Compare: 05_fp8 — MaxView

‹ 04_int8Case: 05_fp806_grad_accum ›

Metrics: `main (e27fc1e97)` vs `feat/nnx-trainstate-and-training-loop (e27fc1e97)`

Metric	main `e27fc1e97`	feat/nnx-trainstate-and-training-loop `e27fc1e97`	Diff (feat/nnx-trainstate-and-training-loop − main)
Parameters	1.105 billion	1.105 billion	—
Final loss	6.1670	6.1670	0
TFLOP/s	85.984	85.808	-0.176
Tok/s	12960.5	12933.9	-26.601
Avg s/step	1.610	1.938	+0.328
Memory %	1.44	1.44	0
JAX	0.9.2	0.9.2	—

Diff = branch value − main value. Green = branch improved. Red = branch regressed.

Field	main	feat/nnx-trainstate-and-training-loop
JAX	0.9.2	0.9.2
Jaxlib	0.9.2	0.9.2
Hardware	PJRT C API / TFRT TPU v6 lite	PJRT C API / TFRT TPU v6 lite
Devices	32	32

Memory Breakdown

Field	main	feat/nnx-trainstate-and-training-loop
Total	1.5 GB	1.5 GB
Output	0.4 GB	0.4 GB
Temp	1.1 GB	1.1 GB
Argument	0.4 GB	0.4 GB

main

Device	Used (GB)	Total (GB)	%
TPU_26(process=7,(2,6,0,0))	0.45	31.25	1.4%
TPU_27(process=7,(3,6,0,0))	0.45	31.25	1.4%
TPU_30(process=7,(2,7,0,0))	0.45	31.25	1.4%
TPU_31(process=7,(3,7,0,0))	0.45	31.25	1.4%

feat/nnx-trainstate-and-training-loop

Device	Used (GB)	Total (GB)	%
TPU_10(process=3,(2,2,0,0))	0.45	31.25	1.4%
TPU_11(process=3,(3,2,0,0))	0.45	31.25	1.4%
TPU_14(process=3,(2,3,0,0))	0.45	31.25	1.4%
TPU_15(process=3,(3,3,0,0))	0.45	31.25	1.4%

main TFLOP/s

feat/nnx-trainstate-and-training-loop TFLOP/s

Step	main Loss	feat/nnx-trainstate-and-training-loop Loss	main TFLOP/s	feat/nnx-trainstate-and-training-loop TFLOP/s	TFLOP/s Δ
0	10.874	10.874	0.883	0.891	+0.0080
1	10.874	10.874	8.263	10.745	+2.4820
2	10.816	10.816	1200.701	438.181	-762.5200
3	10.431	10.431	32.398	31.965	-0.4330
4	9.992	9.992	82.752	82.667	-0.0850
5	9.549	9.549	85.756	85.832	+0.0760
6	9.111	9.111	42.813	42.693	-0.1200
7	8.685	8.685	0.570	0.560	-0.0100
8	8.281	8.281	82.902	83.193	+0.2910
9	7.908	7.908	85.864	85.819	-0.0450
10	7.568	7.568	85.631	85.596	-0.0350
11	7.288	7.288	6.600	1.655	-4.9450
12	7.032	7.032	2037.356	2087.117	+49.7610
13	6.815	6.815	24.217	85.593	+61.3760
14	6.635	6.635	83.040	24.413	-58.6270
15	6.492	6.492	85.888	83.199	-2.6890
16	6.380	6.380	85.913	85.757	-0.1560
17	6.293	6.293	85.849	85.894	+0.0450
18	6.223	6.223	85.493	85.788	+0.2950
19	6.167	6.167	85.984	85.808	-0.1760

main · e27fc1e97 · main_20260421_144122 · full log

XPK Start: Tue Apr 21 14:55:37 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. 
2026-04-21 14:56:02.599470: E external/local_xla/xla/stream_executor/cuda/cuda_platform.cc:51] failed call to cuInit: INTERNAL: CUDA error: Failed call to cuInit: UNKNOWN ERROR (303)
I0421 14:56:02.809980 133091194312512 max_utils.py:273] Attempting to initialize the jax distributed system...
I0421 14:56:11.852323 133091194312512 distributed.py:149] Starting JAX distributed service on [::]:8482
I0421 14:56:11.854670 133091194312512 distributed.py:172] Connecting to JAX distributed service on mt-05-fp8-otau4-slice-job-0-0.mt-05-fp8-otau4:8482
I0421 14:56:13.679855 133091194312512 max_utils.py:284] Jax distributed system initialized!
I0421 14:56:18.664420 133091194312512 max_utils.py:800] System Information: Jax Version: 0.9.2
I0421 14:56:18.664527 133091194312512 max_utils.py:801] System Information: Jaxlib Version: 0.9.2
I0421 14:56:18.664570 133091194312512 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
I0421 14:56:18.664617 133091194312512 train_utils.py:361] WARNING: Sequence packing is essentially ignored for synthetic data. Please use a real dataset to use sequence packing.
I0421 14:56:19.418926 133091194312512 maxtext_utils.py:1565] Num_devices: 32, shape (1, 1, 1, 32, 1, 1, 1, 1, 1, 1, 1, 1, 1)
I0421 14:56:19.419216 133091194312512 checkpointing.py:677] Setting up checkpoint logger...
I0421 14:56:19.419269 133091194312512 checkpointing.py:233] Creating checkpoint manager with ocdbt=True and zarr3=True
I0421 14:56:19.419315 133091194312512 pytree_checkpoint_handler.py:592] save_device_host_concurrent_bytes=None
I0421 14:56:19.419653 133091194312512 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 0x790afdc3c470>, enable_pinned_host_transfer=False, save_concurrent_bytes: 96000000000 (89.4 GiB), restore_concurrent_bytes: 96000000000 (89.4 GiB)
I0421 14:56:22.316349 133091194312512 checkpointing.py:265] Enabling policy for fixed interval checkpointing.
I0421 14:56:22.316587 133091194312512 checkpoint_manager.py:708] [process=7][thread=MainThread] CheckpointManager init: checkpointers=None, item_names=('items',), item_handlers={'items': <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7909f1768230>}, handler_registry=None
I0421 14:56:22.316845 133091194312512 composite_checkpoint_handler.py:237] Deferred registration for item: "items". Adding handler `<orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7909f1768230>` 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`.
I0421 14:56:22.316895 133091194312512 composite_checkpoint_handler.py:237] Deferred registration for item: "metrics". Adding handler `<orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x78f780708110>` for item "metrics" and save args `<class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonSaveArgs'>` and restore args `<class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonRestoreArgs'>` to `_handler_registry`.
I0421 14:56:22.316932 133091194312512 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 0x7909f1768230>, ('items', <class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeRestoreArgs'>): <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7909f1768230>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonSaveArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x78f780708110>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonRestoreArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x78f780708110>}).
I0421 14:56:22.317248 133091194312512 abstract_checkpointer.py:35] orbax-checkpoint version: 0.11.34
I0421 14:56:22.317321 133091194312512 async_checkpointer.py:192] [process=7][thread=MainThread] Using barrier_sync_fn: <function get_barrier_sync_fn.<locals>._fn at 0x78f7584ed580> timeout: 1200 secs and primary_host=0 for async checkpoint writes
I0421 14:56:23.738069 133091194312512 checkpoint_manager.py:1812] Found 0 checkpoint steps in gs://lance-maxtext/linen_ckpt_xpk_main_20260421_144122/linen_xpk_main_20260421_144122_05_fp8/checkpoints
I0421 14:56:24.154440 133091194312512 checkpoint_manager.py:929] [process=7][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_20260421_144122/linen_xpk_main_20260421_144122_05_fp8/checkpoints: <orbax.checkpoint.checkpoint_manager.CheckpointManager object at 0x78f780709610>
I0421 14:56:24.154617 133091194312512 checkpointing.py:301] Checkpoint manager created!
I0421 14:56:25.238074 133091194312512 nnx_wrappers.py:437] Unknown Logical: bfloat16[32,2048,2048]...................................... ('activation_batch', 'activation_norm_length', 'activation_embed').
I0421 14:56:25.238193 133091194312512 nnx_wrappers.py:437] Unknown Physical: bfloat16[32,2048,2048]...................................... ('fsdp', None, None).
I0421 14:56:25.621022 133091194312512 attentions.py:1088] attentions/inputs_q Logical: bfloat16[32,2048,2048]...................................... ('activation_batch', 'activation_attn_length', 'activation_attn_embed').
I0421 14:56:25.621117 133091194312512 attentions.py:1088] attentions/inputs_q Physical: bfloat16[32,2048,2048]...................................... ('fsdp', None, None).
I0421 14:56:25.637345 133091194312512 attentions.py:1089] attentions/inputs_kv Logical: bfloat16[32,2048,2048]...................................... ('activation_batch', 'activation_attn_length', 'activation_attn_embed').
I0421 14:56:25.637404 133091194312512 attentions.py:1089] attentions/inputs_kv Physical: bfloat16[32,2048,2048]...................................... ('fsdp', None, None).
I0421 14:56:25.693845 133091194312512 attentions.py:1154] attentions/query Logical: bfloat16[32,2048,16,128].................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim').
I0421 14:56:25.693944 133091194312512 attentions.py:1154] attentions/query Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None).
I0421 14:56:25.710432 133091194312512 attentions.py:1155] attentions/key Logical: bfloat16[32,2048,16,128].................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim').
I0421 14:56:25.710497 133091194312512 attentions.py:1155] attentions/key Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None).
I0421 14:56:25.726932 133091194312512 attentions.py:1156] attentions/value Logical: bfloat16[32,2048,16,128].................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim').
I0421 14:56:25.726993 133091194312512 attentions.py:1156] attentions/value Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None).
I0421 14:56:25.752547 133091194312512 attentions.py:1198] attentions/out Logical: bfloat16[32,2048,16,128].................................... ('activation_batch', 'activation_attn_length', 'activation_heads', 'activation_kv').
I0421 14:56:25.752617 133091194312512 attentions.py:1198] attentions/out Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None).
I0421 14:56:25.804034 133091194312512 linears.py:525] linears/x Logical: bfloat16[32,2048,7168]...................................... ('activation_batch', 'activation_length', 'activation_mlp').
I0421 14:56:25.804119 133091194312512 linears.py:525] linears/x Physical: bfloat16[32,2048,7168]...................................... ('fsdp', None, None).
I0421 14:56:26.238978 133091194312512 checkpointing.py:577] checkpoint manager exists so trying to load this run's existing checkpoint
I0421 14:56:26.239100 133091194312512 checkpointing.py:665] No existing checkpoints found, not restoring checkpoint.
fsdp: 32
I0421 14:56:28.166382 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_0/Fp8DirectDotGeneralOp_0/input_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.166503 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_0/Fp8DirectDotGeneralOp_0/input_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.166553 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_0/Fp8DirectDotGeneralOp_0/kernel_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.166598 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_0/Fp8DirectDotGeneralOp_0/kernel_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.166635 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_0/Fp8DirectDotGeneralOp_0/output_grad_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.166669 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_0/Fp8DirectDotGeneralOp_0/output_grad_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.166701 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_1/Fp8DirectDotGeneralOp_0/input_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.166745 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_1/Fp8DirectDotGeneralOp_0/input_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.166782 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_1/Fp8DirectDotGeneralOp_0/kernel_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.166813 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_1/Fp8DirectDotGeneralOp_0/kernel_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.166843 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_1/Fp8DirectDotGeneralOp_0/output_grad_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.166873 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_1/Fp8DirectDotGeneralOp_0/output_grad_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.166905 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/mlp/wo/Fp8DirectDotGeneralOp_0/input_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.166934 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/mlp/wo/Fp8DirectDotGeneralOp_0/input_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.166964 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/mlp/wo/Fp8DirectDotGeneralOp_0/kernel_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.166994 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/mlp/wo/Fp8DirectDotGeneralOp_0/kernel_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167021 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/mlp/wo/Fp8DirectDotGeneralOp_0/output_grad_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167055 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/mlp/wo/Fp8DirectDotGeneralOp_0/output_grad_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167084 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/self_attention/key/Fp8DirectDotGeneralOp_0/input_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167111 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/self_attention/key/Fp8DirectDotGeneralOp_0/input_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167138 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/self_attention/key/Fp8DirectDotGeneralOp_0/kernel_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167165 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/self_attention/key/Fp8DirectDotGeneralOp_0/kernel_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167191 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/self_attention/key/Fp8DirectDotGeneralOp_0/output_grad_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167218 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/self_attention/key/Fp8DirectDotGeneralOp_0/output_grad_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167245 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/self_attention/out/Fp8DirectDotGeneralOp_0/input_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167271 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/self_attention/out/Fp8DirectDotGeneralOp_0/input_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167298 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/self_attention/out/Fp8DirectDotGeneralOp_0/kernel_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167324 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/self_attention/out/Fp8DirectDotGeneralOp_0/kernel_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167351 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/self_attention/out/Fp8DirectDotGeneralOp_0/output_grad_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167377 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/self_attention/out/Fp8DirectDotGeneralOp_0/output_grad_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167403 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/self_attention/query/Fp8DirectDotGeneralOp_0/input_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167430 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/self_attention/query/Fp8DirectDotGeneralOp_0/input_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167458 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/self_attention/query/Fp8DirectDotGeneralOp_0/kernel_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167485 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/self_attention/query/Fp8DirectDotGeneralOp_0/kernel_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167513 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/self_attention/query/Fp8DirectDotGeneralOp_0/output_grad_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167547 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/self_attention/query/Fp8DirectDotGeneralOp_0/output_grad_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167574 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/self_attention/value/Fp8DirectDotGeneralOp_0/input_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167601 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/self_attention/value/Fp8DirectDotGeneralOp_0/input_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167627 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/self_attention/value/Fp8DirectDotGeneralOp_0/kernel_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167654 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/self_attention/value/Fp8DirectDotGeneralOp_0/kernel_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167682 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/self_attention/value/Fp8DirectDotGeneralOp_0/output_grad_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167709 133091194312512 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/self_attention/value/Fp8DirectDotGeneralOp_0/output_grad_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 14:56:28.167790 133091194312512 maxtext_utils.py:1668]  params/params/decoder/decoder_norm/scale
    Shape:     float32[2048]
    Logical:   P('norm',)
    Physical:  (None,)
I0421 14:56:28.167850 133091194312512 maxtext_utils.py:1668]  params/params/decoder/layers/mlp/wi_0/kernel
    Shape:     float32[2048,16,7168]
    Logical:   P('embed', 'layers', 'mlp')
    Physical:  ('fsdp', None, None)
I0421 14:56:28.167889 133091194312512 maxtext_utils.py:1668]  params/params/decoder/layers/mlp/wi_1/kernel
    Shape:     float32[2048,16,7168]
    Logical:   P('embed', 'layers', 'mlp')
    Physical:  ('fsdp', None, None)
I0421 14:56:28.167936 133091194312512 maxtext_utils.py:1668]  params/params/decoder/layers/mlp/wo/kernel
    Shape:     float32[7168,16,2048]
    Logical:   P('mlp', 'layers', 'embed')
    Physical:  (None, None, 'fsdp')
I0421 14:56:28.167982 133091194312512 maxtext_utils.py:1668]  params/params/decoder/layers/post_self_attention_layer_norm/scale
    Shape:     float32[2048,16]
    Logical:   P('norm', 'layers')
    Physical:  (None, None)
I0421 14:56:28.168016 133091194312512 maxtext_utils.py:1668]  params/params/decoder/layers/pre_self_attention_layer_norm/scale
    Shape:     float32[2048,16]
    Logical:   P('norm', 'layers')
    Physical:  (None, None)
I0421 14:56:28.168068 133091194312512 maxtext_utils.py:1668]  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)
I0421 14:56:28.168117 133091194312512 maxtext_utils.py:1668]  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')
I0421 14:56:28.168155 133091194312512 maxtext_utils.py:1668]  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)
I0421 14:56:28.168189 133091194312512 maxtext_utils.py:1668]  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)
I0421 14:56:28.168234 133091194312512 maxtext_utils.py:1668]  params/params/decoder/logits_dense/kernel
    Shape:     float32[2048,32000]
    Logical:   P('embed_vocab', 'vocab')
    Physical:  ('fsdp', None)
I0421 14:56:28.168278 133091194312512 maxtext_utils.py:1668]  params/params/token_embedder/embedding
    Shape:     float32[32000,2048]
    Logical:   P('vocab', 'embed_vocab')
    Physical:  (None, 'fsdp')

I0421 14:56:29.787888 133091194312512 train.py:155] train/xent Logical: float32[32,2048]............................................ ('activation_embed_and_logits_batch', 'activation_length').
I0421 14:56:29.787982 133091194312512 train.py:155] train/xent Physical: float32[32,2048]............................................ ('fsdp', None).
I0421 14:56:29.803319 133091194312512 train.py:162] train/z_loss Logical: float32[32,2048]............................................ ('activation_embed_and_logits_batch', 'activation_length').
I0421 14:56:29.803380 133091194312512 train.py:162] train/z_loss Physical: float32[32,2048]............................................ ('fsdp', None).
I0421 14:56:43.642947 133091194312512 max_utils.py:791] Total memory size: 1.5 GB, Output size: 0.4 GB, Temp size: 1.1 GB, Argument size: 0.4 GB, Host temp size: 0.0 GB.
I0421 14:56:43.643857 133091194312512 metric_logger.py:301] number parameters: 1.105 billion
I0421 14:56:59.029362 133091194312512 checkpointing.py:772] Waiting for step 0 to finish before checkpoint...
I0421 14:56:59.201074 133091194312512 checkpointing.py:776] Waited 0.1716902256011963 seconds for step 0 to finish before starting checkpointing.
I0421 14:56:59.203573 133091194312512 checkpoint_manager.py:2009] [process=7][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning.
I0421 14:56:59.205662 133091194312512 checkpoint_manager.py:1512] [process=7] Saving checkpoint at step 0
I0421 14:56:59.207097 133091194312512 event_tracking.py:70] [process=7] [async] Started save checkpoint @ gs://lance-maxtext/linen_ckpt_xpk_main_20260421_144122/linen_xpk_main_20260421_144122_05_fp8/checkpoints/0.
I0421 14:56:59.911985 133091194312512 signaling_client.py:364] Using JaxDistributedSignalingClient
I0421 14:56:59.912912 133091194312512 jax_array_handlers.py:360] Scheduling D2H of 81 prioritized jax.Array.
I0421 14:56:59.912971 133091194312512 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
I0421 14:57:00.324105 133091194312512 base_pytree_checkpoint_handler.py:154] [process=7][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 0.412244s
I0421 14:57:00.324281 133091194312512 base_pytree_checkpoint_handler.py:130] [process=7] /jax/orbax/write/blocking_gbytes_per_sec: 3.668 GiB/s (total gbytes: 1.5 GiB) (time elapsed: 0.42064666748046875 s) (per-host)
I0421 14:57:00.324334 133091194312512 base_pytree_checkpoint_handler.py:768] [process=7][thread=MainThread] Initiated Pytree async_save. Time taken: 0.420709s (batch_requests_ready=0.003581s, total_serialization_initiated=0.417058s, others=0.000070s)
I0421 14:57:00.324441 133091194312512 composite_checkpoint_handler.py:715] [process=7][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 0.425167s (all_items=0.000019s, per_item={'items': '0.00001860'}, temp_paths=0.425148)
I0421 14:57:00.325308 133091194312512 event_tracking.py:125] [process=7] [async] Finished blocking save in 1.12 seconds. Continuing save @ gs://lance-maxtext/linen_ckpt_xpk_main_20260421_144122/linen_xpk_main_20260421_144122_05_fp8/checkpoints/0.
I0421 14:57:00.325613 132963037005568 async_checkpointer.py:76] [process=7][thread=async_save] Background save thread started. Deadline for this save operation is 2026-04-21 15:17:00.325580
I0421 14:57:00.345492 133091194312512 checkpoint_manager.py:1560] [process=7][thread=MainThread][step=0] Starting CheckpointManager Save Finalize thread=save_finalize
I0421 14:57:00.345858 132963011827456 async_checkpointer.py:280] [process=7][thread=save_finalize] Waiting for background save thread=async_save.
I0421 14:57:00.346021 133091194312512 standard_logger.py:34] {'step': 0, 'event_type': 'save', 'directory': 'gs://lance-maxtext/linen_ckpt_xpk_main_20260421_144122/linen_xpk_main_20260421_144122_05_fp8/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776783419.2035546, 'wait_for_prev_duration_secs': 6.151199340820312e-05, 'time_between_consecutive_saves_sec': None, 'checkpointer_blocking_start_time': 1776783419.2057004, 'checkpointer_blocking_duration_secs': 1.1201088428497314, 'get_old_steps_start_time': 1776783420.325836, 'get_old_steps_duration_secs': 3.0517578125e-05, 'checkpoint_manager_blocking_start_time': 1776783419.2017052, 'checkpoint_manager_blocking_duration_secs': 1.1442756652832031}
I0421 14:57:00.346129 133091194312512 checkpointing.py:408] Started an asynchronous checkpoint save for step 0
I0421 14:57:00.346179 133091194312512 max_utils.py:750] 
Memstats: After params initialized:
I0421 14:57:00.346231 133091194312512 max_utils.py:756] 	Using (GB) 0.45 / 31.25 (1.440000%) on TPU_26(process=7,(2,6,0,0))
I0421 14:57:00.346264 133091194312512 max_utils.py:756] 	Using (GB) 0.45 / 31.25 (1.440000%) on TPU_27(process=7,(3,6,0,0))
I0421 14:57:00.346291 133091194312512 max_utils.py:756] 	Using (GB) 0.45 / 31.25 (1.440000%) on TPU_30(process=7,(2,7,0,0))
I0421 14:57:00.346316 133091194312512 max_utils.py:756] 	Using (GB) 0.45 / 31.25 (1.440000%) on TPU_31(process=7,(3,7,0,0))
I0421 14:57:00.675819 133091194312512 metric_logger.py:196] completed step: 0, seconds: 15.385, TFLOP/s/device: 0.883, Tokens/s/device: 133.113, total_weights: 65536, loss: 10.874, lm_loss: 10.874, perplexity: 52776.805
I0421 14:57:00.844683 133091194312512 metric_logger.py:196] completed step: 1, seconds: 1.644, TFLOP/s/device: 8.263, Tokens/s/device: 1245.436, total_weights: 65536, loss: 10.874, lm_loss: 10.874, perplexity: 52761.840
I0421 14:57:01.263677 133091194312512 metric_logger.py:196] completed step: 2, seconds: 0.011, TFLOP/s/device: 1200.701, Tokens/s/device: 180982.679, total_weights: 65536, loss: 10.816, lm_loss: 10.816, perplexity: 49832.398
I0421 14:57:01.421944 133091194312512 metric_logger.py:196] completed step: 3, seconds: 0.419, TFLOP/s/device: 32.398, Tokens/s/device: 4883.434, total_weights: 65536, loss: 10.431, lm_loss: 10.431, perplexity: 33901.820
I0421 14:57:01.739198 133091194312512 metric_logger.py:196] completed step: 4, seconds: 0.164, TFLOP/s/device: 82.752, Tokens/s/device: 12473.278, total_weights: 65536, loss: 9.992, lm_loss: 9.992, perplexity: 21847.639
I0421 14:57:01.745799 133091194312512 metric_logger.py:196] completed step: 5, seconds: 0.158, TFLOP/s/device: 85.756, Tokens/s/device: 12926.029, total_weights: 65536, loss: 9.549, lm_loss: 9.549, perplexity: 14035.006
I0421 14:57:03.320193    2782 google_auth_provider.cc:181] Running on GCE, using service account 562977990677-compute@developer.gserviceaccount.com
I0421 14:57:05.371107 132963020220160 array_metadata_store.py:203] [process=7][thread=array_type_handler] Wrote 81 array_metadata.ArrayMetadata to gs://lance-maxtext/linen_ckpt_xpk_main_20260421_144122/linen_xpk_main_20260421_144122_05_fp8/checkpoints/0/items/array_metadatas/process_7
I0421 14:57:20.218407 132963037005568 base_pytree_checkpoint_handler.py:130] [process=7] /jax/orbax/write/gbytes_per_sec: 77.764 MiB/s (total gbytes: 1.5 GiB) (time elapsed: 20.314732313156128 s) (per-host)
I0421 14:57:20.218532 132963037005568 async_checkpointer.py:90] [process=7][thread=async_save] 3 Handler Commit operations completed. Time taken: 19.892804s.
I0421 14:57:25.722051 133091194312512 metric_logger.py:196] completed step: 6, seconds: 0.317, TFLOP/s/device: 42.813, Tokens/s/device: 6453.239, total_weights: 65536, loss: 9.111, lm_loss: 9.111, perplexity: 9058.688
I0421 14:57:25.880483 133091194312512 metric_logger.py:196] completed step: 7, seconds: 23.819, TFLOP/s/device: 0.570, Tokens/s/device: 85.983, total_weights: 65536, loss: 8.685, lm_loss: 8.685, perplexity: 5914.720
I0421 14:57:26.039039 133091194312512 metric_logger.py:196] completed step: 8, seconds: 0.164, TFLOP/s/device: 82.902, Tokens/s/device: 12495.881, total_weights: 65536, loss: 8.281, lm_loss: 8.281, perplexity: 3950.010
I0421 14:57:26.044228 133091194312512 checkpointing.py:772] Waiting for step 10 to finish before checkpoint...
I0421 14:57:26.355916 133091194312512 checkpointing.py:776] Waited 0.31165385246276855 seconds for step 10 to finish before starting checkpointing.
I0421 14:57:26.358482 133091194312512 checkpoint_manager.py:2020] [process=7][thread=MainThread][step=0][wait_until_finished] Waiting for Save Finalize thread (save_finalize) to complete.
I0421 14:57:27.309042 132963037005568 async_checkpointer.py:160] [process=7][thread=async_save] Background save thread done. Time taken: 26.983296s.
I0421 14:57:27.309320 132963011827456 async_checkpointer.py:288] [process=7][thread=save_finalize] Done with waiting for background save thread=async_save.
I0421 14:57:27.309440 132963011827456 async_checkpointer.py:298] [process=7][thread=save_finalize] No errors found in background save thread=async_save.
I0421 14:57:27.309489 132963011827456 checkpoint_manager.py:2137] [process=7][thread=save_finalize][step=0] CheckpointManager Save Finalize is syncing with other hosts...
I0421 14:57:27.311645 132963011827456 checkpoint_manager.py:2146] [process=7][thread=save_finalize][step=0] CheckpointManager Save Finalize is done on all hosts.
I0421 14:57:27.311854 133091194312512 checkpoint_manager.py:2032] [process=7][thread=MainThread][step=0][wait_until_finished] Done waiting for Save Finalize thread (save_finalize) running at step=0.
I0421 14:57:27.313353 133091194312512 checkpoint_manager.py:1512] [process=7] Saving checkpoint at step 10
I0421 14:57:27.315201 133091194312512 event_tracking.py:70] [process=7] [async] Started save checkpoint @ gs://lance-maxtext/linen_ckpt_xpk_main_20260421_144122/linen_xpk_main_20260421_144122_05_fp8/checkpoints/10.
I0421 14:57:28.038402 133091194312512 jax_array_handlers.py:360] Scheduling D2H of 81 prioritized jax.Array.
I0421 14:57:28.038494 133091194312512 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
I0421 14:57:28.092719 133091194312512 base_pytree_checkpoint_handler.py:154] [process=7][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 0.055356s
I0421 14:57:28.092906 133091194312512 base_pytree_checkpoint_handler.py:130] [process=7] /jax/orbax/write/blocking_gbytes_per_sec: 25.080 GiB/s (total gbytes: 1.5 GiB) (time elapsed: 0.06151294708251953 s) (per-host)
I0421 14:57:28.092957 133091194312512 base_pytree_checkpoint_handler.py:768] [process=7][thread=MainThread] Initiated Pytree async_save. Time taken: 0.061574s (batch_requests_ready=0.003024s, total_serialization_initiated=0.058481s, others=0.000069s)
I0421 14:57:28.093063 133091194312512 composite_checkpoint_handler.py:715] [process=7][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 0.065802s (all_items=0.000017s, per_item={'items': '0.00001717'}, temp_paths=0.065785)
I0421 14:57:28.093790 133091194312512 event_tracking.py:125] [process=7] [async] Finished blocking save in 0.78 seconds. Continuing save @ gs://lance-maxtext/linen_ckpt_xpk_main_20260421_144122/linen_xpk_main_20260421_144122_05_fp8/checkpoints/10.
I0421 14:57:28.094174 132961433278208 async_checkpointer.py:76] [process=7][thread=async_save] Background save thread started. Deadline for this save operation is 2026-04-21 15:17:28.094136
I0421 14:57:28.096343 133091194312512 checkpoint_manager.py:1560] [process=7][thread=MainThread][step=10] Starting CheckpointManager Save Finalize thread=save_finalize
I0421 14:57:28.096614 132963011827456 async_checkpointer.py:280] [process=7][thread=save_finalize] Waiting for background save thread=async_save.
I0421 14:57:28.096793 133091194312512 standard_logger.py:34] {'step': 10, 'event_type': 'save', 'directory': 'gs://lance-maxtext/linen_ckpt_xpk_main_20260421_144122/linen_xpk_main_20260421_144122_05_fp8/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776783446.358453, 'wait_for_prev_duration_secs': 0.9535021781921387, 'time_between_consecutive_saves_sec': None, 'checkpointer_blocking_start_time': 1776783447.3133917, 'checkpointer_blocking_duration_secs': 0.7809386253356934, 'get_old_steps_start_time': 1776783448.0943525, 'get_old_steps_duration_secs': 2.9325485229492188e-05, 'checkpoint_manager_blocking_start_time': 1776783446.3566496, 'checkpoint_manager_blocking_duration_secs': 1.7401056289672852}
I0421 14:57:28.096925 133091194312512 checkpointing.py:408] Started an asynchronous checkpoint save for step 10
I0421 14:57:28.097595 133091194312512 metric_logger.py:196] completed step: 9, seconds: 0.158, TFLOP/s/device: 85.864, Tokens/s/device: 12942.284, total_weights: 65536, loss: 7.908, lm_loss: 7.908, perplexity: 2720.116
I0421 14:57:28.105202 133091194312512 metric_logger.py:196] completed step: 10, seconds: 0.159, TFLOP/s/device: 85.631, Tokens/s/device: 12907.292, total_weights: 65536, loss: 7.568, lm_loss: 7.568, perplexity: 1936.198
I0421 14:57:28.262787 133091194312512 metric_logger.py:196] completed step: 11, seconds: 2.059, TFLOP/s/device: 6.600, Tokens/s/device: 994.837, total_weights: 65536, loss: 7.288, lm_loss: 7.288, perplexity: 1463.272
I0421 14:57:28.828835 133091194312512 metric_logger.py:196] completed step: 12, seconds: 0.007, TFLOP/s/device: 2037.356, Tokens/s/device: 307092.518, total_weights: 65536, loss: 7.032, lm_loss: 7.032, perplexity: 1132.061
I0421 14:57:28.987318 133091194312512 metric_logger.py:196] completed step: 13, seconds: 0.561, TFLOP/s/device: 24.217, Tokens/s/device: 3650.318, total_weights: 65536, loss: 6.815, lm_loss: 6.815, perplexity: 911.031
I0421 14:57:29.145483 133091194312512 metric_logger.py:196] completed step: 14, seconds: 0.164, TFLOP/s/device: 83.040, Tokens/s/device: 12516.654, total_weights: 65536, loss: 6.635, lm_loss: 6.635, perplexity: 761.080
I0421 14:57:29.303700 133091194312512 metric_logger.py:196] completed step: 15, seconds: 0.158, TFLOP/s/device: 85.888, Tokens/s/device: 12946.048, total_weights: 65536, loss: 6.492, lm_loss: 6.492, perplexity: 659.915
I0421 14:57:29.462432 133091194312512 metric_logger.py:196] completed step: 16, seconds: 0.158, TFLOP/s/device: 85.913, Tokens/s/device: 12949.731, total_weights: 65536, loss: 6.380, lm_loss: 6.380, perplexity: 589.948
I0421 14:57:29.620780 133091194312512 metric_logger.py:196] completed step: 17, seconds: 0.158, TFLOP/s/device: 85.849, Tokens/s/device: 12940.158, total_weights: 65536, loss: 6.293, lm_loss: 6.293, perplexity: 540.941
I0421 14:57:29.778961 133091194312512 metric_logger.py:196] completed step: 18, seconds: 0.159, TFLOP/s/device: 85.493, Tokens/s/device: 12886.420, total_weights: 65536, loss: 6.223, lm_loss: 6.223, perplexity: 504.108
I0421 14:57:29.936459 133091194312512 checkpointing.py:772] Waiting for step 19 to finish before checkpoint...
I0421 14:57:29.937474 133091194312512 checkpointing.py:776] Waited 0.0010347366333007812 seconds for step 19 to finish before starting checkpointing.
I0421 14:57:29.940545 133091194312512 checkpoint_manager.py:2020] [process=7][thread=MainThread][step=10][wait_until_finished] Waiting for Save Finalize thread (save_finalize) to complete.
I0421 14:57:33.794583 132960334358272 array_metadata_store.py:203] [process=7][thread=array_type_handler] Wrote 81 array_metadata.ArrayMetadata to gs://lance-maxtext/linen_ckpt_xpk_main_20260421_144122/linen_xpk_main_20260421_144122_05_fp8/checkpoints/10/items/array_metadatas/process_7
I0421 14:58:10.507524 132961433278208 base_pytree_checkpoint_handler.py:130] [process=7] /jax/orbax/write/gbytes_per_sec: 37.192 MiB/s (total gbytes: 1.5 GiB) (time elapsed: 42.476070404052734 s) (per-host)
I0421 14:58:10.507649 132961433278208 async_checkpointer.py:90] [process=7][thread=async_save] 3 Handler Commit operations completed. Time taken: 42.413356s.
I0421 14:58:17.126621 132961433278208 async_checkpointer.py:160] [process=7][thread=async_save] Background save thread done. Time taken: 49.032309s.
I0421 14:58:17.126961 132963011827456 async_checkpointer.py:288] [process=7][thread=save_finalize] Done with waiting for background save thread=async_save.
I0421 14:58:17.127083 132963011827456 async_checkpointer.py:298] [process=7][thread=save_finalize] No errors found in background save thread=async_save.
I0421 14:58:17.127130 132963011827456 checkpoint_manager.py:2137] [process=7][thread=save_finalize][step=10] CheckpointManager Save Finalize is syncing with other hosts...
I0421 14:58:17.129523 132963011827456 checkpoint_manager.py:2146] [process=7][thread=save_finalize][step=10] CheckpointManager Save Finalize is done on all hosts.
I0421 14:58:17.129688 133091194312512 checkpoint_manager.py:2032] [process=7][thread=MainThread][step=10][wait_until_finished] Done waiting for Save Finalize thread (save_finalize) running at step=10.
W0421 14:58:17.129854 133091194312512 checkpoint_manager.py:1452] Waiting for previous save to complete took 47.189306 seconds. If this number is high, consider checkpointing less frequently.
I0421 14:58:17.131581 133091194312512 checkpoint_manager.py:1512] [process=7] Saving checkpoint at step 19
I0421 14:58:17.133552 133091194312512 event_tracking.py:70] [process=7] [async] Started save checkpoint @ gs://lance-maxtext/linen_ckpt_xpk_main_20260421_144122/linen_xpk_main_20260421_144122_05_fp8/checkpoints/19.
I0421 14:58:17.855439 133091194312512 jax_array_handlers.py:360] Scheduling D2H of 81 prioritized jax.Array.
I0421 14:58:17.855535 133091194312512 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
I0421 14:58:17.909365 133091194312512 base_pytree_checkpoint_handler.py:154] [process=7][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 0.054965s
I0421 14:58:17.909517 133091194312512 base_pytree_checkpoint_handler.py:130] [process=7] /jax/orbax/write/blocking_gbytes_per_sec: 25.246 GiB/s (total gbytes: 1.5 GiB) (time elapsed: 0.06110692024230957 s) (per-host)
I0421 14:58:17.909565 133091194312512 base_pytree_checkpoint_handler.py:768] [process=7][thread=MainThread] Initiated Pytree async_save. Time taken: 0.061164s (batch_requests_ready=0.003038s, total_serialization_initiated=0.058064s, others=0.000062s)
I0421 14:58:17.909656 133091194312512 composite_checkpoint_handler.py:715] [process=7][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 0.065272s (all_items=0.000010s, per_item={'items': '0.00001049'}, temp_paths=0.065262)
I0421 14:58:17.910362 133091194312512 event_tracking.py:125] [process=7] [async] Finished blocking save in 0.78 seconds. Continuing save @ gs://lance-maxtext/linen_ckpt_xpk_main_20260421_144122/linen_xpk_main_20260421_144122_05_fp8/checkpoints/19.
I0421 14:58:17.910647 132963011827456 async_checkpointer.py:76] [process=7][thread=async_save] Background save thread started. Deadline for this save operation is 2026-04-21 15:18:17.910614
I0421 14:58:17.912683 133091194312512 checkpoint_manager.py:1560] [process=7][thread=MainThread][step=19] Starting CheckpointManager Save Finalize thread=save_finalize
I0421 14:58:17.913025 132960334358272 async_checkpointer.py:280] [process=7][thread=save_finalize] Waiting for background save thread=async_save.
I0421 14:58:17.913190 133091194312512 standard_logger.py:34] {'step': 19, 'event_type': 'save', 'directory': 'gs://lance-maxtext/linen_ckpt_xpk_main_20260421_144122/linen_xpk_main_20260421_144122_05_fp8/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776783449.9405153, 'wait_for_prev_duration_secs': 47.189306020736694, 'time_between_consecutive_saves_sec': 2.628831624984741, 'checkpointer_blocking_start_time': 1776783497.131618, 'checkpointer_blocking_duration_secs': 0.7791426181793213, 'get_old_steps_start_time': 1776783497.9107797, 'get_old_steps_duration_secs': 2.47955322265625e-05, 'checkpoint_manager_blocking_start_time': 1776783449.937762, 'checkpoint_manager_blocking_duration_secs': 47.97539281845093}
I0421 14:58:17.913311 133091194312512 checkpointing.py:408] Started an asynchronous checkpoint save for step 19
I0421 14:58:17.913357 133091194312512 checkpoint_manager.py:2020] [process=7][thread=MainThread][step=19][wait_until_finished] Waiting for Save Finalize thread (save_finalize) to complete.
I0421 14:58:22.968586 132960325965568 array_metadata_store.py:203] [process=7][thread=array_type_handler] Wrote 81 array_metadata.ArrayMetadata to gs://lance-maxtext/linen_ckpt_xpk_main_20260421_144122/linen_xpk_main_20260421_144122_05_fp8/checkpoints/19/items/array_metadatas/process_7
I0421 14:58:59.332427 132963011827456 base_pytree_checkpoint_handler.py:130] [process=7] /jax/orbax/write/gbytes_per_sec: 38.081 MiB/s (total gbytes: 1.5 GiB) (time elapsed: 41.483994245529175 s) (per-host)
I0421 14:58:59.332544 132963011827456 async_checkpointer.py:90] [process=7][thread=async_save] 3 Handler Commit operations completed. Time taken: 41.421219s.
I0421 14:59:07.841130 132963011827456 async_checkpointer.py:160] [process=7][thread=async_save] Background save thread done. Time taken: 49.929787s.
I0421 14:59:07.841357 132960334358272 async_checkpointer.py:288] [process=7][thread=save_finalize] Done with waiting for background save thread=async_save.
I0421 14:59:07.841415 132960334358272 async_checkpointer.py:298] [process=7][thread=save_finalize] No errors found in background save thread=async_save.
I0421 14:59:07.841456 132960334358272 checkpoint_manager.py:2137] [process=7][thread=save_finalize][step=19] CheckpointManager Save Finalize is syncing with other hosts...
I0421 14:59:07.843129 132960334358272 checkpoint_manager.py:2146] [process=7][thread=save_finalize][step=19] CheckpointManager Save Finalize is done on all hosts.
I0421 14:59:07.843353 133091194312512 checkpoint_manager.py:2032] [process=7][thread=MainThread][step=19][wait_until_finished] Done waiting for Save Finalize thread (save_finalize) running at step=19.
I0421 14:59:07.843512 133091194312512 checkpoint_manager.py:2009] [process=7][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning.
I0421 14:59:07.844414 133091194312512 metric_logger.py:196] completed step: 19, seconds: 0.158, TFLOP/s/device: 85.984, Tokens/s/device: 12960.467, total_weights: 65536, loss: 6.167, lm_loss: 6.167, perplexity: 476.947
Per train step:
 Total TFLOPs: 13.59 
 split as 93.93% learnable weight flops and 6.07% attention flops
XPK End: Tue Apr 21 14:59:18 UTC 2026
EXIT_CODE=0

feat/nnx-trainstate-and-training-loop · e27fc1e97 · feat_nnx_trainstate_and_training_loop_20260421_161900 · full log

XPK Start: Tue Apr 21 16:46:53 UTC 2026
PyTorch was not found. Models won't be available and only tokenizers, configuration and file/data utilities can be used.
2026-04-21 16:47:17.458838: E external/local_xla/xla/stream_executor/cuda/cuda_platform.cc:51] failed call to cuInit: INTERNAL: CUDA error: Failed call to cuInit: UNKNOWN ERROR (303)
I0421 16:47:17.638910 138979969611584 max_utils.py:273] Attempting to initialize the jax distributed system...
I0421 16:47:26.680066 138979969611584 distributed.py:149] Starting JAX distributed service on [::]:8482
I0421 16:47:26.682477 138979969611584 distributed.py:172] Connecting to JAX distributed service on mt-05-fp8-us6m5-slice-job-0-0.mt-05-fp8-us6m5:8482
I0421 16:47:28.286483 138979969611584 max_utils.py:284] Jax distributed system initialized!
I0421 16:47:34.380257 138979969611584 max_utils.py:800] System Information: Jax Version: 0.9.2
I0421 16:47:34.380360 138979969611584 max_utils.py:801] System Information: Jaxlib Version: 0.9.2
I0421 16:47:34.380401 138979969611584 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
I0421 16:47:34.380436 138979969611584 train_utils.py:378] WARNING: Sequence packing is essentially ignored for synthetic data. Please use a real dataset to use sequence packing.
I0421 16:47:35.113396 138979969611584 maxtext_utils.py:1718] Num_devices: 32, shape (1, 1, 1, 32, 1, 1, 1, 1, 1, 1, 1, 1, 1)
I0421 16:47:35.113996 138979969611584 maxtext_utils.py:1718] Num_devices: 32, shape (1, 1, 1, 32, 1, 1, 1, 1, 1, 1, 1, 1, 1)
I0421 16:47:35.114180 138979969611584 checkpointing.py:688] Setting up checkpoint logger...
I0421 16:47:35.114231 138979969611584 checkpointing.py:234] Creating checkpoint manager with ocdbt=True and zarr3=True
I0421 16:47:35.114272 138979969611584 pytree_checkpoint_handler.py:592] save_device_host_concurrent_bytes=None
I0421 16:47:35.114597 138979969611584 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 0x7e663058e840>, enable_pinned_host_transfer=False, save_concurrent_bytes: 96000000000 (89.4 GiB), restore_concurrent_bytes: 96000000000 (89.4 GiB)
I0421 16:47:38.344735 138979969611584 checkpointing.py:266] Enabling policy for fixed interval checkpointing.
I0421 16:47:38.344985 138979969611584 checkpoint_manager.py:708] [process=3][thread=MainThread] CheckpointManager init: checkpointers=None, item_names=('items',), item_handlers={'items': <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7e51d04e3fb0>}, handler_registry=None
I0421 16:47:38.345230 138979969611584 composite_checkpoint_handler.py:237] Deferred registration for item: "items". Adding handler `<orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7e51d04e3fb0>` 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`.
I0421 16:47:38.345279 138979969611584 composite_checkpoint_handler.py:237] Deferred registration for item: "metrics". Adding handler `<orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7e51d04e53d0>` for item "metrics" and save args `<class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonSaveArgs'>` and restore args `<class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonRestoreArgs'>` to `_handler_registry`.
I0421 16:47:38.345314 138979969611584 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 0x7e51d04e3fb0>, ('items', <class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeRestoreArgs'>): <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7e51d04e3fb0>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonSaveArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7e51d04e53d0>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonRestoreArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7e51d04e53d0>}).
I0421 16:47:38.345628 138979969611584 abstract_checkpointer.py:35] orbax-checkpoint version: 0.11.34
I0421 16:47:38.345701 138979969611584 async_checkpointer.py:192] [process=3][thread=MainThread] Using barrier_sync_fn: <function get_barrier_sync_fn.<locals>._fn at 0x7e51d03082c0> timeout: 1200 secs and primary_host=0 for async checkpoint writes
I0421 16:47:39.416587 138979969611584 checkpoint_manager.py:1812] Found 0 checkpoint steps in gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900/linen_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900_05_fp8/checkpoints
I0421 16:47:39.831115 138979969611584 checkpoint_manager.py:929] [process=3][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_20260421_161900/linen_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900_05_fp8/checkpoints: <orbax.checkpoint.checkpoint_manager.CheckpointManager object at 0x7e65081758b0>
I0421 16:47:39.831285 138979969611584 checkpointing.py:302] Checkpoint manager created!
I0421 16:47:40.946399 138979969611584 nnx_wrappers.py:437] Unknown Logical: bfloat16[32,2048,2048]...................................... ('activation_batch', 'activation_norm_length', 'activation_embed').
I0421 16:47:40.946505 138979969611584 nnx_wrappers.py:437] Unknown Physical: bfloat16[32,2048,2048]...................................... ('fsdp', None, None).
I0421 16:47:41.328308 138979969611584 attentions.py:1088] attentions/inputs_q Logical: bfloat16[32,2048,2048]...................................... ('activation_batch', 'activation_attn_length', 'activation_attn_embed').
I0421 16:47:41.328399 138979969611584 attentions.py:1088] attentions/inputs_q Physical: bfloat16[32,2048,2048]...................................... ('fsdp', None, None).
I0421 16:47:41.344769 138979969611584 attentions.py:1089] attentions/inputs_kv Logical: bfloat16[32,2048,2048]...................................... ('activation_batch', 'activation_attn_length', 'activation_attn_embed').
I0421 16:47:41.344829 138979969611584 attentions.py:1089] attentions/inputs_kv Physical: bfloat16[32,2048,2048]...................................... ('fsdp', None, None).
I0421 16:47:41.397119 138979969611584 attentions.py:1154] attentions/query Logical: bfloat16[32,2048,16,128].................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim').
I0421 16:47:41.397200 138979969611584 attentions.py:1154] attentions/query Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None).
I0421 16:47:41.413677 138979969611584 attentions.py:1155] attentions/key Logical: bfloat16[32,2048,16,128].................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim').
I0421 16:47:41.413738 138979969611584 attentions.py:1155] attentions/key Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None).
I0421 16:47:41.430315 138979969611584 attentions.py:1156] attentions/value Logical: bfloat16[32,2048,16,128].................................... ('activation_kv_batch', 'activation_attn_length', 'activation_kv_heads', 'activation_kv_head_dim').
I0421 16:47:41.430378 138979969611584 attentions.py:1156] attentions/value Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None).
I0421 16:47:41.455783 138979969611584 attentions.py:1197] attentions/out Logical: bfloat16[32,2048,16,128].................................... ('activation_batch', 'activation_attn_length', 'activation_heads', 'activation_kv').
I0421 16:47:41.455851 138979969611584 attentions.py:1197] attentions/out Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None).
I0421 16:47:41.507293 138979969611584 linears.py:525] linears/x Logical: bfloat16[32,2048,7168]...................................... ('activation_batch', 'activation_length', 'activation_mlp').
I0421 16:47:41.507371 138979969611584 linears.py:525] linears/x Physical: bfloat16[32,2048,7168]...................................... ('fsdp', None, None).
I0421 16:47:41.953334 138979969611584 checkpointing.py:578] checkpoint manager exists so trying to load this run's existing checkpoint
I0421 16:47:41.953449 138979969611584 checkpointing.py:676] No existing checkpoints found, not restoring checkpoint.
fsdp: 32
I0421 16:47:43.878537 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_0/Fp8DirectDotGeneralOp_0/input_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.878659 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_0/Fp8DirectDotGeneralOp_0/input_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.878708 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_0/Fp8DirectDotGeneralOp_0/kernel_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.878747 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_0/Fp8DirectDotGeneralOp_0/kernel_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.878783 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_0/Fp8DirectDotGeneralOp_0/output_grad_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.878816 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_0/Fp8DirectDotGeneralOp_0/output_grad_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.878848 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_1/Fp8DirectDotGeneralOp_0/input_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.878879 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_1/Fp8DirectDotGeneralOp_0/input_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.878910 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_1/Fp8DirectDotGeneralOp_0/kernel_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.878956 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_1/Fp8DirectDotGeneralOp_0/kernel_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.878989 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_1/Fp8DirectDotGeneralOp_0/output_grad_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879020 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_1/Fp8DirectDotGeneralOp_0/output_grad_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879049 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/mlp/wo/Fp8DirectDotGeneralOp_0/input_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879077 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/mlp/wo/Fp8DirectDotGeneralOp_0/input_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879106 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/mlp/wo/Fp8DirectDotGeneralOp_0/kernel_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879135 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/mlp/wo/Fp8DirectDotGeneralOp_0/kernel_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879167 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/mlp/wo/Fp8DirectDotGeneralOp_0/output_grad_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879195 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/mlp/wo/Fp8DirectDotGeneralOp_0/output_grad_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879223 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/self_attention/key/Fp8DirectDotGeneralOp_0/input_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879250 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/self_attention/key/Fp8DirectDotGeneralOp_0/input_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879276 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/self_attention/key/Fp8DirectDotGeneralOp_0/kernel_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879303 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/self_attention/key/Fp8DirectDotGeneralOp_0/kernel_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879328 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/self_attention/key/Fp8DirectDotGeneralOp_0/output_grad_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879355 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/self_attention/key/Fp8DirectDotGeneralOp_0/output_grad_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879381 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/self_attention/out/Fp8DirectDotGeneralOp_0/input_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879407 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/self_attention/out/Fp8DirectDotGeneralOp_0/input_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879433 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/self_attention/out/Fp8DirectDotGeneralOp_0/kernel_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879460 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/self_attention/out/Fp8DirectDotGeneralOp_0/kernel_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879486 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/self_attention/out/Fp8DirectDotGeneralOp_0/output_grad_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879517 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/self_attention/out/Fp8DirectDotGeneralOp_0/output_grad_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879545 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/self_attention/query/Fp8DirectDotGeneralOp_0/input_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879572 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/self_attention/query/Fp8DirectDotGeneralOp_0/input_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879598 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/self_attention/query/Fp8DirectDotGeneralOp_0/kernel_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879642 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/self_attention/query/Fp8DirectDotGeneralOp_0/kernel_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879672 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/self_attention/query/Fp8DirectDotGeneralOp_0/output_grad_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879699 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/self_attention/query/Fp8DirectDotGeneralOp_0/output_grad_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879725 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/self_attention/value/Fp8DirectDotGeneralOp_0/input_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879751 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/self_attention/value/Fp8DirectDotGeneralOp_0/input_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879777 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/self_attention/value/Fp8DirectDotGeneralOp_0/kernel_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879804 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/self_attention/value/Fp8DirectDotGeneralOp_0/kernel_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879830 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/self_attention/value/Fp8DirectDotGeneralOp_0/output_grad_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879856 138979969611584 maxtext_utils.py:1821]  params/_overwrite_with_gradient/decoder/layers/self_attention/value/Fp8DirectDotGeneralOp_0/output_grad_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 16:47:43.879907 138979969611584 maxtext_utils.py:1821]  params/params/decoder/decoder_norm/scale
    Shape:     float32[2048]
    Logical:   P('norm',)
    Physical:  (None,)
I0421 16:47:43.879973 138979969611584 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)
I0421 16:47:43.880014 138979969611584 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)
I0421 16:47:43.880061 138979969611584 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')
I0421 16:47:43.880109 138979969611584 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)
I0421 16:47:43.880142 138979969611584 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)
I0421 16:47:43.880188 138979969611584 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)
I0421 16:47:43.880234 138979969611584 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')
I0421 16:47:43.880271 138979969611584 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)
I0421 16:47:43.880305 138979969611584 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)
I0421 16:47:43.880349 138979969611584 maxtext_utils.py:1821]  params/params/decoder/logits_dense/kernel
    Shape:     float32[2048,32000]
    Logical:   P('embed_vocab', 'vocab')
    Physical:  ('fsdp', None)

I0421 16:47:43.880393 138979969611584 maxtext_utils.py:1821]  params/params/token_embedder/embedding
    Shape:     float32[32000,2048]
    Logical:   P('vocab', 'embed_vocab')
    Physical:  (None, 'fsdp')
I0421 16:47:45.509187 138979969611584 train.py:157] train/xent Logical: float32[32,2048]............................................ ('activation_embed_and_logits_batch', 'activation_length').
I0421 16:47:45.509285 138979969611584 train.py:157] train/xent Physical: float32[32,2048]............................................ ('fsdp', None).
I0421 16:47:45.524738 138979969611584 train.py:164] train/z_loss Logical: float32[32,2048]............................................ ('activation_embed_and_logits_batch', 'activation_length').
I0421 16:47:45.524797 138979969611584 train.py:164] train/z_loss Physical: float32[32,2048]............................................ ('fsdp', None).
I0421 16:47:59.199489 138979969611584 max_utils.py:791] Total memory size: 1.5 GB, Output size: 0.4 GB, Temp size: 1.1 GB, Argument size: 0.4 GB, Host temp size: 0.0 GB.
I0421 16:47:59.200459 138979969611584 metric_logger.py:301] number parameters: 1.105 billion
I0421 16:48:14.452633 138979969611584 checkpointing.py:794] Waiting for step 0 to finish before checkpoint...
I0421 16:48:14.622602 138979969611584 checkpointing.py:798] Waited 0.16994738578796387 seconds for step 0 to finish before starting checkpointing.
I0421 16:48:14.625045 138979969611584 checkpoint_manager.py:2009] [process=3][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning.
I0421 16:48:14.626944 138979969611584 checkpoint_manager.py:1512] [process=3] Saving checkpoint at step 0
I0421 16:48:14.628355 138979969611584 event_tracking.py:70] [process=3] [async] Started save checkpoint @ gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900/linen_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900_05_fp8/checkpoints/0.
I0421 16:48:14.973866 138979969611584 signaling_client.py:364] Using JaxDistributedSignalingClient
I0421 16:48:14.974840 138979969611584 jax_array_handlers.py:360] Scheduling D2H of 81 prioritized jax.Array.
I0421 16:48:14.974966 138979969611584 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
I0421 16:48:15.400432 138979969611584 base_pytree_checkpoint_handler.py:154] [process=3][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 0.426684s
I0421 16:48:15.400602 138979969611584 base_pytree_checkpoint_handler.py:130] [process=3] /jax/orbax/write/blocking_gbytes_per_sec: 3.548 GiB/s (total gbytes: 1.5 GiB) (time elapsed: 0.43477773666381836 s) (per-host)
I0421 16:48:15.400651 138979969611584 base_pytree_checkpoint_handler.py:768] [process=3][thread=MainThread] Initiated Pytree async_save. Time taken: 0.434838s (batch_requests_ready=0.003352s, total_serialization_initiated=0.431418s, others=0.000068s)
I0421 16:48:15.400759 138979969611584 composite_checkpoint_handler.py:715] [process=3][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 0.439428s (all_items=0.000018s, per_item={'items': '0.00001788'}, temp_paths=0.439410)
I0421 16:48:15.401628 138979969611584 event_tracking.py:125] [process=3] [async] Finished blocking save in 0.77 seconds. Continuing save @ gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900/linen_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900_05_fp8/checkpoints/0.
I0421 16:48:15.401986 138850051553024 async_checkpointer.py:76] [process=3][thread=async_save] Background save thread started. Deadline for this save operation is 2026-04-21 17:08:15.401947
I0421 16:48:15.404064 138979969611584 checkpoint_manager.py:1560] [process=3][thread=MainThread][step=0] Starting CheckpointManager Save Finalize thread=save_finalize
I0421 16:48:15.404313 138850026374912 async_checkpointer.py:280] [process=3][thread=save_finalize] Waiting for background save thread=async_save.
I0421 16:48:15.404434 138979969611584 standard_logger.py:34] {'step': 0, 'event_type': 'save', 'directory': 'gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900/linen_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900_05_fp8/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776790094.6250257, 'wait_for_prev_duration_secs': 6.079673767089844e-05, 'time_between_consecutive_saves_sec': None, 'checkpointer_blocking_start_time': 1776790094.6269834, 'checkpointer_blocking_duration_secs': 0.7751724720001221, 'get_old_steps_start_time': 1776790095.40218, 'get_old_steps_duration_secs': 3.147125244140625e-05, 'checkpoint_manager_blocking_start_time': 1776790094.6232343, 'checkpoint_manager_blocking_duration_secs': 0.7811565399169922}
I0421 16:48:15.404545 138979969611584 checkpointing.py:409] Started an asynchronous checkpoint save for step 0
I0421 16:48:15.404596 138979969611584 max_utils.py:750] 
Memstats: After params initialized:
I0421 16:48:15.404646 138979969611584 max_utils.py:756] 	Using (GB) 0.45 / 31.25 (1.440000%) on TPU_10(process=3,(2,2,0,0))
I0421 16:48:15.404679 138979969611584 max_utils.py:756] 	Using (GB) 0.45 / 31.25 (1.440000%) on TPU_11(process=3,(3,2,0,0))
I0421 16:48:15.404708 138979969611584 max_utils.py:756] 	Using (GB) 0.45 / 31.25 (1.440000%) on TPU_14(process=3,(2,3,0,0))
I0421 16:48:15.404734 138979969611584 max_utils.py:756] 	Using (GB) 0.45 / 31.25 (1.440000%) on TPU_15(process=3,(3,3,0,0))
I0421 16:48:15.719131 138979969611584 metric_logger.py:196] completed step: 0, seconds: 15.252, TFLOP/s/device: 0.891, Tokens/s/device: 134.277, total_weights: 65536, loss: 10.874, lm_loss: 10.874, perplexity: 52776.805
I0421 16:48:15.907658 138979969611584 metric_logger.py:196] completed step: 1, seconds: 1.265, TFLOP/s/device: 10.745, Tokens/s/device: 1619.583, total_weights: 65536, loss: 10.874, lm_loss: 10.874, perplexity: 52761.840
I0421 16:48:16.332610 138979969611584 metric_logger.py:196] completed step: 2, seconds: 0.031, TFLOP/s/device: 438.181, Tokens/s/device: 66047.472, total_weights: 65536, loss: 10.816, lm_loss: 10.816, perplexity: 49832.398
I0421 16:48:16.491272 138979969611584 metric_logger.py:196] completed step: 3, seconds: 0.425, TFLOP/s/device: 31.965, Tokens/s/device: 4818.098, total_weights: 65536, loss: 10.431, lm_loss: 10.431, perplexity: 33901.820
I0421 16:48:16.808818 138979969611584 metric_logger.py:196] completed step: 4, seconds: 0.164, TFLOP/s/device: 82.667, Tokens/s/device: 12460.453, total_weights: 65536, loss: 9.992, lm_loss: 9.992, perplexity: 21847.639
I0421 16:48:16.815352 138979969611584 metric_logger.py:196] completed step: 5, seconds: 0.158, TFLOP/s/device: 85.832, Tokens/s/device: 12937.461, total_weights: 65536, loss: 9.549, lm_loss: 9.549, perplexity: 14035.006
I0421 16:48:23.369302    2818 google_auth_provider.cc:181] Running on GCE, using service account 562977990677-compute@developer.gserviceaccount.com
I0421 16:48:25.560624 138850034767616 array_metadata_store.py:203] [process=3][thread=array_type_handler] Wrote 81 array_metadata.ArrayMetadata to gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900/linen_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900_05_fp8/checkpoints/0/items/array_metadatas/process_3
I0421 16:48:40.112979 138850051553024 base_pytree_checkpoint_handler.py:130] [process=3] /jax/orbax/write/gbytes_per_sec: 62.821 MiB/s (total gbytes: 1.5 GiB) (time elapsed: 25.147104740142822 s) (per-host)
I0421 16:48:40.113106 138850051553024 async_checkpointer.py:90] [process=3][thread=async_save] 3 Handler Commit operations completed. Time taken: 24.711004s.
I0421 16:48:41.235469 138979969611584 metric_logger.py:196] completed step: 6, seconds: 0.318, TFLOP/s/device: 42.693, Tokens/s/device: 6435.091, total_weights: 65536, loss: 9.111, lm_loss: 9.111, perplexity: 9058.688
I0421 16:48:41.393693 138979969611584 metric_logger.py:196] completed step: 7, seconds: 24.263, TFLOP/s/device: 0.560, Tokens/s/device: 84.409, total_weights: 65536, loss: 8.685, lm_loss: 8.685, perplexity: 5914.720
I0421 16:48:41.552536 138979969611584 metric_logger.py:196] completed step: 8, seconds: 0.163, TFLOP/s/device: 83.193, Tokens/s/device: 12539.799, total_weights: 65536, loss: 8.281, lm_loss: 8.281, perplexity: 3950.010
I0421 16:48:41.557542 138979969611584 checkpointing.py:794] Waiting for step 10 to finish before checkpoint...
I0421 16:48:41.870246 138979969611584 checkpointing.py:798] Waited 0.3126704692840576 seconds for step 10 to finish before starting checkpointing.
I0421 16:48:41.873073 138979969611584 checkpoint_manager.py:2020] [process=3][thread=MainThread][step=0][wait_until_finished] Waiting for Save Finalize thread (save_finalize) to complete.
I0421 16:48:48.988143 138850051553024 async_checkpointer.py:160] [process=3][thread=async_save] Background save thread done. Time taken: 33.586024s.
I0421 16:48:48.988367 138850026374912 async_checkpointer.py:288] [process=3][thread=save_finalize] Done with waiting for background save thread=async_save.
I0421 16:48:48.988433 138850026374912 async_checkpointer.py:298] [process=3][thread=save_finalize] No errors found in background save thread=async_save.
I0421 16:48:48.988484 138850026374912 checkpoint_manager.py:2137] [process=3][thread=save_finalize][step=0] CheckpointManager Save Finalize is syncing with other hosts...
I0421 16:48:48.990189 138850026374912 checkpoint_manager.py:2146] [process=3][thread=save_finalize][step=0] CheckpointManager Save Finalize is done on all hosts.
I0421 16:48:48.990361 138979969611584 checkpoint_manager.py:2032] [process=3][thread=MainThread][step=0][wait_until_finished] Done waiting for Save Finalize thread (save_finalize) running at step=0.
W0421 16:48:48.990502 138979969611584 checkpoint_manager.py:1452] Waiting for previous save to complete took 7.117429 seconds. If this number is high, consider checkpointing less frequently.
I0421 16:48:48.992175 138979969611584 checkpoint_manager.py:1512] [process=3] Saving checkpoint at step 10
I0421 16:48:48.994359 138979969611584 event_tracking.py:70] [process=3] [async] Started save checkpoint @ gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900/linen_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900_05_fp8/checkpoints/10.
I0421 16:48:49.696438 138979969611584 jax_array_handlers.py:360] Scheduling D2H of 81 prioritized jax.Array.
I0421 16:48:49.696536 138979969611584 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
I0421 16:48:49.755946 138979969611584 base_pytree_checkpoint_handler.py:154] [process=3][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 0.060439s
I0421 16:48:49.756117 138979969611584 base_pytree_checkpoint_handler.py:130] [process=3] /jax/orbax/write/blocking_gbytes_per_sec: 23.162 GiB/s (total gbytes: 1.5 GiB) (time elapsed: 0.06660747528076172 s) (per-host)
I0421 16:48:49.756168 138979969611584 base_pytree_checkpoint_handler.py:768] [process=3][thread=MainThread] Initiated Pytree async_save. Time taken: 0.066668s (batch_requests_ready=0.003055s, total_serialization_initiated=0.063545s, others=0.000068s)
I0421 16:48:49.756269 138979969611584 composite_checkpoint_handler.py:715] [process=3][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 0.070987s (all_items=0.000015s, per_item={'items': '0.00001502'}, temp_paths=0.070972)
I0421 16:48:49.756989 138979969611584 event_tracking.py:125] [process=3] [async] Finished blocking save in 0.76 seconds. Continuing save @ gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900/linen_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900_05_fp8/checkpoints/10.
I0421 16:48:49.757325 138847358805760 async_checkpointer.py:76] [process=3][thread=async_save] Background save thread started. Deadline for this save operation is 2026-04-21 17:08:49.757287
I0421 16:48:49.759402 138979969611584 checkpoint_manager.py:1560] [process=3][thread=MainThread][step=10] Starting CheckpointManager Save Finalize thread=save_finalize
I0421 16:48:49.759727 138850026374912 async_checkpointer.py:280] [process=3][thread=save_finalize] Waiting for background save thread=async_save.
I0421 16:48:49.759891 138979969611584 standard_logger.py:34] {'step': 10, 'event_type': 'save', 'directory': 'gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900/linen_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900_05_fp8/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776790121.8730435, 'wait_for_prev_duration_secs': 7.117428541183472, 'time_between_consecutive_saves_sec': None, 'checkpointer_blocking_start_time': 1776790128.9922142, 'checkpointer_blocking_duration_secs': 0.7652637958526611, 'get_old_steps_start_time': 1776790129.7575004, 'get_old_steps_duration_secs': 2.9802322387695312e-05, 'checkpoint_manager_blocking_start_time': 1776790121.8709881, 'checkpoint_manager_blocking_duration_secs': 7.888866662979126}
I0421 16:48:49.760015 138979969611584 checkpointing.py:409] Started an asynchronous checkpoint save for step 10
I0421 16:48:49.760705 138979969611584 metric_logger.py:196] completed step: 9, seconds: 0.158, TFLOP/s/device: 85.819, Tokens/s/device: 12935.581, total_weights: 65536, loss: 7.908, lm_loss: 7.908, perplexity: 2720.116
I0421 16:48:49.768042 138979969611584 metric_logger.py:196] completed step: 10, seconds: 0.159, TFLOP/s/device: 85.596, Tokens/s/device: 12902.006, total_weights: 65536, loss: 7.568, lm_loss: 7.568, perplexity: 1936.198
I0421 16:48:49.926009 138979969611584 metric_logger.py:196] completed step: 11, seconds: 8.208, TFLOP/s/device: 1.655, Tokens/s/device: 249.507, total_weights: 65536, loss: 7.288, lm_loss: 7.288, perplexity: 1463.272
I0421 16:48:50.084275 138979969611584 metric_logger.py:196] completed step: 12, seconds: 0.007, TFLOP/s/device: 2087.117, Tokens/s/device: 314592.934, total_weights: 65536, loss: 7.032, lm_loss: 7.032, perplexity: 1132.061
I0421 16:48:50.645906 138979969611584 metric_logger.py:196] completed step: 13, seconds: 0.159, TFLOP/s/device: 85.593, Tokens/s/device: 12901.438, total_weights: 65536, loss: 6.815, lm_loss: 6.815, perplexity: 911.031
I0421 16:48:50.804379 138979969611584 metric_logger.py:196] completed step: 14, seconds: 0.557, TFLOP/s/device: 24.413, Tokens/s/device: 3679.747, total_weights: 65536, loss: 6.635, lm_loss: 6.635, perplexity: 761.080
I0421 16:48:50.962593 138979969611584 metric_logger.py:196] completed step: 15, seconds: 0.163, TFLOP/s/device: 83.199, Tokens/s/device: 12540.644, total_weights: 65536, loss: 6.492, lm_loss: 6.492, perplexity: 659.915
I0421 16:48:51.120856 138979969611584 metric_logger.py:196] completed step: 16, seconds: 0.158, TFLOP/s/device: 85.757, Tokens/s/device: 12926.274, total_weights: 65536, loss: 6.380, lm_loss: 6.380, perplexity: 589.948
I0421 16:48:51.279290 138979969611584 metric_logger.py:196] completed step: 17, seconds: 0.158, TFLOP/s/device: 85.894, Tokens/s/device: 12946.866, total_weights: 65536, loss: 6.293, lm_loss: 6.293, perplexity: 540.941
I0421 16:48:51.437392 138979969611584 metric_logger.py:196] completed step: 18, seconds: 0.158, TFLOP/s/device: 85.788, Tokens/s/device: 12930.844, total_weights: 65536, loss: 6.223, lm_loss: 6.223, perplexity: 504.108
I0421 16:48:51.595086 138979969611584 checkpointing.py:794] Waiting for step 19 to finish before checkpoint...
I0421 16:48:51.596267 138979969611584 checkpointing.py:798] Waited 0.0011970996856689453 seconds for step 19 to finish before starting checkpointing.
I0421 16:48:51.598582 138979969611584 checkpoint_manager.py:2020] [process=3][thread=MainThread][step=10][wait_until_finished] Waiting for Save Finalize thread (save_finalize) to complete.
I0421 16:48:55.968545 138846285063936 array_metadata_store.py:203] [process=3][thread=array_type_handler] Wrote 81 array_metadata.ArrayMetadata to gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900/linen_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900_05_fp8/checkpoints/10/items/array_metadatas/process_3
I0421 16:49:32.779783 138847358805760 base_pytree_checkpoint_handler.py:130] [process=3] /jax/orbax/write/gbytes_per_sec: 36.662 MiB/s (total gbytes: 1.5 GiB) (time elapsed: 43.09022641181946 s) (per-host)
I0421 16:49:32.779913 138847358805760 async_checkpointer.py:90] [process=3][thread=async_save] 3 Handler Commit operations completed. Time taken: 43.022470s.
I0421 16:49:39.986553 138847358805760 async_checkpointer.py:160] [process=3][thread=async_save] Background save thread done. Time taken: 50.229094s.
I0421 16:49:39.986859 138850026374912 async_checkpointer.py:288] [process=3][thread=save_finalize] Done with waiting for background save thread=async_save.
I0421 16:49:39.986981 138850026374912 async_checkpointer.py:298] [process=3][thread=save_finalize] No errors found in background save thread=async_save.
I0421 16:49:39.987031 138850026374912 checkpoint_manager.py:2137] [process=3][thread=save_finalize][step=10] CheckpointManager Save Finalize is syncing with other hosts...
I0421 16:49:39.988660 138850026374912 checkpoint_manager.py:2146] [process=3][thread=save_finalize][step=10] CheckpointManager Save Finalize is done on all hosts.
I0421 16:49:39.988826 138979969611584 checkpoint_manager.py:2032] [process=3][thread=MainThread][step=10][wait_until_finished] Done waiting for Save Finalize thread (save_finalize) running at step=10.
W0421 16:49:39.988975 138979969611584 checkpoint_manager.py:1452] Waiting for previous save to complete took 48.390395 seconds. If this number is high, consider checkpointing less frequently.
I0421 16:49:39.991340 138979969611584 checkpoint_manager.py:1512] [process=3] Saving checkpoint at step 19
I0421 16:49:39.993491 138979969611584 event_tracking.py:70] [process=3] [async] Started save checkpoint @ gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900/linen_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900_05_fp8/checkpoints/19.
I0421 16:49:40.704732 138979969611584 jax_array_handlers.py:360] Scheduling D2H of 81 prioritized jax.Array.
I0421 16:49:40.704834 138979969611584 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
I0421 16:49:40.765172 138979969611584 base_pytree_checkpoint_handler.py:154] [process=3][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 0.061514s
I0421 16:49:40.765332 138979969611584 base_pytree_checkpoint_handler.py:130] [process=3] /jax/orbax/write/blocking_gbytes_per_sec: 22.794 GiB/s (total gbytes: 1.5 GiB) (time elapsed: 0.06768226623535156 s) (per-host)
I0421 16:49:40.765381 138979969611584 base_pytree_checkpoint_handler.py:768] [process=3][thread=MainThread] Initiated Pytree async_save. Time taken: 0.067740s (batch_requests_ready=0.002951s, total_serialization_initiated=0.064727s, others=0.000062s)
I0421 16:49:40.765480 138979969611584 composite_checkpoint_handler.py:715] [process=3][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 0.072911s (all_items=0.000011s, per_item={'items': '0.00001121'}, temp_paths=0.072900)
I0421 16:49:40.766275 138979969611584 event_tracking.py:125] [process=3] [async] Finished blocking save in 0.77 seconds. Continuing save @ gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900/linen_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900_05_fp8/checkpoints/19.
I0421 16:49:40.766537 138850026374912 async_checkpointer.py:76] [process=3][thread=async_save] Background save thread started. Deadline for this save operation is 2026-04-21 17:09:40.766517
I0421 16:49:40.768480 138979969611584 checkpoint_manager.py:1560] [process=3][thread=MainThread][step=19] Starting CheckpointManager Save Finalize thread=save_finalize
I0421 16:49:40.768769 138849471170304 async_checkpointer.py:280] [process=3][thread=save_finalize] Waiting for background save thread=async_save.
I0421 16:49:40.768879 138979969611584 standard_logger.py:34] {'step': 19, 'event_type': 'save', 'directory': 'gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900/linen_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900_05_fp8/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776790131.598556, 'wait_for_prev_duration_secs': 48.39039492607117, 'time_between_consecutive_saves_sec': 2.6083285808563232, 'checkpointer_blocking_start_time': 1776790179.9913805, 'checkpointer_blocking_duration_secs': 0.7752485275268555, 'get_old_steps_start_time': 1776790180.766646, 'get_old_steps_duration_secs': 2.47955322265625e-05, 'checkpoint_manager_blocking_start_time': 1776790131.5966516, 'checkpoint_manager_blocking_duration_secs': 49.17219638824463}
I0421 16:49:40.769001 138979969611584 checkpointing.py:409] Started an asynchronous checkpoint save for step 19
I0421 16:49:40.769045 138979969611584 checkpoint_manager.py:2020] [process=3][thread=MainThread][step=19][wait_until_finished] Waiting for Save Finalize thread (save_finalize) to complete.
I0421 16:49:46.649241 138846285063936 array_metadata_store.py:203] [process=3][thread=array_type_handler] Wrote 81 array_metadata.ArrayMetadata to gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900/linen_xpk_feat_nnx_trainstate_and_training_loop_20260421_161900_05_fp8/checkpoints/19/items/array_metadatas/process_3
I0421 16:50:23.339301 138850026374912 base_pytree_checkpoint_handler.py:130] [process=3] /jax/orbax/write/gbytes_per_sec: 37.047 MiB/s (total gbytes: 1.5 GiB) (time elapsed: 42.64160418510437 s) (per-host)
I0421 16:50:23.339428 138850026374912 async_checkpointer.py:90] [process=3][thread=async_save] 3 Handler Commit operations completed. Time taken: 42.572823s.
I0421 16:50:31.223580 138850026374912 async_checkpointer.py:160] [process=3][thread=async_save] Background save thread done. Time taken: 50.456957s.
I0421 16:50:31.223828 138849471170304 async_checkpointer.py:288] [process=3][thread=save_finalize] Done with waiting for background save thread=async_save.
I0421 16:50:31.223891 138849471170304 async_checkpointer.py:298] [process=3][thread=save_finalize] No errors found in background save thread=async_save.
I0421 16:50:31.223948 138849471170304 checkpoint_manager.py:2137] [process=3][thread=save_finalize][step=19] CheckpointManager Save Finalize is syncing with other hosts...
I0421 16:50:31.226629 138849471170304 checkpoint_manager.py:2146] [process=3][thread=save_finalize][step=19] CheckpointManager Save Finalize is done on all hosts.
I0421 16:50:31.226811 138979969611584 checkpoint_manager.py:2032] [process=3][thread=MainThread][step=19][wait_until_finished] Done waiting for Save Finalize thread (save_finalize) running at step=19.
I0421 16:50:31.226989 138979969611584 checkpoint_manager.py:2009] [process=3][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning.
I0421 16:50:31.228004 138979969611584 metric_logger.py:196] completed step: 19, seconds: 0.158, TFLOP/s/device: 85.808, Tokens/s/device: 12933.866, total_weights: 65536, loss: 6.167, lm_loss: 6.167, perplexity: 476.947
Per train step:
 Total TFLOPs: 13.59 
 split as 93.93% learnable weight flops and 6.07% attention flops
XPK End: Tue Apr 21 16:50:43 UTC 2026
EXIT_CODE=0

Metrics: main (e27fc1e97) vs feat/nnx-trainstate-and-training-loop (e27fc1e97)

Metrics: `main (e27fc1e97)` vs `feat/nnx-trainstate-and-training-loop (e27fc1e97)`