0419 22:11:02.817444 138166740404032 attentions.py:1088] attentions/inputs_q Logical: bfloat16[32,2048,2048]...................................... ('activation_batch', 'activation_attn_length', 'activation_attn_embed'). I0419 22:11:02.817531 138166740404032 attentions.py:1088] attentions/inputs_q Physical: bfloat16[32,2048,2048]...................................... ('fsdp', None, None). I0419 22:11:02.833675 138166740404032 attentions.py:1089] attentions/inputs_kv Logical: bfloat16[32,2048,2048]...................................... ('activation_batch', 'activation_attn_length', 'activation_attn_embed'). I0419 22:11:02.833730 138166740404032 attentions.py:1089] attentions/inputs_kv Physical: bfloat16[32,2048,2048]...................................... ('fsdp', None, None). I0419 22:11:02.857142 138166740404032 attentions.py:1154] attentions/query Logical: bfloat16[32,2048,16,128].................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim'). I0419 22:11:02.857210 138166740404032 attentions.py:1154] attentions/query Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None). I0419 22:11:02.873396 138166740404032 attentions.py:1155] attentions/key Logical: bfloat16[32,2048,16,128].................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim'). I0419 22:11:02.873460 138166740404032 attentions.py:1155] attentions/key Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None). I0419 22:11:02.889742 138166740404032 attentions.py:1156] attentions/value Logical: bfloat16[32,2048,16,128].................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim'). I0419 22:11:02.889800 138166740404032 attentions.py:1156] attentions/value Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None). I0419 22:11:02.914432 138166740404032 attentions.py:1197] attentions/out Logical: bfloat16[32,2048,16,128].................................... ('activation_batch', 'activation_attn_length', 'activation_heads', 'activation_kv'). I0419 22:11:02.914495 138166740404032 attentions.py:1197] attentions/out Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None). I0419 22:11:02.935191 138166740404032 linears.py:525] linears/x Logical: bfloat16[32,2048,7168]...................................... ('activation_batch', 'activation_length', 'activation_mlp'). I0419 22:11:02.935256 138166740404032 linears.py:525] linears/x Physical: bfloat16[32,2048,7168]...................................... ('fsdp', None, None). I0419 22:11:03.151711 138166740404032 checkpointing.py:578] checkpoint manager exists so trying to load this run's existing checkpoint I0419 22:11:03.151815 138166740404032 checkpointing.py:676] No existing checkpoints found, not restoring checkpoint. fsdp: 32 I0419 22:11:04.560827 138166740404032 maxtext_utils.py:1821] params/params/decoder/decoder_norm/scale Shape: float32[2048] Logical: P('norm',) Physical: (None,) I0419 22:11:04.560946 138166740404032 maxtext_utils.py:1821] params/params/decoder/layers/mlp/wi_0/kernel Shape: float32[2048,16,7168] Logical: P('embed', 'layers', 'mlp') Physical: ('fsdp', None, None) I0419 22:11:04.560999 138166740404032 maxtext_utils.py:1821] params/params/decoder/layers/mlp/wi_1/kernel Shape: float32[2048,16,7168] Logical: P('embed', 'layers', 'mlp') Physical: ('fsdp', None, None) I0419 22:11:04.561058 138166740404032 maxtext_utils.py:1821] params/params/decoder/layers/mlp/wo/kernel Shape: float32[7168,16,2048] Logical: P('mlp', 'layers', 'embed') Physical: (None, None, 'fsdp') I0419 22:11:04.561109 138166740404032 maxtext_utils.py:1821] params/params/decoder/layers/post_self_attention_layer_norm/scale Shape: float32[2048,16] Logical: P('norm', 'layers') Physical: (None, None) I0419 22:11:04.561147 138166740404032 maxtext_utils.py:1821] params/params/decoder/layers/pre_self_attention_layer_norm/scale Shape: float32[2048,16] Logical: P('norm', 'layers') Physical: (None, None) I0419 22:11:04.561202 138166740404032 maxtext_utils.py:1821] 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) I0419 22:11:04.561255 138166740404032 maxtext_utils.py:1821] 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') I0419 22:11:04.561295 138166740404032 maxtext_utils.py:1821] 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) I0419 22:11:04.561346 138166740404032 maxtext_utils.py:1821] 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) I0419 22:11:04.561398 138166740404032 maxtext_utils.py:1821] params/params/decoder/logits_dense/kernel Shape: float32[2048,32000] Logical: P('embed_vocab', 'vocab') Physical: ('fsdp', None) I0419 22:11:04.561446 138166740404032 maxtext_utils.py:1821] params/params/token_embedder/embedding Shape: float32[32000,2048] Logical: P('vocab', 'embed_vocab') Physical: (None, 'fsdp') I0419 22:11:04.585793 138166740404032 gradient_accumulation.py:70] gradient_accumulation/inputs Logical: float32[2048]............................................... Unknown. I0419 22:11:04.585856 138166740404032 gradient_accumulation.py:70] gradient_accumulation/inputs Physical: float32[2048]............................................... (None,). I0419 22:11:04.600523 138166740404032 gradient_accumulation.py:70] gradient_accumulation/inputs Logical: float32[2048,16,7168]....................................... Unknown. I0419 22:11:04.600582 138166740404032 gradient_accumulation.py:70] gradient_accumulation/inputs Physical: float32[2048,16,7168]....................................... ('fsdp', None, None). I0419 22:11:04.629822 138166740404032 gradient_accumulation.py:70] gradient_accumulation/inputs Logical: float32[7168,16,2048]....................................... Unknown. I0419 22:11:04.629892 138166740404032 gradient_accumulation.py:70] gradient_accumulation/inputs Physical: float32[7168,16,2048]....................................... (None, None, 'fsdp'). I0419 22:11:04.644535 138166740404032 gradient_accumulation.py:70] gradient_accumulation/inputs Logical: float32[2048,16]............................................ Unknown. I0419 22:11:04.644595 138166740404032 gradient_accumulation.py:70] gradient_accumulation/inputs Physical: float32[2048,16]............................................ (None, None). I0419 22:11:04.673896 138166740404032 gradient_accumulation.py:70] gradient_accumulation/inputs Logical: float32[2048,16,16,128]..................................... Unknown. I0419 22:11:04.673957 138166740404032 gradient_accumulation.py:70] gradient_accumulation/inputs Physical: float32[2048,16,16,128]..................................... ('fsdp', None, None, None). I0419 22:11:04.693709 138166740404032 gradient_accumulation.py:70] gradient_accumulation/inputs Logical: float32[16,16,128,2048]..................................... Unknown. I0419 22:11:04.693810 138166740404032 gradient_accumulation.py:70] gradient_accumulation/inputs Physical: float32[16,16,128,2048]..................................... (None, None, None, 'fsdp'). I0419 22:11:04.737944 138166740404032 gradient_accumulation.py:70] gradient_accumulation/inputs Logical: float32[2048,32000]......................................... Unknown. I0419 22:11:04.738019 138166740404032 gradient_accumulation.py:70] gradient_accumulation/inputs Physical: float32[2048,32000]......................................... ('fsdp', None). I0419 22:11:04.752976 138166740404032 gradient_accumulation.py:70] gradient_accumulation/inputs Logical: float32[32000,2048]......................................... Unknown. I0419 22:11:04.753031 138166740404032 gradient_accumulation.py:70] gradient_accumulation/inputs Physical: float32[32000,2048]......................................... (None, 'fsdp'). I0419 22:11:05.423521 138166740404032 train.py:157] train/xent Logical: float32[32,2048]............................................ ('activation_embed_and_logits_batch', 'activation_length'). I0419 22:11:05.423618 138166740404032 train.py:157] train/xent Physical: float32[32,2048]............................................ ('fsdp', None). I0419 22:11:05.439021 138166740404032 train.py:164] train/z_loss Logical: float32[32,2048]............................................ ('activation_embed_and_logits_batch', 'activation_length'). I0419 22:11:05.439078 138166740404032 train.py:164] train/z_loss Physical: float32[32,2048]............................................ ('fsdp', None). I0419 22:11:16.877161 138166740404032 max_utils.py:791] Total memory size: 1.7 GB, Output size: 0.4 GB, Temp size: 1.3 GB, Argument size: 0.4 GB, Host temp size: 0.0 GB. I0419 22:11:16.877979 138166740404032 metric_logger.py:301] number parameters: 1.104 billion I0419 22:11:18.725232 138166740404032 checkpointing.py:794] Waiting for step 0 to finish before checkpoint... I0419 22:11:29.532099 138166740404032 checkpointing.py:798] Waited 10.806846141815186 seconds for step 0 to finish before starting checkpointing. I0419 22:11:29.534430 138166740404032 checkpoint_manager.py:2009] [process=4][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning. I0419 22:11:29.536237 138166740404032 checkpoint_manager.py:1512] [process=4] Saving checkpoint at step 0 I0419 22:11:29.537678 138166740404032 event_tracking.py:70] [process=4] [async] Started save checkpoint @ gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_trainstate_and_training_loop_20260419_214525/linen_xpk_feat_nnx_trainstate_and_training_loop_20260419_214525_06_grad_accum/checkpoints/0. I0419 22:11:30.736085 138166740404032 signaling_client.py:364] Using JaxDistributedSignalingClient I0419 22:11:30.737011 138166740404032 jax_array_handlers.py:360] Scheduling D2H of 39 prioritized jax.Array. I0419 22:11:30.737068 138166740404032 replica_slices.py:424] 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 I0419 22:11:31.007503 138166740404032 base_pytree_checkpoint_handler.py:154] [process=4][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 0.271495s I0419 22:11:31.007688 138166740404032 base_pytree_checkpoint_handler.py:130] [process=4] /jax/orbax/write/blocking_gbytes_per_sec: 5.572 GiB/s (total gbytes: 1.5 GiB) (time elapsed: 0.2768542766571045 s) (per-host) I0419 22:11:31.007744 138166740404032 base_pytree_checkpoint_handler.py:768] [process=4][thread=MainThread] Initiated Pytree async_save. Time taken: 0.276922s (batch_requests_ready=0.002172s, total_serialization_initiated=0.274674s, others=0.000076s) I0419 22:11:31.007837 138166740404032 composite_checkpoint_handler.py:715] [process=4][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 0.281118s (all_items=0.000018s, per_item={'items': '0.00001764'}, temp_paths=0.281101) I0419 22:11:31.008610 138166740404032 event_tracking.py:125] [process=4] [async] Finished blocking save in 1.47 seconds. Continuing save @ gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_trainstate_and_training_loop_20260419_214525/linen_xpk_feat_nnx_trainstate_and_training_loop_20260419_214525_06_grad_accum/checkpoints/0. I0419 22:11:31.008941 138041372956416 async_checkpointer.py:76] [process=4][thread=async_save] Background save thread started. Deadline for this save operation is 2026-04-19 22:31:31.008905 I0419 22:11:31.028253 138166740404032 checkpoint_manager.py:1560] [process=4][thread=MainThread][step=0] Starting CheckpointManager Save Finalize thread=save_finalize I0419 22:11:31.028538 138037631645440 async_checkpointer.py:280] [process=4][thread=save_finalize] Waiting for background save thread=async_save. I0419 22:11:31.028677 138166740404032 standard_logger.py:34] {'step': 0, 'event_type': 'save', 'directory': 'gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_trainstate_and_training_loop_20260419_214525/linen_xpk_feat_nnx_trainstate_and_training_loop_20260419_214525_06_grad_accum/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776636689.5344129, 'wait_for_prev_duration_secs': 5.6743621826171875e-05, 'time_between_consecutive_saves_sec': None, 'checkpointer_blocking_start_time': 1776636689.5362744, 'checkpointer_blocking_duration_secs': 1.4728164672851562, 'get_old_steps_start_time': 1776636691.0091147, 'get_old_steps_duration_secs': 2.956390380859375e-05, 'checkpoint_manager_blocking_start_time': 1776636689.5326028, 'checkpoint_manager_blocking_duration_secs': 1.4960355758666992} I0419 22:11:31.028781 138166740404032 checkpointing.py:409] Started an asynchronous checkpoint save for step 0 I0419 22:11:31.028832 138166740404032 max_utils.py:750] Memstats: After params initialized: I0419 22:11:31.028882 138166740404032 max_utils.py:756] Using (GB) 0.43 / 31.25 (1.376000%) on TPU_16(process=4,(0,4,0,0)) I0419 22:11:31.028917 138166740404032 max_utils.py:756] Using (GB) 0.43 / 31.25 (1.376000%) on TPU_17(process=4,(1,4,0,0)) I0419 22:11:31.028942 138166740404032 max_utils.py:756] Using (GB) 0.43 / 31.25 (1.376000%) on TPU_20(process=4,(0,5,0,0)) I0419 22:11:31.028966 138166740404032 max_utils.py:756] Using (GB) 0.43 / 31.25 (1.376000%) on TPU_21(process=4,(1,5,0,0)) I0419 22:11:31.341586 138166740404032 metric_logger.py:196] completed step: 0, seconds: 1.847, TFLOP/s/device: 29.423, Tokens/s/device: 4434.957, total_weights: 262144, loss: 10.877, lm_loss: 10.877, perplexity: 52959.059 I0419 22:11:31.946026 138166740404032 metric_logger.py:196] completed step: 1, seconds: 12.615, TFLOP/s/device: 4.308, Tokens/s/device: 649.391, total_weights: 262144, loss: 10.877, lm_loss: 10.877, perplexity: 52959.059 I0419 22:11:32.524776 138166740404032 metric_logger.py:196] completed step: 2, seconds: 0.026, TFLOP/s/device: 2089.765, Tokens/s/device: 314992.118, total_weights: 262144, loss: 10.563, lm_loss: 10.563, perplexity: 38662.707 I0419 22:11:33.103561 138166740404032 metric_logger.py:196] completed step: 3, seconds: 0.584, TFLOP/s/device: 93.005, Tokens/s/device: 14018.780, total_weights: 262144, loss: 10.272, lm_loss: 10.272, perplexity: 28909.668 I0419 22:11:34.260830 138166740404032 metric_logger.py:196] completed step: 4, seconds: 0.579, TFLOP/s/device: 93.921, Tokens/s/device: 14156.845, total_weights: 262144, loss: 10.022, lm_loss: 10.022, perplexity: 22524.992 I0419 22:11:34.266898 138166740404032 metric_logger.py:196] completed step: 5, seconds: 0.579, TFLOP/s/device: 93.826, Tokens/s/device: 14142.548, total_weights: 262144, loss: 9.820, lm_loss: 9.820, perplexity: 18401.865 I0419 22:11:36.540442 2569 google_auth_provider.cc:181] Running on GCE, using service account 562977990677-compute@developer.gserviceaccount.com I0419 22:11:39.859308 138039720855296 array_metadata_store.py:203] [process=4][thread=array_type_handler] Wrote 39 array_metadata.ArrayMetadata to gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_trainstate_and_training_loop_20260419_214525/linen_xpk_feat_nnx_trainstate_and_training_loop_20260419_214525_06_grad_accum/checkpoints/0/items/array_metadatas/process_4 I0419 22:12:11.603717 138041372956416 base_pytree_checkpoint_handler.py:130] [process=4] /jax/orbax/write/gbytes_per_sec: 38.647 MiB/s (total gbytes: 1.5 GiB) (time elapsed: 40.87285017967224 s) (per-host) I0419 22:12:11.603827 138041372956416 async_checkpointer.py:90] [process=4][thread=async_save] 3 Handler Commit operations completed. Time taken: 40.594778s. I0419 22:12:20.453643 138166740404032 metric_logger.py:196] completed step: 6, seconds: 1.157, TFLOP/s/device: 46.954, Tokens/s/device: 7077.359, total_weights: 262144, loss: 9.667, lm_loss: 9.667, perplexity: 15787.604 I0419 22:12:21.031948 138166740404032 metric_logger.py:196] completed step: 7, seconds: 45.608, TFLOP/s/device: 1.192, Tokens/s/device: 179.616, total_weights: 262144, loss: 9.561, lm_loss: 9.561, perplexity: 14203.827 I0419 22:12:21.104387 138041372956416 async_checkpointer.py:160] [process=4][thread=async_save] Background save thread done. Time taken: 50.095319s. I0419 22:12:21.104623 138037631645440 async_checkpointer.py:288] [process=4][thread=save_finalize] Done with waiting for background save thread=async_save. I0419 22:12:21.104695 138037631645440 async_checkpointer.py:298] [process=4][thread=save_finalize] No errors found in background save thread=async_save. I0419 22:12:21.104757 138037631645440 checkpoint_manager.py:2137] [process=4][thread=save_finalize][step=0] CheckpointManager Save Finalize is syncing with other hosts... I0419 22:12:21.106362 138037631645440 checkpoint_manager.py:2146] [process=4][thread=save_finalize][step=0] CheckpointManager Save Finalize is done on all hosts. I0419 22:12:21.610501 138166740404032 metric_logger.py:196] completed step: 8, seconds: 0.584, TFLOP/s/device: 93.085, Tokens/s/device: 14030.761, total_weights: 262144, loss: 9.496, lm_loss: 9.496, perplexity: 13302.920 I0419 22:12:22.188565 138166740404032 checkpointing.py:794] Waiting for step 9 to finish before checkpoint... I0419 22:12:22.189217 138166740404032 checkpointing.py:798] Waited 0.0006794929504394531 seconds for step 9 to finish before starting checkpointing. I0419 22:12:22.191386 138166740404032 checkpoint_manager.py:2009] [process=4][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning. I0419 22:12:22.193011 138166740404032 checkpoint_manager.py:1512] [process=4] Saving checkpoint at step 9 I0419 22:12:22.194747 138166740404032 event_tracking.py:70] [process=4] [async] Started save checkpoint @ gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_trainstate_and_training_loop_20260419_214525/linen_xpk_feat_nnx_trainstate_and_training_loop_20260419_214525_06_grad_accum/checkpoints/9. I0419 22:12:23.333925 138166740404032 jax_array_handlers.py:360] Scheduling D2H of 39 prioritized jax.Array. I0419 22:12:23.334017 138166740404032 replica_slices.py:424] 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 I0419 22:12:23.368819 138166740404032 base_pytree_checkpoint_handler.py:154] [process=4][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 0.035803s I0419 22:12:23.368972 138166740404032 base_pytree_checkpoint_handler.py:130] [process=4] /jax/orbax/write/blocking_gbytes_per_sec: 39.359 GiB/s (total gbytes: 1.5 GiB) (time elapsed: 0.03919243812561035 s) (per-host) I0419 22:12:23.369021 138166740404032 base_pytree_checkpoint_handler.py:768] [process=4][thread=MainThread] Initiated Pytree async_save. Time taken: 0.039251s (batch_requests_ready=0.001635s, total_serialization_initiated=0.037550s, others=0.000066s) I0419 22:12:23.369103 138166740404032 composite_checkpoint_handler.py:715] [process=4][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 0.043253s (all_items=0.000016s, per_item={'items': '0.00001645'}, temp_paths=0.043237) I0419 22:12:23.369766 138166740404032 event_tracking.py:125] [process=4] [async] Finished blocking save in 1.18 seconds. Continuing save @ gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_trainstate_and_training_loop_20260419_214525/linen_xpk_feat_nnx_trainstate_and_training_loop_20260419_214525_06_grad_accum/checkpoints/9. I0419 22:12:23.370127 138037631645440 async_checkpointer.py:76] [process=4][thread=async_save] Background save thread started. Deadline for this save operation is 2026-04-19 22:32:23.370085 I0419 22:12:23.372079 138166740404032 checkpoint_manager.py:1560] [process=4][thread=MainThread][step=9] Starting CheckpointManager Save Finalize thread=save_finalize I0419 22:12:23.372364 138045642745600 async_checkpointer.py:280] [process=4][thread=save_finalize] Waiting for background save thread=async_save. I0419 22:12:23.372521 138166740404032 standard_logger.py:34] {'step': 9, 'event_type': 'save', 'directory': 'gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_trainstate_and_training_loop_20260419_214525/linen_xpk_feat_nnx_trainstate_and_training_loop_20260419_214525_06_grad_accum/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776636742.1913533, 'wait_for_prev_duration_secs': 7.224082946777344e-05, 'time_between_consecutive_saves_sec': 1.0849494934082031, 'checkpointer_blocking_start_time': 1776636742.1930497, 'checkpointer_blocking_duration_secs': 1.177222728729248, 'get_old_steps_start_time': 1776636743.370295, 'get_old_steps_duration_secs': 2.7179718017578125e-05, 'checkpoint_manager_blocking_start_time': 1776636742.1894567, 'checkpoint_manager_blocking_duration_secs': 1.1830291748046875} I0419 22:12:23.372625 138166740404032 checkpointing.py:409] Started an asynchronous checkpoint save for step 9 I0419 22:12:23.372669 138166740404032 checkpoint_manager.py:2020] [process=4][thread=MainThread][step=9][wait_until_finished] Waiting for Save Finalize thread (save_finalize) to complete. I0419 22:12:30.195776 138035442198272 array_metadata_store.py:203] [process=4][thread=array_type_handler] Wrote 39 array_metadata.ArrayMetadata to gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_trainstate_and_training_loop_20260419_214525/linen_xpk_feat_nnx_trainstate_and_training_loop_20260419_214525_06_grad_accum/checkpoints/9/items/array_metadatas/process_4 I0419 22:13:06.347925 138037631645440 base_pytree_checkpoint_handler.py:130] [process=4] /jax/orbax/write/gbytes_per_sec: 36.719 MiB/s (total gbytes: 1.5 GiB) (time elapsed: 43.01810908317566 s) (per-host) I0419 22:13:06.348043 138037631645440 async_checkpointer.py:90] [process=4][thread=async_save] 3 Handler Commit operations completed. Time taken: 42.977801s. I0419 22:13:17.450921 138037631645440 async_checkpointer.py:160] [process=4][thread=async_save] Background save thread done. Time taken: 54.080663s. I0419 22:13:17.451201 138045642745600 async_checkpointer.py:288] [process=4][thread=save_finalize] Done with waiting for background save thread=async_save. I0419 22:13:17.451323 138045642745600 async_checkpointer.py:298] [process=4][thread=save_finalize] No errors found in background save thread=async_save. I0419 22:13:17.451403 138045642745600 checkpoint_manager.py:2137] [process=4][thread=save_finalize][step=9] CheckpointManager Save Finalize is syncing with other hosts... I0419 22:13:17.453255 138045642745600 checkpoint_manager.py:2146] [process=4][thread=save_finalize][step=9] CheckpointManager Save Finalize is done on all hosts. I0419 22:13:17.453438 138166740404032 checkpoint_manager.py:2032] [process=4][thread=MainThread][step=9][wait_until_finished] Done waiting for Save Finalize thread (save_finalize) running at step=9. I0419 22:13:17.453581 138166740404032 checkpoint_manager.py:2009] [process=4][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning. I0419 22:13:17.454603 138166740404032 metric_logger.py:196] completed step: 9, seconds: 0.578, TFLOP/s/device: 93.971, Tokens/s/device: 14164.311, total_weights: 262144, loss: 9.457, lm_loss: 9.457, perplexity: 12802.546 Per train step: Total TFLOPs: 54.35 split as 93.93% learnable weight flops and 6.07% attention flops XPK End: Sun Apr 19 22:13:25 UTC 2026 EXIT_CODE=0