MaxView

‹ 05_fp8Case: 06_grad_accum07_eval ›

Metrics: Linen vs NNX  ·  feat/nnx-post-train-fixes

MetricLinen  28b1e4a16NNX  28b1e4a16Diff (NNX − Linen)
Parameters1.104 billion1.104 billion
Final loss9.45709.45700
TFLOP/s93.97493.876-0.098
Tok/s14164.714149.9-14.778
Avg s/step6.8976.854-0.043
Memory %1.381.380
JAX0.9.20.9.2

Diff = NNX value − Linen value. Green = NNX improved. Red = NNX regressed.

Linen  ·  28b1e4a16  ·  feat_nnx_post_train_fixes_20260425_121355  ·  full log
XPK Start: Sat Apr 25 12:24:11 UTC 2026
PyTorch was not found. Models won't be available and only tokenizers, configuration and file/data utilities can be used.
`rope_parameters`'s factor field must be a float >= 1, got 40
`rope_parameters`'s beta_fast field must be a float, got 32
`rope_parameters`'s beta_slow field must be a float, got 1
DeepseekV32Config got `key=rope_scaling` in kwargs but hasn't set it as attribute. For RoPE standardization you need to set `self.rope_parameters` in model's config. 
2026-04-25 12:24:36.340994: 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)
I0425 12:24:36.556801 134452460345152 max_utils.py:273] Attempting to initialize the jax distributed system...
I0425 12:24:45.600080 134452460345152 distributed.py:149] Starting JAX distributed service on [::]:8482
I0425 12:24:45.602398 134452460345152 distributed.py:172] Connecting to JAX distributed service on mt-06-grad-accum-ndjc5-slice-job-0-0.mt-06-grad-accum-ndjc5:8482
I0425 12:24:46.575775 134452460345152 max_utils.py:284] Jax distributed system initialized!
I0425 12:24:52.773895 134452460345152 max_utils.py:800] System Information: Jax Version: 0.9.2
I0425 12:24:52.774001 134452460345152 max_utils.py:801] System Information: Jaxlib Version: 0.9.2
I0425 12:24:52.774042 134452460345152 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
I0425 12:24:52.774079 134452460345152 train_utils.py:391] WARNING: Sequence packing is essentially ignored for synthetic data. Please use a real dataset to use sequence packing.
I0425 12:24:53.468997 134452460345152 maxtext_utils.py:1771] Num_devices: 32, shape (1, 1, 1, 32, 1, 1, 1, 1, 1, 1, 1, 1, 1)
I0425 12:24:53.469615 134452460345152 maxtext_utils.py:1771] Num_devices: 32, shape (1, 1, 1, 32, 1, 1, 1, 1, 1, 1, 1, 1, 1)
I0425 12:24:53.469799 134452460345152 checkpointing.py:688] Setting up checkpoint logger...
I0425 12:24:53.469851 134452460345152 checkpointing.py:234] Creating checkpoint manager with ocdbt=True and zarr3=True
I0425 12:24:53.469901 134452460345152 pytree_checkpoint_handler.py:592] save_device_host_concurrent_bytes=None
I0425 12:24:53.470258 134452460345152 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 0x7a481a31c200>, enable_pinned_host_transfer=False, save_concurrent_bytes: 96000000000 (89.4 GiB), restore_concurrent_bytes: 96000000000 (89.4 GiB)
I0425 12:24:56.313743 134452460345152 checkpointing.py:266] Enabling policy for fixed interval checkpointing.
I0425 12:24:56.314083 134452460345152 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 0x7a34183df830>}, handler_registry=None
I0425 12:24:56.314345 134452460345152 composite_checkpoint_handler.py:237] Deferred registration for item: "items". Adding handler `<orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7a34183df830>` 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`.
I0425 12:24:56.314396 134452460345152 composite_checkpoint_handler.py:237] Deferred registration for item: "metrics". Adding handler `<orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7a46e3fd5430>` 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`.
I0425 12:24:56.314434 134452460345152 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 0x7a34183df830>, ('items', <class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeRestoreArgs'>): <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7a34183df830>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonSaveArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7a46e3fd5430>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonRestoreArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7a46e3fd5430>}).
I0425 12:24:56.314758 134452460345152 abstract_checkpointer.py:35] orbax-checkpoint version: 0.11.34
I0425 12:24:56.314828 134452460345152 async_checkpointer.py:192] [process=6][thread=MainThread] Using barrier_sync_fn: <function get_barrier_sync_fn.<locals>._fn at 0x7a341816dd00> timeout: 1200 secs and primary_host=0 for async checkpoint writes
I0425 12:24:57.759974 134452460345152 checkpoint_manager.py:1812] Found 0 checkpoint steps in gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_post_train_fixes_20260425_121355/linen_xpk_feat_nnx_post_train_fixes_20260425_121355_06_grad_accum/checkpoints
I0425 12:24:57.762271 134452460345152 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_feat_nnx_post_train_fixes_20260425_121355/linen_xpk_feat_nnx_post_train_fixes_20260425_121355_06_grad_accum/checkpoints: <orbax.checkpoint.checkpoint_manager.CheckpointManager object at 0x7a34183df860>
I0425 12:24:57.762385 134452460345152 checkpointing.py:302] Checkpoint manager created!
I0425 12:24:59.497717 134452460345152 nnx_wrappers.py:437] Unknown Logical: bfloat16[32,2048,2048]...................................... ('activation_batch', 'activation_norm_length', 'activation_embed').
I0425 12:24:59.497821 134452460345152 nnx_wrappers.py:437] Unknown Physical: bfloat16[32,2048,2048]...................................... ('fsdp', None, None).
I0425 12:24:59.887828 134452460345152 attentions.py:1088] attentions/inputs_q Logical: bfloat16[32,2048,2048]...................................... ('activation_batch_attn', 'activation_length_attn', 'activation_embed_attn').
I0425 12:24:59.887927 134452460345152 attentions.py:1088] attentions/inputs_q Physical: bfloat16[32,2048,2048]...................................... ('fsdp', None, None).
I0425 12:24:59.904777 134452460345152 attentions.py:1089] attentions/inputs_kv Logical: bfloat16[32,2048,2048]...................................... ('activation_batch_attn', 'activation_length_attn', 'activation_embed_attn').
I0425 12:24:59.904838 134452460345152 attentions.py:1089] attentions/inputs_kv Physical: bfloat16[32,2048,2048]...................................... ('fsdp', None, None).
I0425 12:24:59.929146 134452460345152 attentions.py:1154] attentions/query Logical: bfloat16[32,2048,16,128].................................... ('activation_kv_batch', 'activation_length_attn', 'activation_kv_heads', 'activation_kv_head_dim').
I0425 12:24:59.929215 134452460345152 attentions.py:1154] attentions/query Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None).
I0425 12:24:59.945888 134452460345152 attentions.py:1155] attentions/key Logical: bfloat16[32,2048,16,128].................................... ('activation_kv_batch', 'activation_length_attn', 'activation_kv_heads', 'activation_kv_head_dim').
I0425 12:24:59.945955 134452460345152 attentions.py:1155] attentions/key Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None).
I0425 12:24:59.962740 134452460345152 attentions.py:1156] attentions/value Logical: bfloat16[32,2048,16,128].................................... ('activation_kv_batch', 'activation_length_attn', 'activation_kv_heads', 'activation_kv_head_dim').
I0425 12:24:59.962802 134452460345152 attentions.py:1156] attentions/value Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None).
I0425 12:24:59.987932 134452460345152 attentions.py:1198] attentions/out Logical: bfloat16[32,2048,16,128].................................... ('activation_batch_attn', 'activation_length_attn', 'activation_heads', 'activation_kv').
I0425 12:24:59.988003 134452460345152 attentions.py:1198] attentions/out Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None).
I0425 12:25:00.009121 134452460345152 linears.py:525] linears/x Logical: bfloat16[32,2048,7168]...................................... ('activation_batch', 'activation_length', 'activation_mlp').
I0425 12:25:00.009185 134452460345152 linears.py:525] linears/x Physical: bfloat16[32,2048,7168]...................................... ('fsdp', None, None).
I0425 12:25:00.231319 134452460345152 checkpointing.py:578] checkpoint manager exists so trying to load this run's existing checkpoint
I0425 12:25:00.231433 134452460345152 checkpointing.py:676] No existing checkpoints found, not restoring checkpoint.
fsdp: 32
I0425 12:25:01.666321 134452460345152 maxtext_utils.py:1880]  params/params/decoder/decoder_norm/scale
    Shape:     float32[2048]
    Logical:   P('norm',)
    Physical:  (None,)
I0425 12:25:01.666452 134452460345152 maxtext_utils.py:1880]  params/params/decoder/layers/mlp/wi_0/kernel
    Shape:     float32[2048,16,7168]
    Logical:   P('embed', 'layers', 'mlp')
    Physical:  ('fsdp', None, None)
I0425 12:25:01.666507 134452460345152 maxtext_utils.py:1880]  params/params/decoder/layers/mlp/wi_1/kernel
    Shape:     float32[2048,16,7168]
    Logical:   P('embed', 'layers', 'mlp')
    Physical:  ('fsdp', None, None)
I0425 12:25:01.666572 134452460345152 maxtext_utils.py:1880]  params/params/decoder/layers/mlp/wo/kernel
    Shape:     float32[7168,16,2048]
    Logical:   P('mlp', 'layers', 'embed')
    Physical:  (None, None, 'fsdp')
I0425 12:25:01.666627 134452460345152 maxtext_utils.py:1880]  params/params/decoder/layers/post_self_attention_layer_norm/scale
    Shape:     float32[2048,16]
    Logical:   P('norm', 'layers')
    Physical:  (None, None)
I0425 12:25:01.666665 134452460345152 maxtext_utils.py:1880]  params/params/decoder/layers/pre_self_attention_layer_norm/scale
    Shape:     float32[2048,16]
    Logical:   P('norm', 'layers')
    Physical:  (None, None)
I0425 12:25:01.666717 134452460345152 maxtext_utils.py:1880]  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)
I0425 12:25:01.666768 134452460345152 maxtext_utils.py:1880]  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')
I0425 12:25:01.666809 134452460345152 maxtext_utils.py:1880]  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)
I0425 12:25:01.666846 134452460345152 maxtext_utils.py:1880]  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)
I0425 12:25:01.666892 134452460345152 maxtext_utils.py:1880]  params/params/decoder/logits_dense/kernel
    Shape:     float32[2048,32000]
    Logical:   P('embed_vocab', 'vocab')
    Physical:  ('fsdp', None)
I0425 12:25:01.666942 134452460345152 maxtext_utils.py:1880]  params/params/token_embedder/embedding
    Shape:     float32[32000,2048]
    Logical:   P('vocab', 'embed_vocab')
    Physical:  (None, 'fsdp')

I0425 12:25:01.695855 134452460345152 gradient_accumulation.py:70] gradient_accumulation/inputs Logical: float32[2048]............................................... Unknown.
I0425 12:25:01.695957 134452460345152 gradient_accumulation.py:70] gradient_accumulation/inputs Physical: float32[2048]............................................... (None,).
I0425 12:25:01.711204 134452460345152 gradient_accumulation.py:70] gradient_accumulation/inputs Logical: float32[2048,16,7168]....................................... Unknown.
I0425 12:25:01.711263 134452460345152 gradient_accumulation.py:70] gradient_accumulation/inputs Physical: float32[2048,16,7168]....................................... ('fsdp', None, None).
I0425 12:25:01.741261 134452460345152 gradient_accumulation.py:70] gradient_accumulation/inputs Logical: float32[7168,16,2048]....................................... Unknown.
I0425 12:25:01.741325 134452460345152 gradient_accumulation.py:70] gradient_accumulation/inputs Physical: float32[7168,16,2048]....................................... (None, None, 'fsdp').
I0425 12:25:01.756428 134452460345152 gradient_accumulation.py:70] gradient_accumulation/inputs Logical: float32[2048,16]............................................ Unknown.
I0425 12:25:01.756484 134452460345152 gradient_accumulation.py:70] gradient_accumulation/inputs Physical: float32[2048,16]............................................ (None, None).
I0425 12:25:01.786473 134452460345152 gradient_accumulation.py:70] gradient_accumulation/inputs Logical: float32[2048,16,16,128]..................................... Unknown.
I0425 12:25:01.786544 134452460345152 gradient_accumulation.py:70] gradient_accumulation/inputs Physical: float32[2048,16,16,128]..................................... ('fsdp', None, None, None).
I0425 12:25:01.801588 134452460345152 gradient_accumulation.py:70] gradient_accumulation/inputs Logical: float32[16,16,128,2048]..................................... Unknown.
I0425 12:25:01.801645 134452460345152 gradient_accumulation.py:70] gradient_accumulation/inputs Physical: float32[16,16,128,2048]..................................... (None, None, None, 'fsdp').
I0425 12:25:01.846640 134452460345152 gradient_accumulation.py:70] gradient_accumulation/inputs Logical: float32[2048,32000]......................................... Unknown.
I0425 12:25:01.846711 134452460345152 gradient_accumulation.py:70] gradient_accumulation/inputs Physical: float32[2048,32000]......................................... ('fsdp', None).
I0425 12:25:01.861717 134452460345152 gradient_accumulation.py:70] gradient_accumulation/inputs Logical: float32[32000,2048]......................................... Unknown.
I0425 12:25:01.861775 134452460345152 gradient_accumulation.py:70] gradient_accumulation/inputs Physical: float32[32000,2048]......................................... (None, 'fsdp').
I0425 12:25:02.547379 134452460345152 train.py:157] train/xent Logical: float32[32,2048]............................................ ('activation_embed_and_logits_batch', 'activation_length').
I0425 12:25:02.547475 134452460345152 train.py:157] train/xent Physical: float32[32,2048]............................................ ('fsdp', None).
I0425 12:25:02.563363 134452460345152 train.py:164] train/z_loss Logical: float32[32,2048]............................................ ('activation_embed_and_logits_batch', 'activation_length').
I0425 12:25:02.563423 134452460345152 train.py:164] train/z_loss Physical: float32[32,2048]............................................ ('fsdp', None).
I0425 12:25:14.165463 134452460345152 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.
I0425 12:25:14.166346 134452460345152 metric_logger.py:301] number parameters: 1.104 billion
I0425 12:25:16.069761 134452460345152 checkpointing.py:794] Waiting for step 0 to finish before checkpoint...
I0425 12:25:27.168294 134452460345152 checkpointing.py:798] Waited 11.098504066467285 seconds for step 0 to finish before starting checkpointing.
I0425 12:25:27.170757 134452460345152 checkpoint_manager.py:2009] [process=6][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning.
I0425 12:25:27.172700 134452460345152 checkpoint_manager.py:1512] [process=6] Saving checkpoint at step 0
I0425 12:25:27.174220 134452460345152 event_tracking.py:70] [process=6] [async] Started save checkpoint @ gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_post_train_fixes_20260425_121355/linen_xpk_feat_nnx_post_train_fixes_20260425_121355_06_grad_accum/checkpoints/0.
I0425 12:25:27.535589 134452460345152 signaling_client.py:364] Using JaxDistributedSignalingClient
I0425 12:25:27.536538 134452460345152 jax_array_handlers.py:360] Scheduling D2H of 39 prioritized jax.Array.
I0425 12:25:27.536657 134452460345152 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
I0425 12:25:27.809545 134452460345152 base_pytree_checkpoint_handler.py:154] [process=6][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 0.274058s
I0425 12:25:27.809718 134452460345152 base_pytree_checkpoint_handler.py:130] [process=6] /jax/orbax/write/blocking_gbytes_per_sec: 5.512 GiB/s (total gbytes: 1.5 GiB) (time elapsed: 0.2798466682434082 s) (per-host)
I0425 12:25:27.809770 134452460345152 base_pytree_checkpoint_handler.py:768] [process=6][thread=MainThread] Initiated Pytree async_save. Time taken: 0.279908s (batch_requests_ready=0.002362s, total_serialization_initiated=0.277476s, others=0.000070s)
I0425 12:25:27.809860 134452460345152 composite_checkpoint_handler.py:715] [process=6][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 0.284106s (all_items=0.000018s, per_item={'items': '0.00001788'}, temp_paths=0.284088)
I0425 12:25:27.810608 134452460345152 event_tracking.py:125] [process=6] [async] Finished blocking save in 0.64 seconds. Continuing save @ gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_post_train_fixes_20260425_121355/linen_xpk_feat_nnx_post_train_fixes_20260425_121355_06_grad_accum/checkpoints/0.
I0425 12:25:27.810944 134324740683520 async_checkpointer.py:76] [process=6][thread=async_save] Background save thread started. Deadline for this save operation is 2026-04-25 12:45:27.810907
I0425 12:25:27.819964 134452460345152 checkpoint_manager.py:1560] [process=6][thread=MainThread][step=0] Starting CheckpointManager Save Finalize thread=save_finalize
I0425 12:25:27.820220 134324221372160 async_checkpointer.py:280] [process=6][thread=save_finalize] Waiting for background save thread=async_save.
I0425 12:25:27.820347 134452460345152 standard_logger.py:34] {'step': 0, 'event_type': 'save', 'directory': 'gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_post_train_fixes_20260425_121355/linen_xpk_feat_nnx_post_train_fixes_20260425_121355_06_grad_accum/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1777119927.170739, 'wait_for_prev_duration_secs': 6.318092346191406e-05, 'time_between_consecutive_saves_sec': None, 'checkpointer_blocking_start_time': 1777119927.1727388, 'checkpointer_blocking_duration_secs': 0.638369083404541, 'get_old_steps_start_time': 1777119927.811132, 'get_old_steps_duration_secs': 3.147125244140625e-05, 'checkpoint_manager_blocking_start_time': 1777119927.1688166, 'checkpoint_manager_blocking_duration_secs': 0.6514883041381836}
I0425 12:25:27.820452 134452460345152 checkpointing.py:409] Started an asynchronous checkpoint save for step 0
I0425 12:25:27.820499 134452460345152 max_utils.py:750] 
Memstats: After params initialized:
I0425 12:25:27.820545 134452460345152 max_utils.py:756] 	Using (GB) 0.43 / 31.25 (1.376000%) on TPU_24(process=6,(0,6,0,0))
I0425 12:25:27.820578 134452460345152 max_utils.py:756] 	Using (GB) 0.43 / 31.25 (1.376000%) on TPU_25(process=6,(1,6,0,0))
I0425 12:25:27.820604 134452460345152 max_utils.py:756] 	Using (GB) 0.43 / 31.25 (1.376000%) on TPU_28(process=6,(0,7,0,0))
I0425 12:25:27.820628 134452460345152 max_utils.py:756] 	Using (GB) 0.43 / 31.25 (1.376000%) on TPU_29(process=6,(1,7,0,0))
I0425 12:25:28.132968 134452460345152 metric_logger.py:196] completed step: 0, seconds: 1.903, TFLOP/s/device: 28.555, Tokens/s/device: 4304.126, total_weights: 262144, loss: 10.877, lm_loss: 10.877, perplexity: 52959.059
I0425 12:25:28.737223 134452460345152 metric_logger.py:196] completed step: 1, seconds: 12.062, TFLOP/s/device: 4.506, Tokens/s/device: 679.181, total_weights: 262144, loss: 10.877, lm_loss: 10.877, perplexity: 52959.059
I0425 12:25:29.315138 134452460345152 metric_logger.py:196] completed step: 2, seconds: 0.026, TFLOP/s/device: 2069.080, Tokens/s/device: 311874.215, total_weights: 262144, loss: 10.563, lm_loss: 10.563, perplexity: 38662.707
I0425 12:25:29.893062 134452460345152 metric_logger.py:196] completed step: 3, seconds: 0.584, TFLOP/s/device: 93.040, Tokens/s/device: 14023.963, total_weights: 262144, loss: 10.272, lm_loss: 10.272, perplexity: 28909.668
I0425 12:25:30.684459    2564 google_auth_provider.cc:181] Running on GCE, using service account 562977990677-compute@developer.gserviceaccount.com
I0425 12:25:31.049549 134452460345152 metric_logger.py:196] completed step: 4, seconds: 0.578, TFLOP/s/device: 94.063, Tokens/s/device: 14178.211, total_weights: 262144, loss: 10.022, lm_loss: 10.022, perplexity: 22524.992
I0425 12:25:31.055671 134452460345152 metric_logger.py:196] completed step: 5, seconds: 0.578, TFLOP/s/device: 94.046, Tokens/s/device: 14175.635, total_weights: 262144, loss: 9.820, lm_loss: 9.820, perplexity: 18401.865
I0425 12:25:32.742630 134324229764864 array_metadata_store.py:203] [process=6][thread=array_type_handler] Wrote 39 array_metadata.ArrayMetadata to gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_post_train_fixes_20260425_121355/linen_xpk_feat_nnx_post_train_fixes_20260425_121355_06_grad_accum/checkpoints/0/items/array_metadatas/process_6
I0425 12:26:05.643560 134324740683520 base_pytree_checkpoint_handler.py:130] [process=6] /jax/orbax/write/gbytes_per_sec: 41.444 MiB/s (total gbytes: 1.5 GiB) (time elapsed: 38.11365032196045 s) (per-host)
I0425 12:26:05.643691 134324740683520 async_checkpointer.py:90] [process=6][thread=async_save] 3 Handler Commit operations completed. Time taken: 37.832636s.
I0425 12:26:11.427745 134324740683520 async_checkpointer.py:160] [process=6][thread=async_save] Background save thread done. Time taken: 43.616674s.
I0425 12:26:11.427977 134324221372160 async_checkpointer.py:288] [process=6][thread=save_finalize] Done with waiting for background save thread=async_save.
I0425 12:26:11.428040 134324221372160 async_checkpointer.py:298] [process=6][thread=save_finalize] No errors found in background save thread=async_save.
I0425 12:26:11.428113 134324221372160 checkpoint_manager.py:2137] [process=6][thread=save_finalize][step=0] CheckpointManager Save Finalize is syncing with other hosts...
I0425 12:26:11.430015 134324221372160 checkpoint_manager.py:2146] [process=6][thread=save_finalize][step=0] CheckpointManager Save Finalize is done on all hosts.
I0425 12:26:17.562240 134452460345152 metric_logger.py:196] completed step: 6, seconds: 1.157, TFLOP/s/device: 46.969, Tokens/s/device: 7079.732, total_weights: 262144, loss: 9.667, lm_loss: 9.667, perplexity: 15787.604
I0425 12:26:18.140390 134452460345152 metric_logger.py:196] completed step: 7, seconds: 45.929, TFLOP/s/device: 1.183, Tokens/s/device: 178.362, total_weights: 262144, loss: 9.561, lm_loss: 9.561, perplexity: 14203.827
I0425 12:26:18.718675 134452460345152 metric_logger.py:196] completed step: 8, seconds: 0.583, TFLOP/s/device: 93.225, Tokens/s/device: 14051.868, total_weights: 262144, loss: 9.496, lm_loss: 9.496, perplexity: 13302.920
I0425 12:26:19.296545 134452460345152 checkpointing.py:794] Waiting for step 9 to finish before checkpoint...
I0425 12:26:19.297224 134452460345152 checkpointing.py:798] Waited 0.0007004737854003906 seconds for step 9 to finish before starting checkpointing.
I0425 12:26:19.299318 134452460345152 checkpoint_manager.py:2009] [process=6][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning.
I0425 12:26:19.301256 134452460345152 checkpoint_manager.py:1512] [process=6] Saving checkpoint at step 9
I0425 12:26:19.302597 134452460345152 event_tracking.py:70] [process=6] [async] Started save checkpoint @ gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_post_train_fixes_20260425_121355/linen_xpk_feat_nnx_post_train_fixes_20260425_121355_06_grad_accum/checkpoints/9.
I0425 12:26:20.030508 134452460345152 jax_array_handlers.py:360] Scheduling D2H of 39 prioritized jax.Array.
I0425 12:26:20.030619 134452460345152 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
I0425 12:26:20.065685 134452460345152 base_pytree_checkpoint_handler.py:154] [process=6][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 0.036127s
I0425 12:26:20.065855 134452460345152 base_pytree_checkpoint_handler.py:130] [process=6] /jax/orbax/write/blocking_gbytes_per_sec: 38.787 GiB/s (total gbytes: 1.5 GiB) (time elapsed: 0.03977012634277344 s) (per-host)
I0425 12:26:20.065908 134452460345152 base_pytree_checkpoint_handler.py:768] [process=6][thread=MainThread] Initiated Pytree async_save. Time taken: 0.039833s (batch_requests_ready=0.001812s, total_serialization_initiated=0.037950s, others=0.000071s)
I0425 12:26:20.065997 134452460345152 composite_checkpoint_handler.py:715] [process=6][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 0.044005s (all_items=0.000015s, per_item={'items': '0.00001526'}, temp_paths=0.043990)
I0425 12:26:20.066739 134452460345152 event_tracking.py:125] [process=6] [async] Finished blocking save in 0.77 seconds. Continuing save @ gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_post_train_fixes_20260425_121355/linen_xpk_feat_nnx_post_train_fixes_20260425_121355_06_grad_accum/checkpoints/9.
I0425 12:26:20.067088 134329019639552 async_checkpointer.py:76] [process=6][thread=async_save] Background save thread started. Deadline for this save operation is 2026-04-25 12:46:20.067050
I0425 12:26:20.069007 134452460345152 checkpoint_manager.py:1560] [process=6][thread=MainThread][step=9] Starting CheckpointManager Save Finalize thread=save_finalize
I0425 12:26:20.069308 134324204586752 async_checkpointer.py:280] [process=6][thread=save_finalize] Waiting for background save thread=async_save.
I0425 12:26:20.069448 134452460345152 standard_logger.py:34] {'step': 9, 'event_type': 'save', 'directory': 'gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_post_train_fixes_20260425_121355/linen_xpk_feat_nnx_post_train_fixes_20260425_121355_06_grad_accum/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1777119979.299284, 'wait_for_prev_duration_secs': 8.082389831542969e-05, 'time_between_consecutive_saves_sec': 7.869229316711426, 'checkpointer_blocking_start_time': 1777119979.3012962, 'checkpointer_blocking_duration_secs': 0.7659118175506592, 'get_old_steps_start_time': 1777119980.0672355, 'get_old_steps_duration_secs': 3.0279159545898438e-05, 'checkpoint_manager_blocking_start_time': 1777119979.2974646, 'checkpoint_manager_blocking_duration_secs': 0.7719497680664062}
I0425 12:26:20.069601 134452460345152 checkpointing.py:409] Started an asynchronous checkpoint save for step 9
I0425 12:26:20.069648 134452460345152 checkpoint_manager.py:2020] [process=6][thread=MainThread][step=9][wait_until_finished] Waiting for Save Finalize thread (save_finalize) to complete.
I0425 12:26:25.256240 134324212979456 array_metadata_store.py:203] [process=6][thread=array_type_handler] Wrote 39 array_metadata.ArrayMetadata to gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_post_train_fixes_20260425_121355/linen_xpk_feat_nnx_post_train_fixes_20260425_121355_06_grad_accum/checkpoints/9/items/array_metadatas/process_6
I0425 12:27:02.676635 134329019639552 base_pytree_checkpoint_handler.py:130] [process=6] /jax/orbax/write/gbytes_per_sec: 37.036 MiB/s (total gbytes: 1.5 GiB) (time elapsed: 42.65051198005676 s) (per-host)
I0425 12:27:02.676766 134329019639552 async_checkpointer.py:90] [process=6][thread=async_save] 3 Handler Commit operations completed. Time taken: 42.609045s.
I0425 12:27:08.702904 134329019639552 async_checkpointer.py:160] [process=6][thread=async_save] Background save thread done. Time taken: 48.635167s.
I0425 12:27:08.703188 134324204586752 async_checkpointer.py:288] [process=6][thread=save_finalize] Done with waiting for background save thread=async_save.
I0425 12:27:08.703313 134324204586752 async_checkpointer.py:298] [process=6][thread=save_finalize] No errors found in background save thread=async_save.
I0425 12:27:08.703372 134324204586752 checkpoint_manager.py:2137] [process=6][thread=save_finalize][step=9] CheckpointManager Save Finalize is syncing with other hosts...
I0425 12:27:08.704844 134324204586752 checkpoint_manager.py:2146] [process=6][thread=save_finalize][step=9] CheckpointManager Save Finalize is done on all hosts.
I0425 12:27:08.705012 134452460345152 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.
I0425 12:27:08.705170 134452460345152 checkpoint_manager.py:2009] [process=6][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning.
I0425 12:27:08.706063 134452460345152 metric_logger.py:196] completed step: 9, seconds: 0.578, TFLOP/s/device: 93.974, Tokens/s/device: 14164.727, 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: Sat Apr 25 12:27:21 UTC 2026
EXIT_CODE=0
NNX  ·  28b1e4a16  ·  feat_nnx_post_train_fixes_20260425_121355  ·  full log
XPK Start: Sat Apr 25 13:49:43 UTC 2026
PyTorch was not found. Models won't be available and only tokenizers, configuration and file/data utilities can be used.
`rope_parameters`'s factor field must be a float >= 1, got 40
`rope_parameters`'s beta_fast field must be a float, got 32
`rope_parameters`'s beta_slow field must be a float, got 1
DeepseekV32Config got `key=rope_scaling` in kwargs but hasn't set it as attribute. For RoPE standardization you need to set `self.rope_parameters` in model's config. 
2026-04-25 13:50:07.870539: 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)
I0425 13:50:08.436819 138672244799296 max_utils.py:273] Attempting to initialize the jax distributed system...
I0425 13:50:17.476965 138672244799296 distributed.py:149] Starting JAX distributed service on [::]:8482
I0425 13:50:17.479286 138672244799296 distributed.py:172] Connecting to JAX distributed service on mt-06-grad-accum-ysknz-slice-job-0-0.mt-06-grad-accum-ysknz:8482
I0425 13:50:18.797026 138672244799296 max_utils.py:284] Jax distributed system initialized!
I0425 13:50:24.792316 138672244799296 max_utils.py:800] System Information: Jax Version: 0.9.2
I0425 13:50:24.792424 138672244799296 max_utils.py:801] System Information: Jaxlib Version: 0.9.2
I0425 13:50:24.792465 138672244799296 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
I0425 13:50:24.792504 138672244799296 train_utils.py:391] WARNING: Sequence packing is essentially ignored for synthetic data. Please use a real dataset to use sequence packing.
I0425 13:50:25.500897 138672244799296 maxtext_utils.py:1771] Num_devices: 32, shape (1, 1, 1, 32, 1, 1, 1, 1, 1, 1, 1, 1, 1)
I0425 13:50:25.615007 138672244799296 checkpointing.py:688] Setting up checkpoint logger...
I0425 13:50:25.615125 138672244799296 checkpointing.py:234] Creating checkpoint manager with ocdbt=True and zarr3=True
I0425 13:50:25.615171 138672244799296 pytree_checkpoint_handler.py:592] save_device_host_concurrent_bytes=None
I0425 13:50:25.615374 138672244799296 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 0x7e1e8ab7a5d0>, enable_pinned_host_transfer=False, save_concurrent_bytes: 96000000000 (89.4 GiB), restore_concurrent_bytes: 96000000000 (89.4 GiB)
I0425 13:50:28.475950 138672244799296 checkpointing.py:266] Enabling policy for fixed interval checkpointing.
I0425 13:50:28.476192 138672244799296 checkpoint_manager.py:708] [process=5][thread=MainThread] CheckpointManager init: checkpointers=None, item_names=('items',), item_handlers={'items': <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7e09fc392750>}, handler_registry=None
I0425 13:50:28.476431 138672244799296 composite_checkpoint_handler.py:237] Deferred registration for item: "items". Adding handler `<orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7e09fc392750>` 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`.
I0425 13:50:28.476480 138672244799296 composite_checkpoint_handler.py:237] Deferred registration for item: "metrics". Adding handler `<orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7e09fc337da0>` 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`.
I0425 13:50:28.476516 138672244799296 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 0x7e09fc392750>, ('items', <class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeRestoreArgs'>): <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7e09fc392750>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonSaveArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7e09fc337da0>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonRestoreArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7e09fc337da0>}).
I0425 13:50:28.476853 138672244799296 abstract_checkpointer.py:35] orbax-checkpoint version: 0.11.34
I0425 13:50:28.476924 138672244799296 async_checkpointer.py:192] [process=5][thread=MainThread] Using barrier_sync_fn: <function get_barrier_sync_fn.<locals>._fn at 0x7e09fc30b1a0> timeout: 1200 secs and primary_host=0 for async checkpoint writes
I0425 13:50:29.208391 138672244799296 checkpoint_manager.py:1812] Found 0 checkpoint steps in gs://lance-maxtext/nnx_ckpt_xpk_feat_nnx_post_train_fixes_20260425_121355/nnx_xpk_feat_nnx_post_train_fixes_20260425_121355_06_grad_accum/checkpoints
I0425 13:50:29.399814 138672244799296 checkpoint_manager.py:929] [process=5][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/nnx_ckpt_xpk_feat_nnx_post_train_fixes_20260425_121355/nnx_xpk_feat_nnx_post_train_fixes_20260425_121355_06_grad_accum/checkpoints: <orbax.checkpoint.checkpoint_manager.CheckpointManager object at 0x7e09fc3362a0>
I0425 13:50:29.399986 138672244799296 checkpointing.py:302] Checkpoint manager created!
I0425 13:50:31.150587 138672244799296 checkpointing.py:578] checkpoint manager exists so trying to load this run's existing checkpoint
I0425 13:50:31.150715 138672244799296 checkpointing.py:676] No existing checkpoints found, not restoring checkpoint.
fsdp: 32
I0425 13:50:32.519670 138672244799296 maxtext_utils.py:1889]  decoder/decoder_norm/scale/value
    Shape:     float32[2048]
    Physical:  (None,)
I0425 13:50:32.519792 138672244799296 maxtext_utils.py:1889]  decoder/layers/mlp/wi_0/kernel/value
    Shape:     float32[2048,16,7168]
    Physical:  ('fsdp', None, None)
I0425 13:50:32.519844 138672244799296 maxtext_utils.py:1889]  decoder/layers/mlp/wi_1/kernel/value
    Shape:     float32[2048,16,7168]
    Physical:  ('fsdp', None, None)
I0425 13:50:32.519888 138672244799296 maxtext_utils.py:1889]  decoder/layers/mlp/wo/kernel/value
    Shape:     float32[7168,16,2048]
    Physical:  (None, None, 'fsdp')
I0425 13:50:32.519926 138672244799296 maxtext_utils.py:1889]  decoder/layers/post_self_attention_layer_norm/scale/value
    Shape:     float32[2048,16]
    Physical:  (None, None)
I0425 13:50:32.519960 138672244799296 maxtext_utils.py:1889]  decoder/layers/pre_self_attention_layer_norm/scale/value
    Shape:     float32[2048,16]
    Physical:  (None, None)
I0425 13:50:32.519996 138672244799296 maxtext_utils.py:1889]  decoder/layers/self_attention/key/kernel/value
    Shape:     float32[2048,16,16,128]
    Physical:  ('fsdp', None, None, None)
I0425 13:50:32.520033 138672244799296 maxtext_utils.py:1889]  decoder/layers/self_attention/out/kernel/value
    Shape:     float32[16,16,128,2048]
    Physical:  (None, None, None, 'fsdp')
I0425 13:50:32.520068 138672244799296 maxtext_utils.py:1889]  decoder/layers/self_attention/query/kernel/value
    Shape:     float32[2048,16,16,128]
    Physical:  ('fsdp', None, None, None)
I0425 13:50:32.520102 138672244799296 maxtext_utils.py:1889]  decoder/layers/self_attention/value/kernel/value
    Shape:     float32[2048,16,16,128]
    Physical:  ('fsdp', None, None, None)
I0425 13:50:32.520135 138672244799296 maxtext_utils.py:1889]  decoder/logits_dense/kernel/value
    Shape:     float32[2048,32000]
    Physical:  ('fsdp', None)

I0425 13:50:32.520166 138672244799296 maxtext_utils.py:1889]  token_embedder/embedding/value
    Shape:     float32[32000,2048]
    Physical:  (None, 'fsdp')
I0425 13:50:32.907552 138672244799296 gradient_accumulation.py:70] gradient_accumulation/inputs Logical: float32[2048]............................................... Unknown.
I0425 13:50:32.907646 138672244799296 gradient_accumulation.py:70] gradient_accumulation/inputs Physical: float32[2048]............................................... (None,).
I0425 13:50:32.922995 138672244799296 gradient_accumulation.py:70] gradient_accumulation/inputs Logical: float32[2048,16,7168]....................................... Unknown.
I0425 13:50:32.923060 138672244799296 gradient_accumulation.py:70] gradient_accumulation/inputs Physical: float32[2048,16,7168]....................................... ('fsdp', None, None).
I0425 13:50:32.952704 138672244799296 gradient_accumulation.py:70] gradient_accumulation/inputs Logical: float32[7168,16,2048]....................................... Unknown.
I0425 13:50:32.952769 138672244799296 gradient_accumulation.py:70] gradient_accumulation/inputs Physical: float32[7168,16,2048]....................................... (None, None, 'fsdp').
I0425 13:50:32.967584 138672244799296 gradient_accumulation.py:70] gradient_accumulation/inputs Logical: float32[2048,16]............................................ Unknown.
I0425 13:50:32.967642 138672244799296 gradient_accumulation.py:70] gradient_accumulation/inputs Physical: float32[2048,16]............................................ (None, None).
I0425 13:50:32.997481 138672244799296 gradient_accumulation.py:70] gradient_accumulation/inputs Logical: float32[2048,16,16,128]..................................... Unknown.
I0425 13:50:32.997551 138672244799296 gradient_accumulation.py:70] gradient_accumulation/inputs Physical: float32[2048,16,16,128]..................................... ('fsdp', None, None, None).
I0425 13:50:33.012447 138672244799296 gradient_accumulation.py:70] gradient_accumulation/inputs Logical: float32[16,16,128,2048]..................................... Unknown.
I0425 13:50:33.012502 138672244799296 gradient_accumulation.py:70] gradient_accumulation/inputs Physical: float32[16,16,128,2048]..................................... (None, None, None, 'fsdp').
I0425 13:50:33.057322 138672244799296 gradient_accumulation.py:70] gradient_accumulation/inputs Logical: float32[2048,32000]......................................... Unknown.
I0425 13:50:33.057396 138672244799296 gradient_accumulation.py:70] gradient_accumulation/inputs Physical: float32[2048,32000]......................................... ('fsdp', None).
I0425 13:50:33.072208 138672244799296 gradient_accumulation.py:70] gradient_accumulation/inputs Logical: float32[32000,2048]......................................... Unknown.
I0425 13:50:33.072274 138672244799296 gradient_accumulation.py:70] gradient_accumulation/inputs Physical: float32[32000,2048]......................................... (None, 'fsdp').
I0425 13:50:33.300643 138672244799296 nnx_decoders.py:465] nnx_decoders/carry Logical: bfloat16[32,2048,2048]...................................... ('activation_batch', 'activation_norm_length', 'activation_embed').
I0425 13:50:33.300801 138672244799296 nnx_decoders.py:465] nnx_decoders/carry Physical: bfloat16[32,2048,2048]...................................... ('fsdp', None, None).
I0425 13:50:33.307493 138672244799296 nnx_decoders.py:465] Unknown Logical: bfloat16[32,2048,2048]...................................... ('activation_batch', 'activation_norm_length', 'activation_embed').
I0425 13:50:33.307599 138672244799296 nnx_decoders.py:465] Unknown Physical: bfloat16[32,2048,2048]...................................... ('fsdp', None, None).
I0425 13:50:33.324306 138672244799296 attentions.py:1088] attentions/inputs_q Logical: bfloat16[32,2048,2048]...................................... ('activation_batch', 'activation_attn_length', 'activation_attn_embed').
I0425 13:50:33.324450 138672244799296 attentions.py:1088] attentions/inputs_q Physical: bfloat16[32,2048,2048]...................................... ('fsdp', None, None).
I0425 13:50:33.340541 138672244799296 attentions.py:1089] attentions/inputs_kv Logical: bfloat16[32,2048,2048]...................................... ('activation_batch', 'activation_attn_length', 'activation_attn_embed').
I0425 13:50:33.340683 138672244799296 attentions.py:1089] attentions/inputs_kv Physical: bfloat16[32,2048,2048]...................................... ('fsdp', None, None).
I0425 13:50:33.378011 138672244799296 attentions.py:1154] attentions/query Logical: bfloat16[32,2048,16,128].................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim').
I0425 13:50:33.378199 138672244799296 attentions.py:1154] attentions/query Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None).
I0425 13:50:33.394442 138672244799296 attentions.py:1155] attentions/key Logical: bfloat16[32,2048,16,128].................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim').
I0425 13:50:33.394580 138672244799296 attentions.py:1155] attentions/key Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None).
I0425 13:50:33.410554 138672244799296 attentions.py:1156] attentions/value Logical: bfloat16[32,2048,16,128].................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim').
I0425 13:50:33.410711 138672244799296 attentions.py:1156] attentions/value Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None).
I0425 13:50:33.443072 138672244799296 attentions.py:1198] attentions/out Logical: bfloat16[32,2048,16,128].................................... ('activation_batch', 'activation_attn_length', 'activation_heads', 'activation_kv').
I0425 13:50:33.443260 138672244799296 attentions.py:1198] attentions/out Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None).
I0425 13:50:33.465321 138672244799296 linears.py:525] linears/x Logical: bfloat16[32,2048,7168]...................................... ('activation_batch', 'activation_length', 'activation_mlp').
I0425 13:50:33.465430 138672244799296 linears.py:525] linears/x Physical: bfloat16[32,2048,7168]...................................... ('fsdp', None, None).
I0425 13:50:44.566596 138672244799296 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.
I0425 13:50:44.569016 138672244799296 metric_logger.py:301] number parameters: 1.104 billion
I0425 13:50:46.216036 138672244799296 checkpointing.py:794] Waiting for step 0 to finish before checkpoint...
I0425 13:50:58.182707 138672244799296 checkpointing.py:798] Waited 11.966617345809937 seconds for step 0 to finish before starting checkpointing.
I0425 13:50:58.185109 138672244799296 checkpoint_manager.py:2009] [process=5][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning.
I0425 13:50:58.186723 138672244799296 checkpoint_manager.py:1512] [process=5] Saving checkpoint at step 0
I0425 13:50:58.188163 138672244799296 event_tracking.py:70] [process=5] [async] Started save checkpoint @ gs://lance-maxtext/nnx_ckpt_xpk_feat_nnx_post_train_fixes_20260425_121355/nnx_xpk_feat_nnx_post_train_fixes_20260425_121355_06_grad_accum/checkpoints/0.
I0425 13:50:58.515727 138672244799296 signaling_client.py:364] Using JaxDistributedSignalingClient
I0425 13:50:58.516777 138672244799296 jax_array_handlers.py:360] Scheduling D2H of 69 prioritized jax.Array.
I0425 13:50:58.516834 138672244799296 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
I0425 13:50:58.790699 138672244799296 base_pytree_checkpoint_handler.py:154] [process=5][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 0.275746s
I0425 13:50:58.790869 138672244799296 base_pytree_checkpoint_handler.py:130] [process=5] /jax/orbax/write/blocking_gbytes_per_sec: 5.415 GiB/s (total gbytes: 1.5 GiB) (time elapsed: 0.28488922119140625 s) (per-host)
I0425 13:50:58.790920 138672244799296 base_pytree_checkpoint_handler.py:768] [process=5][thread=MainThread] Initiated Pytree async_save. Time taken: 0.284950s (batch_requests_ready=0.002942s, total_serialization_initiated=0.281940s, others=0.000068s)
I0425 13:50:58.791014 138672244799296 composite_checkpoint_handler.py:715] [process=5][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 0.288911s (all_items=0.000017s, per_item={'items': '0.00001693'}, temp_paths=0.288894)
I0425 13:50:58.791811 138672244799296 event_tracking.py:125] [process=5] [async] Finished blocking save in 0.60 seconds. Continuing save @ gs://lance-maxtext/nnx_ckpt_xpk_feat_nnx_post_train_fixes_20260425_121355/nnx_xpk_feat_nnx_post_train_fixes_20260425_121355_06_grad_accum/checkpoints/0.
I0425 13:50:58.792080 138543123035904 async_checkpointer.py:76] [process=5][thread=async_save] Background save thread started. Deadline for this save operation is 2026-04-25 14:10:58.792056
I0425 13:50:58.802561 138672244799296 checkpoint_manager.py:1560] [process=5][thread=MainThread][step=0] Starting CheckpointManager Save Finalize thread=save_finalize
I0425 13:50:58.802856 138546795648768 async_checkpointer.py:280] [process=5][thread=save_finalize] Waiting for background save thread=async_save.
I0425 13:50:58.803011 138672244799296 standard_logger.py:34] {'step': 0, 'event_type': 'save', 'directory': 'gs://lance-maxtext/nnx_ckpt_xpk_feat_nnx_post_train_fixes_20260425_121355/nnx_xpk_feat_nnx_post_train_fixes_20260425_121355_06_grad_accum/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1777125058.1850913, 'wait_for_prev_duration_secs': 5.8650970458984375e-05, 'time_between_consecutive_saves_sec': None, 'checkpointer_blocking_start_time': 1777125058.1867611, 'checkpointer_blocking_duration_secs': 0.6054284572601318, 'get_old_steps_start_time': 1777125058.7922127, 'get_old_steps_duration_secs': 2.8371810913085938e-05, 'checkpoint_manager_blocking_start_time': 1777125058.1830935, 'checkpoint_manager_blocking_duration_secs': 0.6198809146881104}
I0425 13:50:58.803121 138672244799296 checkpointing.py:409] Started an asynchronous checkpoint save for step 0
I0425 13:50:58.803186 138672244799296 max_utils.py:750] 
Memstats: After params initialized:
I0425 13:50:58.803234 138672244799296 max_utils.py:756] 	Using (GB) 0.43 / 31.25 (1.376000%) on TPU_18(process=5,(2,4,0,0))
I0425 13:50:58.803266 138672244799296 max_utils.py:756] 	Using (GB) 0.43 / 31.25 (1.376000%) on TPU_19(process=5,(3,4,0,0))
I0425 13:50:58.803292 138672244799296 max_utils.py:756] 	Using (GB) 0.43 / 31.25 (1.376000%) on TPU_22(process=5,(2,5,0,0))
I0425 13:50:58.803316 138672244799296 max_utils.py:756] 	Using (GB) 0.43 / 31.25 (1.376000%) on TPU_23(process=5,(3,5,0,0))
I0425 13:50:59.112791 138672244799296 metric_logger.py:196] completed step: 0, seconds: 1.646, TFLOP/s/device: 33.019, Tokens/s/device: 4977.041, total_weights: 262144, loss: 10.876, lm_loss: 10.876, perplexity: 52883.598
I0425 13:50:59.718693 138672244799296 metric_logger.py:196] completed step: 1, seconds: 12.895, TFLOP/s/device: 4.215, Tokens/s/device: 635.293, total_weights: 262144, loss: 10.876, lm_loss: 10.876, perplexity: 52883.598
I0425 13:51:00.297227 138672244799296 metric_logger.py:196] completed step: 2, seconds: 0.029, TFLOP/s/device: 1874.863, Tokens/s/device: 282599.696, total_weights: 262144, loss: 10.561, lm_loss: 10.561, perplexity: 38601.980
I0425 13:51:00.875505 138672244799296 metric_logger.py:196] completed step: 3, seconds: 0.584, TFLOP/s/device: 93.026, Tokens/s/device: 14021.899, total_weights: 262144, loss: 10.270, lm_loss: 10.270, perplexity: 28865.557
I0425 13:51:02.032714 138672244799296 metric_logger.py:196] completed step: 4, seconds: 0.579, TFLOP/s/device: 93.860, Tokens/s/device: 14147.628, total_weights: 262144, loss: 10.021, lm_loss: 10.021, perplexity: 22497.775
I0425 13:51:02.039813 138672244799296 metric_logger.py:196] completed step: 5, seconds: 0.578, TFLOP/s/device: 94.030, Tokens/s/device: 14173.182, total_weights: 262144, loss: 9.819, lm_loss: 9.819, perplexity: 18383.324
I0425 13:51:02.658630    2584 google_auth_provider.cc:181] Running on GCE, using service account 562977990677-compute@developer.gserviceaccount.com
I0425 13:51:04.688260 138546813982464 array_metadata_store.py:203] [process=5][thread=array_type_handler] Wrote 69 array_metadata.ArrayMetadata to gs://lance-maxtext/nnx_ckpt_xpk_feat_nnx_post_train_fixes_20260425_121355/nnx_xpk_feat_nnx_post_train_fixes_20260425_121355_06_grad_accum/checkpoints/0/items/array_metadatas/process_5
I0425 13:51:37.435819 138543123035904 base_pytree_checkpoint_handler.py:130] [process=5] /jax/orbax/write/gbytes_per_sec: 40.575 MiB/s (total gbytes: 1.5 GiB) (time elapsed: 38.92981839179993 s) (per-host)
I0425 13:51:37.435931 138543123035904 async_checkpointer.py:90] [process=5][thread=async_save] 3 Handler Commit operations completed. Time taken: 38.643771s.
I0425 13:51:45.470040 138543123035904 async_checkpointer.py:160] [process=5][thread=async_save] Background save thread done. Time taken: 46.677864s.
I0425 13:51:45.470363 138546795648768 async_checkpointer.py:288] [process=5][thread=save_finalize] Done with waiting for background save thread=async_save.
I0425 13:51:45.470484 138546795648768 async_checkpointer.py:298] [process=5][thread=save_finalize] No errors found in background save thread=async_save.
I0425 13:51:45.470537 138546795648768 checkpoint_manager.py:2137] [process=5][thread=save_finalize][step=0] CheckpointManager Save Finalize is syncing with other hosts...
I0425 13:51:45.570288 138546795648768 checkpoint_manager.py:2146] [process=5][thread=save_finalize][step=0] CheckpointManager Save Finalize is done on all hosts.
I0425 13:51:47.319409 138672244799296 metric_logger.py:196] completed step: 6, seconds: 1.158, TFLOP/s/device: 46.947, Tokens/s/device: 7076.386, total_weights: 262144, loss: 9.666, lm_loss: 9.666, perplexity: 15775.117
I0425 13:51:47.897933 138672244799296 metric_logger.py:196] completed step: 7, seconds: 44.702, TFLOP/s/device: 1.216, Tokens/s/device: 183.257, total_weights: 262144, loss: 9.560, lm_loss: 9.560, perplexity: 14192.143
I0425 13:51:48.476548 138672244799296 metric_logger.py:196] completed step: 8, seconds: 0.584, TFLOP/s/device: 93.039, Tokens/s/device: 14023.819, total_weights: 262144, loss: 9.495, lm_loss: 9.495, perplexity: 13292.714
I0425 13:51:49.055621 138672244799296 checkpointing.py:794] Waiting for step 9 to finish before checkpoint...
I0425 13:51:49.056262 138672244799296 checkpointing.py:798] Waited 0.0006530284881591797 seconds for step 9 to finish before starting checkpointing.
I0425 13:51:49.058477 138672244799296 checkpoint_manager.py:2009] [process=5][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning.
I0425 13:51:49.060171 138672244799296 checkpoint_manager.py:1512] [process=5] Saving checkpoint at step 9
I0425 13:51:49.061512 138672244799296 event_tracking.py:70] [process=5] [async] Started save checkpoint @ gs://lance-maxtext/nnx_ckpt_xpk_feat_nnx_post_train_fixes_20260425_121355/nnx_xpk_feat_nnx_post_train_fixes_20260425_121355_06_grad_accum/checkpoints/9.
I0425 13:51:49.363119 138672244799296 jax_array_handlers.py:360] Scheduling D2H of 69 prioritized jax.Array.
I0425 13:51:49.363209 138672244799296 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
I0425 13:51:49.404022 138672244799296 base_pytree_checkpoint_handler.py:154] [process=5][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 0.042538s
I0425 13:51:49.404185 138672244799296 base_pytree_checkpoint_handler.py:130] [process=5] /jax/orbax/write/blocking_gbytes_per_sec: 31.172 GiB/s (total gbytes: 1.5 GiB) (time elapsed: 0.04948616027832031 s) (per-host)
I0425 13:51:49.404235 138672244799296 base_pytree_checkpoint_handler.py:768] [process=5][thread=MainThread] Initiated Pytree async_save. Time taken: 0.049545s (batch_requests_ready=0.002588s, total_serialization_initiated=0.046891s, others=0.000066s)
I0425 13:51:49.404336 138672244799296 composite_checkpoint_handler.py:715] [process=5][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 0.053647s (all_items=0.000015s, per_item={'items': '0.00001550'}, temp_paths=0.053632)
I0425 13:51:49.405013 138672244799296 event_tracking.py:125] [process=5] [async] Finished blocking save in 0.34 seconds. Continuing save @ gs://lance-maxtext/nnx_ckpt_xpk_feat_nnx_post_train_fixes_20260425_121355/nnx_xpk_feat_nnx_post_train_fixes_20260425_121355_06_grad_accum/checkpoints/9.
I0425 13:51:49.405365 138546795648768 async_checkpointer.py:76] [process=5][thread=async_save] Background save thread started. Deadline for this save operation is 2026-04-25 14:11:49.405327
I0425 13:51:49.407254 138672244799296 checkpoint_manager.py:1560] [process=5][thread=MainThread][step=9] Starting CheckpointManager Save Finalize thread=save_finalize
I0425 13:51:49.407563 138546813982464 async_checkpointer.py:280] [process=5][thread=save_finalize] Waiting for background save thread=async_save.
I0425 13:51:49.407711 138672244799296 standard_logger.py:34] {'step': 9, 'event_type': 'save', 'directory': 'gs://lance-maxtext/nnx_ckpt_xpk_feat_nnx_post_train_fixes_20260425_121355/nnx_xpk_feat_nnx_post_train_fixes_20260425_121355_06_grad_accum/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1777125109.0584435, 'wait_for_prev_duration_secs': 7.510185241699219e-05, 'time_between_consecutive_saves_sec': 3.4880623817443848, 'checkpointer_blocking_start_time': 1777125109.0602157, 'checkpointer_blocking_duration_secs': 0.3452930450439453, 'get_old_steps_start_time': 1777125109.4055321, 'get_old_steps_duration_secs': 3.0279159545898438e-05, 'checkpoint_manager_blocking_start_time': 1777125109.0564854, 'checkpoint_manager_blocking_duration_secs': 0.35118937492370605}
I0425 13:51:49.407884 138672244799296 checkpointing.py:409] Started an asynchronous checkpoint save for step 9
I0425 13:51:49.407930 138672244799296 checkpoint_manager.py:2020] [process=5][thread=MainThread][step=9][wait_until_finished] Waiting for Save Finalize thread (save_finalize) to complete.
I0425 13:51:54.455385 138547390576384 array_metadata_store.py:203] [process=5][thread=array_type_handler] Wrote 69 array_metadata.ArrayMetadata to gs://lance-maxtext/nnx_ckpt_xpk_feat_nnx_post_train_fixes_20260425_121355/nnx_xpk_feat_nnx_post_train_fixes_20260425_121355_06_grad_accum/checkpoints/9/items/array_metadatas/process_5
I0425 13:52:31.786423 138546795648768 base_pytree_checkpoint_handler.py:130] [process=5] /jax/orbax/write/gbytes_per_sec: 37.227 MiB/s (total gbytes: 1.5 GiB) (time elapsed: 42.431676149368286 s) (per-host)
I0425 13:52:31.786555 138546795648768 async_checkpointer.py:90] [process=5][thread=async_save] 3 Handler Commit operations completed. Time taken: 42.381074s.
I0425 13:52:39.544386 138546795648768 async_checkpointer.py:160] [process=5][thread=async_save] Background save thread done. Time taken: 50.138890s.
I0425 13:52:39.544697 138546813982464 async_checkpointer.py:288] [process=5][thread=save_finalize] Done with waiting for background save thread=async_save.
I0425 13:52:39.544832 138546813982464 async_checkpointer.py:298] [process=5][thread=save_finalize] No errors found in background save thread=async_save.
I0425 13:52:39.544892 138546813982464 checkpoint_manager.py:2137] [process=5][thread=save_finalize][step=9] CheckpointManager Save Finalize is syncing with other hosts...
I0425 13:52:39.546713 138546813982464 checkpoint_manager.py:2146] [process=5][thread=save_finalize][step=9] CheckpointManager Save Finalize is done on all hosts.
I0425 13:52:39.546887 138672244799296 checkpoint_manager.py:2032] [process=5][thread=MainThread][step=9][wait_until_finished] Done waiting for Save Finalize thread (save_finalize) running at step=9.
I0425 13:52:39.547074 138672244799296 checkpoint_manager.py:2009] [process=5][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning.
I0425 13:52:39.547955 138672244799296 metric_logger.py:196] completed step: 9, seconds: 0.579, TFLOP/s/device: 93.876, Tokens/s/device: 14149.949, total_weights: 262144, loss: 9.457, lm_loss: 9.457, perplexity: 12793.173
Per train step:
 Total TFLOPs: 54.35 
 split as 93.93% learnable weight flops and 6.07% attention flops
XPK End: Sat Apr 25 13:52:48 UTC 2026
EXIT_CODE=0