Compare: 05_fp8 — MaxView

‹ 04_int8Case: 05_fp806_grad_accum ›

Metrics: `main (e27fc1e97)` vs `feat/nnx-post-train-fixes (e27fc1e97)`

Metric	main `e27fc1e97`	feat/nnx-post-train-fixes `e27fc1e97`	Diff (feat/nnx-post-train-fixes − main)
Parameters	1.105 billion	1.105 billion	—
Final loss	6.1670	6.1670	0
TFLOP/s	85.902	85.682	-0.22
Tok/s	12948.0	12914.9	-33.069
Avg s/step	1.789	1.630	-0.159
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-post-train-fixes
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-post-train-fixes
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-post-train-fixes

Device	Used (GB)	Total (GB)	%
TPU_16(process=4,(0,4,0,0))	0.45	31.25	1.4%
TPU_17(process=4,(1,4,0,0))	0.45	31.25	1.4%
TPU_20(process=4,(0,5,0,0))	0.45	31.25	1.4%
TPU_21(process=4,(1,5,0,0))	0.45	31.25	1.4%

main TFLOP/s

feat/nnx-post-train-fixes TFLOP/s

Step	main Loss	feat/nnx-post-train-fixes Loss	main TFLOP/s	feat/nnx-post-train-fixes TFLOP/s	TFLOP/s Δ
0	10.874	10.874	0.895	0.896	+0.0010
1	10.874	10.874	8.101	10.850	+2.7490
2	10.816	10.816	319.885	441.169	+121.2840
3	10.431	10.431	15.745	32.122	+16.3770
4	9.992	9.992	82.512	82.971	+0.4590
5	9.549	9.549	85.975	85.994	+0.0190
6	9.111	9.111	42.760	42.778	+0.0180
7	8.685	8.685	0.559	0.571	+0.0120
8	8.281	8.281	83.129	83.272	+0.1430
9	7.908	7.908	84.286	85.763	+1.4770
10	7.568	7.568	87.359	85.751	-1.6080
11	7.288	7.288	3.047	4.805	+1.7580
12	7.032	7.032	2242.841	2241.731	-1.1100
13	6.815	6.815	24.435	85.239	+60.8040
14	6.635	6.635	83.664	24.838	-58.8260
15	6.492	6.492	85.948	83.433	-2.5150
16	6.380	6.380	85.898	85.739	-0.1590
17	6.293	6.293	85.904	85.707	-0.1970
18	6.223	6.223	85.821	85.906	+0.0850
19	6.167	6.167	85.902	85.682	-0.2200

main · e27fc1e97 · main_20260421_061359 · full log

XPK Start: Tue Apr 21 06:27:57 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 06:28:23.011727: 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 06:28:23.220046 132163276240704 max_utils.py:273] Attempting to initialize the jax distributed system...
I0421 06:28:32.261011 132163276240704 distributed.py:149] Starting JAX distributed service on [::]:8482
I0421 06:28:32.263340 132163276240704 distributed.py:172] Connecting to JAX distributed service on mt-05-fp8-1cd78-slice-job-0-0.mt-05-fp8-1cd78:8482
I0421 06:28:33.647696 132163276240704 max_utils.py:284] Jax distributed system initialized!
I0421 06:28:39.707912 132163276240704 max_utils.py:800] System Information: Jax Version: 0.9.2
I0421 06:28:39.708024 132163276240704 max_utils.py:801] System Information: Jaxlib Version: 0.9.2
I0421 06:28:39.708064 132163276240704 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 06:28:39.708099 132163276240704 train_utils.py:361] WARNING: Sequence packing is essentially ignored for synthetic data. Please use a real dataset to use sequence packing.
I0421 06:28:40.404944 132163276240704 maxtext_utils.py:1565] Num_devices: 32, shape (1, 1, 1, 32, 1, 1, 1, 1, 1, 1, 1, 1, 1)
I0421 06:28:40.405224 132163276240704 checkpointing.py:677] Setting up checkpoint logger...
I0421 06:28:40.405278 132163276240704 checkpointing.py:233] Creating checkpoint manager with ocdbt=True and zarr3=True
I0421 06:28:40.405322 132163276240704 pytree_checkpoint_handler.py:592] save_device_host_concurrent_bytes=None
I0421 06:28:40.405647 132163276240704 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 0x7832f1748740>, enable_pinned_host_transfer=False, save_concurrent_bytes: 96000000000 (89.4 GiB), restore_concurrent_bytes: 96000000000 (89.4 GiB)
I0421 06:28:44.084786 132163276240704 checkpointing.py:265] Enabling policy for fixed interval checkpointing.
I0421 06:28:44.085026 132163276240704 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 0x781ebc63dbb0>}, handler_registry=None
I0421 06:28:44.085264 132163276240704 composite_checkpoint_handler.py:237] Deferred registration for item: "items". Adding handler `<orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x781ebc63dbb0>` 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 06:28:44.085313 132163276240704 composite_checkpoint_handler.py:237] Deferred registration for item: "metrics". Adding handler `<orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x781ebc641790>` 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 06:28:44.085350 132163276240704 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 0x781ebc63dbb0>, ('items', <class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeRestoreArgs'>): <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x781ebc63dbb0>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonSaveArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x781ebc641790>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonRestoreArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x781ebc641790>}).
I0421 06:28:44.085663 132163276240704 abstract_checkpointer.py:35] orbax-checkpoint version: 0.11.34
I0421 06:28:44.085744 132163276240704 async_checkpointer.py:192] [process=7][thread=MainThread] Using barrier_sync_fn: <function get_barrier_sync_fn.<locals>._fn at 0x781ebc3d9440> timeout: 1200 secs and primary_host=0 for async checkpoint writes
I0421 06:28:45.347954 132163276240704 checkpoint_manager.py:1812] Found 0 checkpoint steps in gs://lance-maxtext/linen_ckpt_xpk_main_20260421_061359/linen_xpk_main_20260421_061359_05_fp8/checkpoints
I0421 06:28:45.768409 132163276240704 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_061359/linen_xpk_main_20260421_061359_05_fp8/checkpoints: <orbax.checkpoint.checkpoint_manager.CheckpointManager object at 0x781ebc63d130>
I0421 06:28:45.768589 132163276240704 checkpointing.py:301] Checkpoint manager created!
I0421 06:28:46.860480 132163276240704 nnx_wrappers.py:437] Unknown Logical: bfloat16[32,2048,2048]...................................... ('activation_batch', 'activation_norm_length', 'activation_embed').
I0421 06:28:46.860601 132163276240704 nnx_wrappers.py:437] Unknown Physical: bfloat16[32,2048,2048]...................................... ('fsdp', None, None).
I0421 06:28:47.245419 132163276240704 attentions.py:1088] attentions/inputs_q Logical: bfloat16[32,2048,2048]...................................... ('activation_batch', 'activation_attn_length', 'activation_attn_embed').
I0421 06:28:47.245517 132163276240704 attentions.py:1088] attentions/inputs_q Physical: bfloat16[32,2048,2048]...................................... ('fsdp', None, None).
I0421 06:28:47.261877 132163276240704 attentions.py:1089] attentions/inputs_kv Logical: bfloat16[32,2048,2048]...................................... ('activation_batch', 'activation_attn_length', 'activation_attn_embed').
I0421 06:28:47.261939 132163276240704 attentions.py:1089] attentions/inputs_kv Physical: bfloat16[32,2048,2048]...................................... ('fsdp', None, None).
I0421 06:28:47.314231 132163276240704 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 06:28:47.314315 132163276240704 attentions.py:1154] attentions/query Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None).
I0421 06:28:47.330748 132163276240704 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 06:28:47.330813 132163276240704 attentions.py:1155] attentions/key Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None).
I0421 06:28:47.347155 132163276240704 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 06:28:47.347217 132163276240704 attentions.py:1156] attentions/value Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None).
I0421 06:28:47.372632 132163276240704 attentions.py:1198] attentions/out Logical: bfloat16[32,2048,16,128].................................... ('activation_batch', 'activation_attn_length', 'activation_heads', 'activation_kv').
I0421 06:28:47.372709 132163276240704 attentions.py:1198] attentions/out Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None).
I0421 06:28:47.430418 132163276240704 linears.py:525] linears/x Logical: bfloat16[32,2048,7168]...................................... ('activation_batch', 'activation_length', 'activation_mlp').
I0421 06:28:47.430517 132163276240704 linears.py:525] linears/x Physical: bfloat16[32,2048,7168]...................................... ('fsdp', None, None).
I0421 06:28:47.864310 132163276240704 checkpointing.py:577] checkpoint manager exists so trying to load this run's existing checkpoint
I0421 06:28:47.864430 132163276240704 checkpointing.py:665] No existing checkpoints found, not restoring checkpoint.
fsdp: 32
I0421 06:28:49.805718 132163276240704 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 06:28:49.805852 132163276240704 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 06:28:49.805899 132163276240704 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 06:28:49.805943 132163276240704 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 06:28:49.805978 132163276240704 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 06:28:49.806015 132163276240704 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 06:28:49.806048 132163276240704 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 06:28:49.806077 132163276240704 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 06:28:49.806106 132163276240704 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 06:28:49.806134 132163276240704 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 06:28:49.806163 132163276240704 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 06:28:49.806191 132163276240704 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 06:28:49.806222 132163276240704 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 06:28:49.806251 132163276240704 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/mlp/wo/Fp8DirectDotGeneralOp_0/input_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 06:28:49.806279 132163276240704 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 06:28:49.806306 132163276240704 maxtext_utils.py:1668]  params/_overwrite_with_gradient/decoder/layers/mlp/wo/Fp8DirectDotGeneralOp_0/kernel_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 06:28:49.806333 132163276240704 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 06:28:49.806361 132163276240704 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 06:28:49.806389 132163276240704 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 06:28:49.806416 132163276240704 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 06:28:49.806442 132163276240704 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 06:28:49.806469 132163276240704 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 06:28:49.806495 132163276240704 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 06:28:49.806521 132163276240704 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 06:28:49.806547 132163276240704 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 06:28:49.806573 132163276240704 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 06:28:49.806599 132163276240704 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 06:28:49.806624 132163276240704 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 06:28:49.806650 132163276240704 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 06:28:49.806676 132163276240704 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 06:28:49.806703 132163276240704 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 06:28:49.806729 132163276240704 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 06:28:49.806777 132163276240704 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 06:28:49.806825 132163276240704 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 06:28:49.806856 132163276240704 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 06:28:49.806884 132163276240704 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 06:28:49.806910 132163276240704 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 06:28:49.806935 132163276240704 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 06:28:49.806963 132163276240704 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 06:28:49.807004 132163276240704 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 06:28:49.807059 132163276240704 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 06:28:49.807092 132163276240704 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 06:28:49.807144 132163276240704 maxtext_utils.py:1668]  params/params/decoder/decoder_norm/scale
    Shape:     float32[2048]
    Logical:   P('norm',)
    Physical:  (None,)
I0421 06:28:49.807196 132163276240704 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 06:28:49.807232 132163276240704 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 06:28:49.807276 132163276240704 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 06:28:49.807320 132163276240704 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 06:28:49.807352 132163276240704 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 06:28:49.807397 132163276240704 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 06:28:49.807442 132163276240704 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 06:28:49.807475 132163276240704 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 06:28:49.807506 132163276240704 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 06:28:49.807547 132163276240704 maxtext_utils.py:1668]  params/params/decoder/logits_dense/kernel
    Shape:     float32[2048,32000]
    Logical:   P('embed_vocab', 'vocab')
    Physical:  ('fsdp', None)
I0421 06:28:49.807587 132163276240704 maxtext_utils.py:1668]  params/params/token_embedder/embedding
    Shape:     float32[32000,2048]
    Logical:   P('vocab', 'embed_vocab')
    Physical:  (None, 'fsdp')

I0421 06:28:51.425977 132163276240704 train.py:155] train/xent Logical: float32[32,2048]............................................ ('activation_embed_and_logits_batch', 'activation_length').
I0421 06:28:51.426072 132163276240704 train.py:155] train/xent Physical: float32[32,2048]............................................ ('fsdp', None).
I0421 06:28:51.441353 132163276240704 train.py:162] train/z_loss Logical: float32[32,2048]............................................ ('activation_embed_and_logits_batch', 'activation_length').
I0421 06:28:51.441415 132163276240704 train.py:162] train/z_loss Physical: float32[32,2048]............................................ ('fsdp', None).
I0421 06:29:05.556449 132163276240704 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 06:29:05.557339 132163276240704 metric_logger.py:301] number parameters: 1.105 billion
I0421 06:29:20.740784 132163276240704 checkpointing.py:772] Waiting for step 0 to finish before checkpoint...
I0421 06:29:20.908869 132163276240704 checkpointing.py:776] Waited 0.1680917739868164 seconds for step 0 to finish before starting checkpointing.
I0421 06:29:20.911437 132163276240704 checkpoint_manager.py:2009] [process=7][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning.
I0421 06:29:20.913308 132163276240704 checkpoint_manager.py:1512] [process=7] Saving checkpoint at step 0
I0421 06:29:20.914724 132163276240704 event_tracking.py:70] [process=7] [async] Started save checkpoint @ gs://lance-maxtext/linen_ckpt_xpk_main_20260421_061359/linen_xpk_main_20260421_061359_05_fp8/checkpoints/0.
I0421 06:29:21.681885 132163276240704 signaling_client.py:364] Using JaxDistributedSignalingClient
I0421 06:29:21.682889 132163276240704 jax_array_handlers.py:360] Scheduling D2H of 81 prioritized jax.Array.
I0421 06:29:21.682985 132163276240704 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 06:29:22.090966 132163276240704 base_pytree_checkpoint_handler.py:154] [process=7][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 0.409202s
I0421 06:29:22.091136 132163276240704 base_pytree_checkpoint_handler.py:130] [process=7] /jax/orbax/write/blocking_gbytes_per_sec: 3.693 GiB/s (total gbytes: 1.5 GiB) (time elapsed: 0.41774606704711914 s) (per-host)
I0421 06:29:22.091188 132163276240704 base_pytree_checkpoint_handler.py:768] [process=7][thread=MainThread] Initiated Pytree async_save. Time taken: 0.417809s (batch_requests_ready=0.003831s, total_serialization_initiated=0.413906s, others=0.000071s)
I0421 06:29:22.091312 132163276240704 composite_checkpoint_handler.py:715] [process=7][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 0.422017s (all_items=0.000018s, per_item={'items': '0.00001812'}, temp_paths=0.421999)
I0421 06:29:22.092077 132163276240704 event_tracking.py:125] [process=7] [async] Finished blocking save in 1.18 seconds. Continuing save @ gs://lance-maxtext/linen_ckpt_xpk_main_20260421_061359/linen_xpk_main_20260421_061359_05_fp8/checkpoints/0.
I0421 06:29:22.092348 132032193623808 async_checkpointer.py:76] [process=7][thread=async_save] Background save thread started. Deadline for this save operation is 2026-04-21 06:49:22.092323
I0421 06:29:22.106103 132163276240704 checkpoint_manager.py:1560] [process=7][thread=MainThread][step=0] Starting CheckpointManager Save Finalize thread=save_finalize
I0421 06:29:22.106351 132032168445696 async_checkpointer.py:280] [process=7][thread=save_finalize] Waiting for background save thread=async_save.
I0421 06:29:22.106475 132163276240704 standard_logger.py:34] {'step': 0, 'event_type': 'save', 'directory': 'gs://lance-maxtext/linen_ckpt_xpk_main_20260421_061359/linen_xpk_main_20260421_061359_05_fp8/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776752960.9114192, 'wait_for_prev_duration_secs': 6.008148193359375e-05, 'time_between_consecutive_saves_sec': None, 'checkpointer_blocking_start_time': 1776752960.9133472, 'checkpointer_blocking_duration_secs': 1.1791143417358398, 'get_old_steps_start_time': 1776752962.0924857, 'get_old_steps_duration_secs': 3.0279159545898438e-05, 'checkpoint_manager_blocking_start_time': 1776752960.9095075, 'checkpoint_manager_blocking_duration_secs': 1.1969287395477295}
I0421 06:29:22.106578 132163276240704 checkpointing.py:408] Started an asynchronous checkpoint save for step 0
I0421 06:29:22.106629 132163276240704 max_utils.py:750] 
Memstats: After params initialized:
I0421 06:29:22.106678 132163276240704 max_utils.py:756] 	Using (GB) 0.45 / 31.25 (1.440000%) on TPU_26(process=7,(2,6,0,0))
I0421 06:29:22.106711 132163276240704 max_utils.py:756] 	Using (GB) 0.45 / 31.25 (1.440000%) on TPU_27(process=7,(3,6,0,0))
I0421 06:29:22.106749 132163276240704 max_utils.py:756] 	Using (GB) 0.45 / 31.25 (1.440000%) on TPU_30(process=7,(2,7,0,0))
I0421 06:29:22.106777 132163276240704 max_utils.py:756] 	Using (GB) 0.45 / 31.25 (1.440000%) on TPU_31(process=7,(3,7,0,0))
I0421 06:29:22.420022 132163276240704 metric_logger.py:196] completed step: 0, seconds: 15.183, TFLOP/s/device: 0.895, Tokens/s/device: 134.885, total_weights: 65536, loss: 10.874, lm_loss: 10.874, perplexity: 52776.805
I0421 06:29:22.620060 132163276240704 metric_logger.py:196] completed step: 1, seconds: 1.677, TFLOP/s/device: 8.101, Tokens/s/device: 1221.065, total_weights: 65536, loss: 10.874, lm_loss: 10.874, perplexity: 52761.840
I0421 06:29:23.482835 132163276240704 metric_logger.py:196] completed step: 2, seconds: 0.042, TFLOP/s/device: 319.885, Tokens/s/device: 48216.598, total_weights: 65536, loss: 10.816, lm_loss: 10.816, perplexity: 49832.398
I0421 06:29:23.641300 132163276240704 metric_logger.py:196] completed step: 3, seconds: 0.863, TFLOP/s/device: 15.745, Tokens/s/device: 2373.282, total_weights: 65536, loss: 10.431, lm_loss: 10.431, perplexity: 33901.820
I0421 06:29:23.958712 132163276240704 metric_logger.py:196] completed step: 4, seconds: 0.165, TFLOP/s/device: 82.512, Tokens/s/device: 12437.146, total_weights: 65536, loss: 9.992, lm_loss: 9.992, perplexity: 21847.639
I0421 06:29:23.965051 132163276240704 metric_logger.py:196] completed step: 5, seconds: 0.158, TFLOP/s/device: 85.975, Tokens/s/device: 12959.155, total_weights: 65536, loss: 9.549, lm_loss: 9.549, perplexity: 14035.006
I0421 06:29:26.137092    2840 google_auth_provider.cc:181] Running on GCE, using service account 562977990677-compute@developer.gserviceaccount.com
I0421 06:29:28.602801 132032176838400 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_061359/linen_xpk_main_20260421_061359_05_fp8/checkpoints/0/items/array_metadatas/process_7
I0421 06:29:43.029149 132032193623808 base_pytree_checkpoint_handler.py:130] [process=7] /jax/orbax/write/gbytes_per_sec: 73.974 MiB/s (total gbytes: 1.5 GiB) (time elapsed: 21.355700254440308 s) (per-host)
I0421 06:29:43.029272 132032193623808 async_checkpointer.py:90] [process=7][thread=async_save] 3 Handler Commit operations completed. Time taken: 20.936846s.
I0421 06:29:48.429259 132163276240704 metric_logger.py:196] completed step: 6, seconds: 0.318, TFLOP/s/device: 42.760, Tokens/s/device: 6445.298, total_weights: 65536, loss: 9.111, lm_loss: 9.111, perplexity: 9058.688
I0421 06:29:48.587963 132163276240704 metric_logger.py:196] completed step: 7, seconds: 24.307, TFLOP/s/device: 0.559, Tokens/s/device: 84.257, total_weights: 65536, loss: 8.685, lm_loss: 8.685, perplexity: 5914.720
I0421 06:29:48.746113 132163276240704 metric_logger.py:196] completed step: 8, seconds: 0.163, TFLOP/s/device: 83.129, Tokens/s/device: 12530.132, total_weights: 65536, loss: 8.281, lm_loss: 8.281, perplexity: 3950.010
I0421 06:29:48.750975 132163276240704 checkpointing.py:772] Waiting for step 10 to finish before checkpoint...
I0421 06:29:49.062427 132163276240704 checkpointing.py:776] Waited 0.3114185333251953 seconds for step 10 to finish before starting checkpointing.
I0421 06:29:49.065165 132163276240704 checkpoint_manager.py:2020] [process=7][thread=MainThread][step=0][wait_until_finished] Waiting for Save Finalize thread (save_finalize) to complete.
I0421 06:29:52.422053 132032193623808 async_checkpointer.py:160] [process=7][thread=async_save] Background save thread done. Time taken: 30.329609s.
I0421 06:29:52.422265 132032168445696 async_checkpointer.py:288] [process=7][thread=save_finalize] Done with waiting for background save thread=async_save.
I0421 06:29:52.422330 132032168445696 async_checkpointer.py:298] [process=7][thread=save_finalize] No errors found in background save thread=async_save.
I0421 06:29:52.422379 132032168445696 checkpoint_manager.py:2137] [process=7][thread=save_finalize][step=0] CheckpointManager Save Finalize is syncing with other hosts...
I0421 06:29:52.424604 132032168445696 checkpoint_manager.py:2146] [process=7][thread=save_finalize][step=0] CheckpointManager Save Finalize is done on all hosts.
I0421 06:29:52.424793 132163276240704 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.
W0421 06:29:52.424934 132163276240704 checkpoint_manager.py:1452] Waiting for previous save to complete took 3.359771 seconds. If this number is high, consider checkpointing less frequently.
I0421 06:29:52.426844 132163276240704 checkpoint_manager.py:1512] [process=7] Saving checkpoint at step 10
I0421 06:29:52.428817 132163276240704 event_tracking.py:70] [process=7] [async] Started save checkpoint @ gs://lance-maxtext/linen_ckpt_xpk_main_20260421_061359/linen_xpk_main_20260421_061359_05_fp8/checkpoints/10.
I0421 06:29:53.148622 132163276240704 jax_array_handlers.py:360] Scheduling D2H of 81 prioritized jax.Array.
I0421 06:29:53.148716 132163276240704 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 06:29:53.200686 132163276240704 base_pytree_checkpoint_handler.py:154] [process=7][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 0.052956s
I0421 06:29:53.200871 132163276240704 base_pytree_checkpoint_handler.py:130] [process=7] /jax/orbax/write/blocking_gbytes_per_sec: 26.013 GiB/s (total gbytes: 1.5 GiB) (time elapsed: 0.059307098388671875 s) (per-host)
I0421 06:29:53.200923 132163276240704 base_pytree_checkpoint_handler.py:768] [process=7][thread=MainThread] Initiated Pytree async_save. Time taken: 0.059368s (batch_requests_ready=0.003294s, total_serialization_initiated=0.056005s, others=0.000069s)
I0421 06:29:53.201033 132163276240704 composite_checkpoint_handler.py:715] [process=7][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 0.063605s (all_items=0.000016s, per_item={'items': '0.00001574'}, temp_paths=0.063589)
I0421 06:29:53.201761 132163276240704 event_tracking.py:125] [process=7] [async] Finished blocking save in 0.77 seconds. Continuing save @ gs://lance-maxtext/linen_ckpt_xpk_main_20260421_061359/linen_xpk_main_20260421_061359_05_fp8/checkpoints/10.
I0421 06:29:53.202055 132030037747456 async_checkpointer.py:76] [process=7][thread=async_save] Background save thread started. Deadline for this save operation is 2026-04-21 06:49:53.202019
I0421 06:29:53.204074 132163276240704 checkpoint_manager.py:1560] [process=7][thread=MainThread][step=10] Starting CheckpointManager Save Finalize thread=save_finalize
I0421 06:29:53.204394 132032168445696 async_checkpointer.py:280] [process=7][thread=save_finalize] Waiting for background save thread=async_save.
I0421 06:29:53.204500 132163276240704 standard_logger.py:34] {'step': 10, 'event_type': 'save', 'directory': 'gs://lance-maxtext/linen_ckpt_xpk_main_20260421_061359/linen_xpk_main_20260421_061359_05_fp8/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776752989.0651329, 'wait_for_prev_duration_secs': 3.3597710132598877, 'time_between_consecutive_saves_sec': None, 'checkpointer_blocking_start_time': 1776752992.4268832, 'checkpointer_blocking_duration_secs': 0.7753148078918457, 'get_old_steps_start_time': 1776752993.2022212, 'get_old_steps_duration_secs': 2.8848648071289062e-05, 'checkpoint_manager_blocking_start_time': 1776752989.063141, 'checkpoint_manager_blocking_duration_secs': 4.141326665878296}
I0421 06:29:53.204682 132163276240704 checkpointing.py:408] Started an asynchronous checkpoint save for step 10
I0421 06:29:53.205248 132163276240704 metric_logger.py:196] completed step: 9, seconds: 0.161, TFLOP/s/device: 84.286, Tokens/s/device: 12704.478, total_weights: 65536, loss: 7.908, lm_loss: 7.908, perplexity: 2720.116
I0421 06:29:53.212261 132163276240704 metric_logger.py:196] completed step: 10, seconds: 0.156, TFLOP/s/device: 87.359, Tokens/s/device: 13167.708, total_weights: 65536, loss: 7.568, lm_loss: 7.568, perplexity: 1936.198
I0421 06:29:53.370654 132163276240704 metric_logger.py:196] completed step: 11, seconds: 4.459, TFLOP/s/device: 3.047, Tokens/s/device: 459.271, total_weights: 65536, loss: 7.288, lm_loss: 7.288, perplexity: 1463.272
I0421 06:29:53.929492 132163276240704 metric_logger.py:196] completed step: 12, seconds: 0.006, TFLOP/s/device: 2242.841, Tokens/s/device: 338065.368, total_weights: 65536, loss: 7.032, lm_loss: 7.032, perplexity: 1132.061
I0421 06:29:54.087828 132163276240704 metric_logger.py:196] completed step: 13, seconds: 0.556, TFLOP/s/device: 24.435, Tokens/s/device: 3683.182, total_weights: 65536, loss: 6.815, lm_loss: 6.815, perplexity: 911.031
I0421 06:29:54.245974 132163276240704 metric_logger.py:196] completed step: 14, seconds: 0.162, TFLOP/s/device: 83.664, Tokens/s/device: 12610.682, total_weights: 65536, loss: 6.635, lm_loss: 6.635, perplexity: 761.080
I0421 06:29:54.404190 132163276240704 metric_logger.py:196] completed step: 15, seconds: 0.158, TFLOP/s/device: 85.948, Tokens/s/device: 12954.974, total_weights: 65536, loss: 6.492, lm_loss: 6.492, perplexity: 659.915
I0421 06:29:54.562529 132163276240704 metric_logger.py:196] completed step: 16, seconds: 0.158, TFLOP/s/device: 85.898, Tokens/s/device: 12947.521, total_weights: 65536, loss: 6.380, lm_loss: 6.380, perplexity: 589.948
I0421 06:29:54.720666 132163276240704 metric_logger.py:196] completed step: 17, seconds: 0.158, TFLOP/s/device: 85.904, Tokens/s/device: 12948.339, total_weights: 65536, loss: 6.293, lm_loss: 6.293, perplexity: 540.941
I0421 06:29:54.878768 132163276240704 metric_logger.py:196] completed step: 18, seconds: 0.158, TFLOP/s/device: 85.821, Tokens/s/device: 12935.826, total_weights: 65536, loss: 6.223, lm_loss: 6.223, perplexity: 504.108
I0421 06:29:55.036547 132163276240704 checkpointing.py:772] Waiting for step 19 to finish before checkpoint...
I0421 06:29:55.037562 132163276240704 checkpointing.py:776] Waited 0.0010323524475097656 seconds for step 19 to finish before starting checkpointing.
I0421 06:29:55.039658 132163276240704 checkpoint_manager.py:2020] [process=7][thread=MainThread][step=10][wait_until_finished] Waiting for Save Finalize thread (save_finalize) to complete.
I0421 06:29:59.985913 132031614789376 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_061359/linen_xpk_main_20260421_061359_05_fp8/checkpoints/10/items/array_metadatas/process_7
I0421 06:30:36.254052 132030037747456 base_pytree_checkpoint_handler.py:130] [process=7] /jax/orbax/write/gbytes_per_sec: 36.643 MiB/s (total gbytes: 1.5 GiB) (time elapsed: 43.11244034767151 s) (per-host)
I0421 06:30:36.254178 132030037747456 async_checkpointer.py:90] [process=7][thread=async_save] 3 Handler Commit operations completed. Time taken: 43.052007s.
I0421 06:30:44.605104 132030037747456 async_checkpointer.py:160] [process=7][thread=async_save] Background save thread done. Time taken: 51.402917s.
I0421 06:30:44.605330 132032168445696 async_checkpointer.py:288] [process=7][thread=save_finalize] Done with waiting for background save thread=async_save.
I0421 06:30:44.605387 132032168445696 async_checkpointer.py:298] [process=7][thread=save_finalize] No errors found in background save thread=async_save.
I0421 06:30:44.605436 132032168445696 checkpoint_manager.py:2137] [process=7][thread=save_finalize][step=10] CheckpointManager Save Finalize is syncing with other hosts...
I0421 06:30:44.607124 132032168445696 checkpoint_manager.py:2146] [process=7][thread=save_finalize][step=10] CheckpointManager Save Finalize is done on all hosts.
I0421 06:30:44.607305 132163276240704 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 06:30:44.607433 132163276240704 checkpoint_manager.py:1452] Waiting for previous save to complete took 49.567781 seconds. If this number is high, consider checkpointing less frequently.
I0421 06:30:44.608998 132163276240704 checkpoint_manager.py:1512] [process=7] Saving checkpoint at step 19
I0421 06:30:44.610955 132163276240704 event_tracking.py:70] [process=7] [async] Started save checkpoint @ gs://lance-maxtext/linen_ckpt_xpk_main_20260421_061359/linen_xpk_main_20260421_061359_05_fp8/checkpoints/19.
I0421 06:30:45.572218 132163276240704 jax_array_handlers.py:360] Scheduling D2H of 81 prioritized jax.Array.
I0421 06:30:45.572316 132163276240704 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 06:30:45.625988 132163276240704 base_pytree_checkpoint_handler.py:154] [process=7][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 0.054770s
I0421 06:30:45.626160 132163276240704 base_pytree_checkpoint_handler.py:130] [process=7] /jax/orbax/write/blocking_gbytes_per_sec: 25.268 GiB/s (total gbytes: 1.5 GiB) (time elapsed: 0.061055660247802734 s) (per-host)
I0421 06:30:45.626213 132163276240704 base_pytree_checkpoint_handler.py:768] [process=7][thread=MainThread] Initiated Pytree async_save. Time taken: 0.061118s (batch_requests_ready=0.003225s, total_serialization_initiated=0.057826s, others=0.000067s)
I0421 06:30:45.626311 132163276240704 composite_checkpoint_handler.py:715] [process=7][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 0.065336s (all_items=0.000011s, per_item={'items': '0.00001073'}, temp_paths=0.065325)
I0421 06:30:45.627056 132163276240704 event_tracking.py:125] [process=7] [async] Finished blocking save in 1.02 seconds. Continuing save @ gs://lance-maxtext/linen_ckpt_xpk_main_20260421_061359/linen_xpk_main_20260421_061359_05_fp8/checkpoints/19.
I0421 06:30:45.627313 132031614789376 async_checkpointer.py:76] [process=7][thread=async_save] Background save thread started. Deadline for this save operation is 2026-04-21 06:50:45.627291
I0421 06:30:45.629249 132163276240704 checkpoint_manager.py:1560] [process=7][thread=MainThread][step=19] Starting CheckpointManager Save Finalize thread=save_finalize
I0421 06:30:45.629610 132032168445696 async_checkpointer.py:280] [process=7][thread=save_finalize] Waiting for background save thread=async_save.
I0421 06:30:45.629754 132163276240704 standard_logger.py:34] {'step': 19, 'event_type': 'save', 'directory': 'gs://lance-maxtext/linen_ckpt_xpk_main_20260421_061359/linen_xpk_main_20260421_061359_05_fp8/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776752995.0396278, 'wait_for_prev_duration_secs': 49.56778120994568, 'time_between_consecutive_saves_sec': 2.6149816513061523, 'checkpointer_blocking_start_time': 1776753044.6090386, 'checkpointer_blocking_duration_secs': 1.018376111984253, 'get_old_steps_start_time': 1776753045.6274323, 'get_old_steps_duration_secs': 2.574920654296875e-05, 'checkpoint_manager_blocking_start_time': 1776752995.0378532, 'checkpoint_manager_blocking_duration_secs': 50.59185981750488}
I0421 06:30:45.629923 132163276240704 checkpointing.py:408] Started an asynchronous checkpoint save for step 19
I0421 06:30:45.629968 132163276240704 checkpoint_manager.py:2020] [process=7][thread=MainThread][step=19][wait_until_finished] Waiting for Save Finalize thread (save_finalize) to complete.
I0421 06:30:50.685370 132026950739712 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_061359/linen_xpk_main_20260421_061359_05_fp8/checkpoints/19/items/array_metadatas/process_7
I0421 06:31:27.395551 132031614789376 base_pytree_checkpoint_handler.py:130] [process=7] /jax/orbax/write/gbytes_per_sec: 37.766 MiB/s (total gbytes: 1.5 GiB) (time elapsed: 41.83040142059326 s) (per-host)
I0421 06:31:27.395679 132031614789376 async_checkpointer.py:90] [process=7][thread=async_save] 3 Handler Commit operations completed. Time taken: 41.768289s.
I0421 06:31:35.979569 132031614789376 async_checkpointer.py:160] [process=7][thread=async_save] Background save thread done. Time taken: 50.352162s.
I0421 06:31:35.979911 132032168445696 async_checkpointer.py:288] [process=7][thread=save_finalize] Done with waiting for background save thread=async_save.
I0421 06:31:35.980038 132032168445696 async_checkpointer.py:298] [process=7][thread=save_finalize] No errors found in background save thread=async_save.
I0421 06:31:35.980079 132032168445696 checkpoint_manager.py:2137] [process=7][thread=save_finalize][step=19] CheckpointManager Save Finalize is syncing with other hosts...
I0421 06:31:35.982316 132032168445696 checkpoint_manager.py:2146] [process=7][thread=save_finalize][step=19] CheckpointManager Save Finalize is done on all hosts.
I0421 06:31:35.982473 132163276240704 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 06:31:35.982618 132163276240704 checkpoint_manager.py:2009] [process=7][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning.
I0421 06:31:35.983495 132163276240704 metric_logger.py:196] completed step: 19, seconds: 0.158, TFLOP/s/device: 85.902, Tokens/s/device: 12948.012, 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 06:31:48 UTC 2026
EXIT_CODE=0

feat/nnx-post-train-fixes · e27fc1e97 · feat_nnx_post_train_fixes_20260421_114057 · full log

XPK Start: Tue Apr 21 11:54:44 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 11:55:07.922156: 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 11:55:08.458266 137145608030016 max_utils.py:273] Attempting to initialize the jax distributed system...
I0421 11:55:17.499878 137145608030016 distributed.py:149] Starting JAX distributed service on [::]:8482
I0421 11:55:17.502157 137145608030016 distributed.py:172] Connecting to JAX distributed service on mt-05-fp8-4nww1-slice-job-0-0.mt-05-fp8-4nww1:8482
I0421 11:55:19.062081 137145608030016 max_utils.py:284] Jax distributed system initialized!
I0421 11:55:25.188642 137145608030016 max_utils.py:800] System Information: Jax Version: 0.9.2
I0421 11:55:25.188760 137145608030016 max_utils.py:801] System Information: Jaxlib Version: 0.9.2
I0421 11:55:25.188801 137145608030016 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 11:55:25.188837 137145608030016 train_utils.py:378] WARNING: Sequence packing is essentially ignored for synthetic data. Please use a real dataset to use sequence packing.
I0421 11:55:25.884599 137145608030016 maxtext_utils.py:1718] Num_devices: 32, shape (1, 1, 1, 32, 1, 1, 1, 1, 1, 1, 1, 1, 1)
I0421 11:55:25.885189 137145608030016 maxtext_utils.py:1718] Num_devices: 32, shape (1, 1, 1, 32, 1, 1, 1, 1, 1, 1, 1, 1, 1)
I0421 11:55:25.885370 137145608030016 checkpointing.py:688] Setting up checkpoint logger...
I0421 11:55:25.885425 137145608030016 checkpointing.py:234] Creating checkpoint manager with ocdbt=True and zarr3=True
I0421 11:55:25.885473 137145608030016 pytree_checkpoint_handler.py:592] save_device_host_concurrent_bytes=None
I0421 11:55:25.885815 137145608030016 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 0x7cbafb64c9b0>, enable_pinned_host_transfer=False, save_concurrent_bytes: 96000000000 (89.4 GiB), restore_concurrent_bytes: 96000000000 (89.4 GiB)
I0421 11:55:28.832622 137145608030016 checkpointing.py:266] Enabling policy for fixed interval checkpointing.
I0421 11:55:28.832995 137145608030016 checkpoint_manager.py:708] [process=4][thread=MainThread] CheckpointManager init: checkpointers=None, item_names=('items',), item_handlers={'items': <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7ca5e8752ea0>}, handler_registry=None
I0421 11:55:28.833227 137145608030016 composite_checkpoint_handler.py:237] Deferred registration for item: "items". Adding handler `<orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7ca5e8752ea0>` 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 11:55:28.833276 137145608030016 composite_checkpoint_handler.py:237] Deferred registration for item: "metrics". Adding handler `<orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7ca5e8755f10>` 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 11:55:28.833311 137145608030016 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 0x7ca5e8752ea0>, ('items', <class 'orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeRestoreArgs'>): <orbax.checkpoint._src.handlers.pytree_checkpoint_handler.PyTreeCheckpointHandler object at 0x7ca5e8752ea0>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonSaveArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7ca5e8755f10>, ('metrics', <class 'orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonRestoreArgs'>): <orbax.checkpoint._src.handlers.json_checkpoint_handler.JsonCheckpointHandler object at 0x7ca5e8755f10>}).
I0421 11:55:28.833639 137145608030016 abstract_checkpointer.py:35] orbax-checkpoint version: 0.11.34
I0421 11:55:28.833721 137145608030016 async_checkpointer.py:192] [process=4][thread=MainThread] Using barrier_sync_fn: <function get_barrier_sync_fn.<locals>._fn at 0x7ca5e84f0180> timeout: 1200 secs and primary_host=0 for async checkpoint writes
I0421 11:55:30.402199 137145608030016 checkpoint_manager.py:1812] Found 0 checkpoint steps in gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114057/linen_xpk_feat_nnx_post_train_fixes_20260421_114057_05_fp8/checkpoints
I0421 11:55:30.813775 137145608030016 checkpoint_manager.py:929] [process=4][thread=MainThread] CheckpointManager created,  primary_host=0, CheckpointManagerOptions=CheckpointManagerOptions(save_interval_steps=1, max_to_keep=None, keep_time_interval=None, keep_period=None, should_keep_fn=None, best_fn=None, best_mode='max', keep_checkpoints_without_metrics=True, step_prefix=None, step_format_fixed_length=None, step_name_format=None, create=True, cleanup_tmp_directories=False, save_on_steps=frozenset(), single_host_load_and_broadcast=False, todelete_subdir=None, todelete_full_path=None, enable_background_delete=False, read_only=False, enable_async_checkpointing=True, async_options=None, multiprocessing_options=MultiprocessingOptions(primary_host=0, active_processes=None, barrier_sync_key_prefix=None), should_save_fn=None, file_options=FileOptions(path_permission_mode=None), save_root_metadata=True, temporary_path_class=None, save_decision_policy=FixedIntervalPolicy(interval=10), preservation_policy=LatestN(n=None), prevent_write_metrics=False, enable_should_save_is_saving_in_progress_check=True, enable_per_process_directory_creation=False, lightweight_initialize=False), root_directory=gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114057/linen_xpk_feat_nnx_post_train_fixes_20260421_114057_05_fp8/checkpoints: <orbax.checkpoint.checkpoint_manager.CheckpointManager object at 0x7ca6c82429c0>
I0421 11:55:30.813931 137145608030016 checkpointing.py:302] Checkpoint manager created!
I0421 11:55:31.915634 137145608030016 nnx_wrappers.py:437] Unknown Logical: bfloat16[32,2048,2048]...................................... ('activation_batch', 'activation_norm_length', 'activation_embed').
I0421 11:55:31.915751 137145608030016 nnx_wrappers.py:437] Unknown Physical: bfloat16[32,2048,2048]...................................... ('fsdp', None, None).
I0421 11:55:32.294453 137145608030016 attentions.py:1088] attentions/inputs_q Logical: bfloat16[32,2048,2048]...................................... ('activation_batch', 'activation_attn_length', 'activation_attn_embed').
I0421 11:55:32.294542 137145608030016 attentions.py:1088] attentions/inputs_q Physical: bfloat16[32,2048,2048]...................................... ('fsdp', None, None).
I0421 11:55:32.310728 137145608030016 attentions.py:1089] attentions/inputs_kv Logical: bfloat16[32,2048,2048]...................................... ('activation_batch', 'activation_attn_length', 'activation_attn_embed').
I0421 11:55:32.310784 137145608030016 attentions.py:1089] attentions/inputs_kv Physical: bfloat16[32,2048,2048]...................................... ('fsdp', None, None).
I0421 11:55:32.363005 137145608030016 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 11:55:32.363090 137145608030016 attentions.py:1154] attentions/query Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None).
I0421 11:55:32.379459 137145608030016 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 11:55:32.379519 137145608030016 attentions.py:1155] attentions/key Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None).
I0421 11:55:32.395780 137145608030016 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 11:55:32.395840 137145608030016 attentions.py:1156] attentions/value Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None).
I0421 11:55:32.420335 137145608030016 attentions.py:1197] attentions/out Logical: bfloat16[32,2048,16,128].................................... ('activation_batch', 'activation_attn_length', 'activation_heads', 'activation_kv').
I0421 11:55:32.420398 137145608030016 attentions.py:1197] attentions/out Physical: bfloat16[32,2048,16,128].................................... ('fsdp', None, None, None).
I0421 11:55:32.471717 137145608030016 linears.py:525] linears/x Logical: bfloat16[32,2048,7168]...................................... ('activation_batch', 'activation_length', 'activation_mlp').
I0421 11:55:32.471794 137145608030016 linears.py:525] linears/x Physical: bfloat16[32,2048,7168]...................................... ('fsdp', None, None).
I0421 11:55:32.909113 137145608030016 checkpointing.py:578] checkpoint manager exists so trying to load this run's existing checkpoint
I0421 11:55:32.909232 137145608030016 checkpointing.py:676] No existing checkpoints found, not restoring checkpoint.
fsdp: 32
I0421 11:55:34.824510 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_0/Fp8DirectDotGeneralOp_0/input_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.824625 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_0/Fp8DirectDotGeneralOp_0/input_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.824673 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_0/Fp8DirectDotGeneralOp_0/kernel_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.824727 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_0/Fp8DirectDotGeneralOp_0/kernel_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.824764 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_0/Fp8DirectDotGeneralOp_0/output_grad_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.824798 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_0/Fp8DirectDotGeneralOp_0/output_grad_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.824830 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_1/Fp8DirectDotGeneralOp_0/input_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.824862 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_1/Fp8DirectDotGeneralOp_0/input_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.824892 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_1/Fp8DirectDotGeneralOp_0/kernel_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.824921 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_1/Fp8DirectDotGeneralOp_0/kernel_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.824951 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_1/Fp8DirectDotGeneralOp_0/output_grad_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.824980 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/mlp/wi_1/Fp8DirectDotGeneralOp_0/output_grad_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825009 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/mlp/wo/Fp8DirectDotGeneralOp_0/input_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825038 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/mlp/wo/Fp8DirectDotGeneralOp_0/input_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825068 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/mlp/wo/Fp8DirectDotGeneralOp_0/kernel_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825100 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/mlp/wo/Fp8DirectDotGeneralOp_0/kernel_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825130 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/mlp/wo/Fp8DirectDotGeneralOp_0/output_grad_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825158 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/mlp/wo/Fp8DirectDotGeneralOp_0/output_grad_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825190 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/self_attention/key/Fp8DirectDotGeneralOp_0/input_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825241 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/self_attention/key/Fp8DirectDotGeneralOp_0/input_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825272 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/self_attention/key/Fp8DirectDotGeneralOp_0/kernel_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825300 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/self_attention/key/Fp8DirectDotGeneralOp_0/kernel_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825328 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/self_attention/key/Fp8DirectDotGeneralOp_0/output_grad_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825356 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/self_attention/key/Fp8DirectDotGeneralOp_0/output_grad_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825383 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/self_attention/out/Fp8DirectDotGeneralOp_0/input_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825410 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/self_attention/out/Fp8DirectDotGeneralOp_0/input_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825437 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/self_attention/out/Fp8DirectDotGeneralOp_0/kernel_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825464 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/self_attention/out/Fp8DirectDotGeneralOp_0/kernel_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825491 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/self_attention/out/Fp8DirectDotGeneralOp_0/output_grad_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825518 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/self_attention/out/Fp8DirectDotGeneralOp_0/output_grad_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825546 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/self_attention/query/Fp8DirectDotGeneralOp_0/input_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825573 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/self_attention/query/Fp8DirectDotGeneralOp_0/input_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825600 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/self_attention/query/Fp8DirectDotGeneralOp_0/kernel_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825627 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/self_attention/query/Fp8DirectDotGeneralOp_0/kernel_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825654 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/self_attention/query/Fp8DirectDotGeneralOp_0/output_grad_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825682 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/self_attention/query/Fp8DirectDotGeneralOp_0/output_grad_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825756 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/self_attention/value/Fp8DirectDotGeneralOp_0/input_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825790 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/self_attention/value/Fp8DirectDotGeneralOp_0/input_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825819 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/self_attention/value/Fp8DirectDotGeneralOp_0/kernel_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825847 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/self_attention/value/Fp8DirectDotGeneralOp_0/kernel_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825874 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/self_attention/value/Fp8DirectDotGeneralOp_0/output_grad_amax_history
    Shape:     float32[16,1024]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825901 137145608030016 maxtext_utils.py:1827]  params/_overwrite_with_gradient/decoder/layers/self_attention/value/Fp8DirectDotGeneralOp_0/output_grad_scale
    Shape:     float32[16,1]
    Logical:   P()
    Physical:  ()
I0421 11:55:34.825952 137145608030016 maxtext_utils.py:1827]  params/params/decoder/decoder_norm/scale
    Shape:     float32[2048]
    Logical:   P('norm',)
    Physical:  (None,)
I0421 11:55:34.826004 137145608030016 maxtext_utils.py:1827]  params/params/decoder/layers/mlp/wi_0/kernel
    Shape:     float32[2048,16,7168]
    Logical:   P('embed', 'layers', 'mlp')
    Physical:  ('fsdp', None, None)
I0421 11:55:34.826042 137145608030016 maxtext_utils.py:1827]  params/params/decoder/layers/mlp/wi_1/kernel
    Shape:     float32[2048,16,7168]
    Logical:   P('embed', 'layers', 'mlp')
    Physical:  ('fsdp', None, None)
I0421 11:55:34.826090 137145608030016 maxtext_utils.py:1827]  params/params/decoder/layers/mlp/wo/kernel
    Shape:     float32[7168,16,2048]
    Logical:   P('mlp', 'layers', 'embed')
    Physical:  (None, None, 'fsdp')
I0421 11:55:34.826136 137145608030016 maxtext_utils.py:1827]  params/params/decoder/layers/post_self_attention_layer_norm/scale
    Shape:     float32[2048,16]
    Logical:   P('norm', 'layers')
    Physical:  (None, None)
I0421 11:55:34.826168 137145608030016 maxtext_utils.py:1827]  params/params/decoder/layers/pre_self_attention_layer_norm/scale
    Shape:     float32[2048,16]
    Logical:   P('norm', 'layers')
    Physical:  (None, None)
I0421 11:55:34.826221 137145608030016 maxtext_utils.py:1827]  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 11:55:34.826272 137145608030016 maxtext_utils.py:1827]  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 11:55:34.826309 137145608030016 maxtext_utils.py:1827]  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 11:55:34.826343 137145608030016 maxtext_utils.py:1827]  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 11:55:34.826388 137145608030016 maxtext_utils.py:1827]  params/params/decoder/logits_dense/kernel
    Shape:     float32[2048,32000]
    Logical:   P('embed_vocab', 'vocab')
    Physical:  ('fsdp', None)
I0421 11:55:34.826432 137145608030016 maxtext_utils.py:1827]  params/params/token_embedder/embedding
    Shape:     float32[32000,2048]
    Logical:   P('vocab', 'embed_vocab')
    Physical:  (None, 'fsdp')

I0421 11:55:36.458385 137145608030016 train.py:157] train/xent Logical: float32[32,2048]............................................ ('activation_embed_and_logits_batch', 'activation_length').
I0421 11:55:36.458490 137145608030016 train.py:157] train/xent Physical: float32[32,2048]............................................ ('fsdp', None).
I0421 11:55:36.473890 137145608030016 train.py:164] train/z_loss Logical: float32[32,2048]............................................ ('activation_embed_and_logits_batch', 'activation_length').
I0421 11:55:36.473955 137145608030016 train.py:164] train/z_loss Physical: float32[32,2048]............................................ ('fsdp', None).
I0421 11:55:49.925239 137145608030016 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 11:55:49.926102 137145608030016 metric_logger.py:301] number parameters: 1.105 billion
I0421 11:56:05.092931 137145608030016 checkpointing.py:794] Waiting for step 0 to finish before checkpoint...
I0421 11:56:05.262457 137145608030016 checkpointing.py:798] Waited 0.16951203346252441 seconds for step 0 to finish before starting checkpointing.
I0421 11:56:05.265083 137145608030016 checkpoint_manager.py:2009] [process=4][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning.
I0421 11:56:05.266856 137145608030016 checkpoint_manager.py:1512] [process=4] Saving checkpoint at step 0
I0421 11:56:05.268426 137145608030016 event_tracking.py:70] [process=4] [async] Started save checkpoint @ gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114057/linen_xpk_feat_nnx_post_train_fixes_20260421_114057_05_fp8/checkpoints/0.
I0421 11:56:05.605294 137145608030016 signaling_client.py:364] Using JaxDistributedSignalingClient
I0421 11:56:05.606241 137145608030016 jax_array_handlers.py:360] Scheduling D2H of 81 prioritized jax.Array.
I0421 11:56:05.606357 137145608030016 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 11:56:06.019849 137145608030016 base_pytree_checkpoint_handler.py:154] [process=4][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 0.414674s
I0421 11:56:06.020020 137145608030016 base_pytree_checkpoint_handler.py:130] [process=4] /jax/orbax/write/blocking_gbytes_per_sec: 3.646 GiB/s (total gbytes: 1.5 GiB) (time elapsed: 0.42318201065063477 s) (per-host)
I0421 11:56:06.020075 137145608030016 base_pytree_checkpoint_handler.py:768] [process=4][thread=MainThread] Initiated Pytree async_save. Time taken: 0.423248s (batch_requests_ready=0.003652s, total_serialization_initiated=0.419522s, others=0.000074s)
I0421 11:56:06.020180 137145608030016 composite_checkpoint_handler.py:715] [process=4][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 0.427330s (all_items=0.000019s, per_item={'items': '0.00001860'}, temp_paths=0.427311)
I0421 11:56:06.021004 137145608030016 event_tracking.py:125] [process=4] [async] Finished blocking save in 0.75 seconds. Continuing save @ gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114057/linen_xpk_feat_nnx_post_train_fixes_20260421_114057_05_fp8/checkpoints/0.
I0421 11:56:06.021349 137016231679744 async_checkpointer.py:76] [process=4][thread=async_save] Background save thread started. Deadline for this save operation is 2026-04-21 12:16:06.021310
I0421 11:56:06.032199 137145608030016 checkpoint_manager.py:1560] [process=4][thread=MainThread][step=0] Starting CheckpointManager Save Finalize thread=save_finalize
I0421 11:56:06.032507 137016204396288 async_checkpointer.py:280] [process=4][thread=save_finalize] Waiting for background save thread=async_save.
I0421 11:56:06.032661 137145608030016 standard_logger.py:34] {'step': 0, 'event_type': 'save', 'directory': 'gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114057/linen_xpk_feat_nnx_post_train_fixes_20260421_114057_05_fp8/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776772565.2650647, 'wait_for_prev_duration_secs': 6.389617919921875e-05, 'time_between_consecutive_saves_sec': None, 'checkpointer_blocking_start_time': 1776772565.2668931, 'checkpointer_blocking_duration_secs': 0.7546074390411377, 'get_old_steps_start_time': 1776772566.021524, 'get_old_steps_duration_secs': 3.0517578125e-05, 'checkpoint_manager_blocking_start_time': 1776772565.2630558, 'checkpoint_manager_blocking_duration_secs': 0.7695667743682861}
I0421 11:56:06.032782 137145608030016 checkpointing.py:409] Started an asynchronous checkpoint save for step 0
I0421 11:56:06.032834 137145608030016 max_utils.py:750] 
Memstats: After params initialized:
I0421 11:56:06.032886 137145608030016 max_utils.py:756] 	Using (GB) 0.45 / 31.25 (1.440000%) on TPU_16(process=4,(0,4,0,0))
I0421 11:56:06.032919 137145608030016 max_utils.py:756] 	Using (GB) 0.45 / 31.25 (1.440000%) on TPU_17(process=4,(1,4,0,0))
I0421 11:56:06.032946 137145608030016 max_utils.py:756] 	Using (GB) 0.45 / 31.25 (1.440000%) on TPU_20(process=4,(0,5,0,0))
I0421 11:56:06.032971 137145608030016 max_utils.py:756] 	Using (GB) 0.45 / 31.25 (1.440000%) on TPU_21(process=4,(1,5,0,0))
I0421 11:56:06.347244 137145608030016 metric_logger.py:196] completed step: 0, seconds: 15.167, TFLOP/s/device: 0.896, Tokens/s/device: 135.033, total_weights: 65536, loss: 10.874, lm_loss: 10.874, perplexity: 52776.805
I0421 11:56:06.535559 137145608030016 metric_logger.py:196] completed step: 1, seconds: 1.252, TFLOP/s/device: 10.850, Tokens/s/device: 1635.387, total_weights: 65536, loss: 10.874, lm_loss: 10.874, perplexity: 52761.840
I0421 11:56:06.958270 137145608030016 metric_logger.py:196] completed step: 2, seconds: 0.031, TFLOP/s/device: 441.169, Tokens/s/device: 66497.825, total_weights: 65536, loss: 10.816, lm_loss: 10.816, perplexity: 49832.398
I0421 11:56:07.116400 137145608030016 metric_logger.py:196] completed step: 3, seconds: 0.423, TFLOP/s/device: 32.122, Tokens/s/device: 4841.814, total_weights: 65536, loss: 10.431, lm_loss: 10.431, perplexity: 33901.820
I0421 11:56:07.433610 137145608030016 metric_logger.py:196] completed step: 4, seconds: 0.164, TFLOP/s/device: 82.971, Tokens/s/device: 12506.336, total_weights: 65536, loss: 9.992, lm_loss: 9.992, perplexity: 21847.639
I0421 11:56:07.439606 137145608030016 metric_logger.py:196] completed step: 5, seconds: 0.158, TFLOP/s/device: 85.994, Tokens/s/device: 12961.943, total_weights: 65536, loss: 9.549, lm_loss: 9.549, perplexity: 14035.006
I0421 11:56:09.298443    2850 google_auth_provider.cc:181] Running on GCE, using service account 562977990677-compute@developer.gserviceaccount.com
I0421 11:56:11.258092 137016212788992 array_metadata_store.py:203] [process=4][thread=array_type_handler] Wrote 81 array_metadata.ArrayMetadata to gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114057/linen_xpk_feat_nnx_post_train_fixes_20260421_114057_05_fp8/checkpoints/0/items/array_metadatas/process_4
I0421 11:56:26.006079 137016231679744 base_pytree_checkpoint_handler.py:130] [process=4] /jax/orbax/write/gbytes_per_sec: 77.404 MiB/s (total gbytes: 1.5 GiB) (time elapsed: 20.409206867218018 s) (per-host)
I0421 11:56:26.006202 137016231679744 async_checkpointer.py:90] [process=4][thread=async_save] 3 Handler Commit operations completed. Time taken: 19.984741s.
I0421 11:56:31.398446 137145608030016 metric_logger.py:196] completed step: 6, seconds: 0.318, TFLOP/s/device: 42.778, Tokens/s/device: 6448.038, total_weights: 65536, loss: 9.111, lm_loss: 9.111, perplexity: 9058.688
I0421 11:56:31.556929 137145608030016 metric_logger.py:196] completed step: 7, seconds: 23.802, TFLOP/s/device: 0.571, Tokens/s/device: 86.045, total_weights: 65536, loss: 8.685, lm_loss: 8.685, perplexity: 5914.720
I0421 11:56:31.715339 137145608030016 metric_logger.py:196] completed step: 8, seconds: 0.163, TFLOP/s/device: 83.272, Tokens/s/device: 12551.635, total_weights: 65536, loss: 8.281, lm_loss: 8.281, perplexity: 3950.010
I0421 11:56:31.720130 137145608030016 checkpointing.py:794] Waiting for step 10 to finish before checkpoint...
I0421 11:56:32.032146 137145608030016 checkpointing.py:798] Waited 0.3119814395904541 seconds for step 10 to finish before starting checkpointing.
I0421 11:56:32.035110 137145608030016 checkpoint_manager.py:2020] [process=4][thread=MainThread][step=0][wait_until_finished] Waiting for Save Finalize thread (save_finalize) to complete.
I0421 11:56:33.766348 137016231679744 async_checkpointer.py:160] [process=4][thread=async_save] Background save thread done. Time taken: 27.744870s.
I0421 11:56:33.766615 137016204396288 async_checkpointer.py:288] [process=4][thread=save_finalize] Done with waiting for background save thread=async_save.
I0421 11:56:33.766766 137016204396288 async_checkpointer.py:298] [process=4][thread=save_finalize] No errors found in background save thread=async_save.
I0421 11:56:33.766820 137016204396288 checkpoint_manager.py:2137] [process=4][thread=save_finalize][step=0] CheckpointManager Save Finalize is syncing with other hosts...
I0421 11:56:33.768384 137016204396288 checkpoint_manager.py:2146] [process=4][thread=save_finalize][step=0] CheckpointManager Save Finalize is done on all hosts.
I0421 11:56:33.768574 137145608030016 checkpoint_manager.py:2032] [process=4][thread=MainThread][step=0][wait_until_finished] Done waiting for Save Finalize thread (save_finalize) running at step=0.
W0421 11:56:33.768731 137145608030016 checkpoint_manager.py:1452] Waiting for previous save to complete took 1.733601 seconds. If this number is high, consider checkpointing less frequently.
I0421 11:56:33.770545 137145608030016 checkpoint_manager.py:1512] [process=4] Saving checkpoint at step 10
I0421 11:56:33.772632 137145608030016 event_tracking.py:70] [process=4] [async] Started save checkpoint @ gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114057/linen_xpk_feat_nnx_post_train_fixes_20260421_114057_05_fp8/checkpoints/10.
I0421 11:56:34.484388 137145608030016 jax_array_handlers.py:360] Scheduling D2H of 81 prioritized jax.Array.
I0421 11:56:34.484482 137145608030016 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 11:56:34.538142 137145608030016 base_pytree_checkpoint_handler.py:154] [process=4][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 0.054633s
I0421 11:56:34.538312 137145608030016 base_pytree_checkpoint_handler.py:130] [process=4] /jax/orbax/write/blocking_gbytes_per_sec: 25.256 GiB/s (total gbytes: 1.5 GiB) (time elapsed: 0.06108450889587402 s) (per-host)
I0421 11:56:34.538362 137145608030016 base_pytree_checkpoint_handler.py:768] [process=4][thread=MainThread] Initiated Pytree async_save. Time taken: 0.061144s (batch_requests_ready=0.003333s, total_serialization_initiated=0.057745s, others=0.000067s)
I0421 11:56:34.538465 137145608030016 composite_checkpoint_handler.py:715] [process=4][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 0.065253s (all_items=0.000017s, per_item={'items': '0.00001669'}, temp_paths=0.065237)
I0421 11:56:34.539217 137145608030016 event_tracking.py:125] [process=4] [async] Finished blocking save in 0.77 seconds. Continuing save @ gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114057/linen_xpk_feat_nnx_post_train_fixes_20260421_114057_05_fp8/checkpoints/10.
I0421 11:56:34.539557 137012979939072 async_checkpointer.py:76] [process=4][thread=async_save] Background save thread started. Deadline for this save operation is 2026-04-21 12:16:34.539520
I0421 11:56:34.541568 137145608030016 checkpoint_manager.py:1560] [process=4][thread=MainThread][step=10] Starting CheckpointManager Save Finalize thread=save_finalize
I0421 11:56:34.541943 137016204396288 async_checkpointer.py:280] [process=4][thread=save_finalize] Waiting for background save thread=async_save.
I0421 11:56:34.542079 137145608030016 standard_logger.py:34] {'step': 10, 'event_type': 'save', 'directory': 'gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114057/linen_xpk_feat_nnx_post_train_fixes_20260421_114057_05_fp8/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776772592.0350802, 'wait_for_prev_duration_secs': 1.7336010932922363, 'time_between_consecutive_saves_sec': None, 'checkpointer_blocking_start_time': 1776772593.7705843, 'checkpointer_blocking_duration_secs': 0.7691340446472168, 'get_old_steps_start_time': 1776772594.5397418, 'get_old_steps_duration_secs': 3.0994415283203125e-05, 'checkpoint_manager_blocking_start_time': 1776772592.0328681, 'checkpoint_manager_blocking_duration_secs': 2.5091769695281982}
I0421 11:56:34.542255 137145608030016 checkpointing.py:409] Started an asynchronous checkpoint save for step 10
I0421 11:56:34.542843 137145608030016 metric_logger.py:196] completed step: 9, seconds: 0.158, TFLOP/s/device: 85.763, Tokens/s/device: 12927.171, total_weights: 65536, loss: 7.908, lm_loss: 7.908, perplexity: 2720.116
I0421 11:56:34.549892 137145608030016 metric_logger.py:196] completed step: 10, seconds: 0.158, TFLOP/s/device: 85.751, Tokens/s/device: 12925.376, total_weights: 65536, loss: 7.568, lm_loss: 7.568, perplexity: 1936.198
I0421 11:56:34.708269 137145608030016 metric_logger.py:196] completed step: 11, seconds: 2.828, TFLOP/s/device: 4.805, Tokens/s/device: 724.294, total_weights: 65536, loss: 7.288, lm_loss: 7.288, perplexity: 1463.272
I0421 11:56:34.866472 137145608030016 metric_logger.py:196] completed step: 12, seconds: 0.006, TFLOP/s/device: 2241.731, Tokens/s/device: 337898.037, total_weights: 65536, loss: 7.032, lm_loss: 7.032, perplexity: 1132.061
I0421 11:56:35.418172 137145608030016 metric_logger.py:196] completed step: 13, seconds: 0.159, TFLOP/s/device: 85.239, Tokens/s/device: 12848.100, total_weights: 65536, loss: 6.815, lm_loss: 6.815, perplexity: 911.031
I0421 11:56:35.576601 137145608030016 metric_logger.py:196] completed step: 14, seconds: 0.547, TFLOP/s/device: 24.838, Tokens/s/device: 3743.887, total_weights: 65536, loss: 6.635, lm_loss: 6.635, perplexity: 761.080
I0421 11:56:35.735311 137145608030016 metric_logger.py:196] completed step: 15, seconds: 0.163, TFLOP/s/device: 83.433, Tokens/s/device: 12575.913, total_weights: 65536, loss: 6.492, lm_loss: 6.492, perplexity: 659.915
I0421 11:56:35.893397 137145608030016 metric_logger.py:196] completed step: 16, seconds: 0.158, TFLOP/s/device: 85.739, Tokens/s/device: 12923.500, total_weights: 65536, loss: 6.380, lm_loss: 6.380, perplexity: 589.948
I0421 11:56:36.051824 137145608030016 metric_logger.py:196] completed step: 17, seconds: 0.159, TFLOP/s/device: 85.707, Tokens/s/device: 12918.690, total_weights: 65536, loss: 6.293, lm_loss: 6.293, perplexity: 540.941
I0421 11:56:36.210159 137145608030016 metric_logger.py:196] completed step: 18, seconds: 0.158, TFLOP/s/device: 85.906, Tokens/s/device: 12948.667, total_weights: 65536, loss: 6.223, lm_loss: 6.223, perplexity: 504.108
I0421 11:56:36.367809 137145608030016 checkpointing.py:794] Waiting for step 19 to finish before checkpoint...
I0421 11:56:36.368815 137145608030016 checkpointing.py:798] Waited 0.0010249614715576172 seconds for step 19 to finish before starting checkpointing.
I0421 11:56:36.371027 137145608030016 checkpoint_manager.py:2020] [process=4][thread=MainThread][step=10][wait_until_finished] Waiting for Save Finalize thread (save_finalize) to complete.
I0421 11:56:40.251822 137014598944512 array_metadata_store.py:203] [process=4][thread=array_type_handler] Wrote 81 array_metadata.ArrayMetadata to gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114057/linen_xpk_feat_nnx_post_train_fixes_20260421_114057_05_fp8/checkpoints/10/items/array_metadatas/process_4
I0421 11:57:17.045736 137012979939072 base_pytree_checkpoint_handler.py:130] [process=4] /jax/orbax/write/gbytes_per_sec: 37.111 MiB/s (total gbytes: 1.5 GiB) (time elapsed: 42.56844091415405 s) (per-host)
I0421 11:57:17.045864 137012979939072 async_checkpointer.py:90] [process=4][thread=async_save] 3 Handler Commit operations completed. Time taken: 42.506187s.
I0421 11:57:25.174582 137012979939072 async_checkpointer.py:160] [process=4][thread=async_save] Background save thread done. Time taken: 50.634889s.
I0421 11:57:25.174882 137016204396288 async_checkpointer.py:288] [process=4][thread=save_finalize] Done with waiting for background save thread=async_save.
I0421 11:57:25.175016 137016204396288 async_checkpointer.py:298] [process=4][thread=save_finalize] No errors found in background save thread=async_save.
I0421 11:57:25.175065 137016204396288 checkpoint_manager.py:2137] [process=4][thread=save_finalize][step=10] CheckpointManager Save Finalize is syncing with other hosts...
I0421 11:57:25.176632 137016204396288 checkpoint_manager.py:2146] [process=4][thread=save_finalize][step=10] CheckpointManager Save Finalize is done on all hosts.
I0421 11:57:25.176843 137145608030016 checkpoint_manager.py:2032] [process=4][thread=MainThread][step=10][wait_until_finished] Done waiting for Save Finalize thread (save_finalize) running at step=10.
W0421 11:57:25.176972 137145608030016 checkpoint_manager.py:1452] Waiting for previous save to complete took 48.805949 seconds. If this number is high, consider checkpointing less frequently.
I0421 11:57:25.178524 137145608030016 checkpoint_manager.py:1512] [process=4] Saving checkpoint at step 19
I0421 11:57:25.180688 137145608030016 event_tracking.py:70] [process=4] [async] Started save checkpoint @ gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114057/linen_xpk_feat_nnx_post_train_fixes_20260421_114057_05_fp8/checkpoints/19.
I0421 11:57:25.881733 137145608030016 jax_array_handlers.py:360] Scheduling D2H of 81 prioritized jax.Array.
I0421 11:57:25.881827 137145608030016 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 11:57:25.936812 137145608030016 base_pytree_checkpoint_handler.py:154] [process=4][thread=MainThread] Initiated "orbax.checkpoint._src.serialization.jax_array_handlers.ArrayHandler".serialize. Time taken: 0.055991s
I0421 11:57:25.936972 137145608030016 base_pytree_checkpoint_handler.py:130] [process=4] /jax/orbax/write/blocking_gbytes_per_sec: 24.712 GiB/s (total gbytes: 1.5 GiB) (time elapsed: 0.062429189682006836 s) (per-host)
I0421 11:57:25.937020 137145608030016 base_pytree_checkpoint_handler.py:768] [process=4][thread=MainThread] Initiated Pytree async_save. Time taken: 0.062488s (batch_requests_ready=0.003265s, total_serialization_initiated=0.059159s, others=0.000064s)
I0421 11:57:25.937135 137145608030016 composite_checkpoint_handler.py:715] [process=4][thread=MainThread] Initiated CompositeCheckpointHandler.async_save. Time taken: 0.066545s (all_items=0.000011s, per_item={'items': '0.00001121'}, temp_paths=0.066534)
I0421 11:57:25.937918 137145608030016 event_tracking.py:125] [process=4] [async] Finished blocking save in 0.76 seconds. Continuing save @ gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114057/linen_xpk_feat_nnx_post_train_fixes_20260421_114057_05_fp8/checkpoints/19.
I0421 11:57:25.938243 137014598944512 async_checkpointer.py:76] [process=4][thread=async_save] Background save thread started. Deadline for this save operation is 2026-04-21 12:17:25.938210
I0421 11:57:25.940201 137145608030016 checkpoint_manager.py:1560] [process=4][thread=MainThread][step=19] Starting CheckpointManager Save Finalize thread=save_finalize
I0421 11:57:25.940497 137016204396288 async_checkpointer.py:280] [process=4][thread=save_finalize] Waiting for background save thread=async_save.
I0421 11:57:25.940633 137145608030016 standard_logger.py:34] {'step': 19, 'event_type': 'save', 'directory': 'gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114057/linen_xpk_feat_nnx_post_train_fixes_20260421_114057_05_fp8/checkpoints', 'reached_preemption': False, 'preemption_received_at': None, 'synchronous': False, 'wait_for_prev_start_time': 1776772596.370999, 'wait_for_prev_duration_secs': 48.80594873428345, 'time_between_consecutive_saves_sec': 2.602570056915283, 'checkpointer_blocking_start_time': 1776772645.178563, 'checkpointer_blocking_duration_secs': 0.75982666015625, 'get_old_steps_start_time': 1776772645.9384093, 'get_old_steps_duration_secs': 2.956390380859375e-05, 'checkpoint_manager_blocking_start_time': 1776772596.3690963, 'checkpoint_manager_blocking_duration_secs': 49.57150411605835}
I0421 11:57:25.940812 137145608030016 checkpointing.py:409] Started an asynchronous checkpoint save for step 19
I0421 11:57:25.940855 137145608030016 checkpoint_manager.py:2020] [process=4][thread=MainThread][step=19][wait_until_finished] Waiting for Save Finalize thread (save_finalize) to complete.
I0421 11:57:30.981530 137013525202688 array_metadata_store.py:203] [process=4][thread=array_type_handler] Wrote 81 array_metadata.ArrayMetadata to gs://lance-maxtext/linen_ckpt_xpk_feat_nnx_post_train_fixes_20260421_114057/linen_xpk_feat_nnx_post_train_fixes_20260421_114057_05_fp8/checkpoints/19/items/array_metadatas/process_4
I0421 11:58:07.457842 137014598944512 base_pytree_checkpoint_handler.py:130] [process=4] /jax/orbax/write/gbytes_per_sec: 37.990 MiB/s (total gbytes: 1.5 GiB) (time elapsed: 41.583253383636475 s) (per-host)
I0421 11:58:07.457966 137014598944512 async_checkpointer.py:90] [process=4][thread=async_save] 3 Handler Commit operations completed. Time taken: 41.519611s.
I0421 11:58:16.874091 137014598944512 async_checkpointer.py:160] [process=4][thread=async_save] Background save thread done. Time taken: 50.935719s.
I0421 11:58:16.874400 137016204396288 async_checkpointer.py:288] [process=4][thread=save_finalize] Done with waiting for background save thread=async_save.
I0421 11:58:16.874528 137016204396288 async_checkpointer.py:298] [process=4][thread=save_finalize] No errors found in background save thread=async_save.
I0421 11:58:16.874569 137016204396288 checkpoint_manager.py:2137] [process=4][thread=save_finalize][step=19] CheckpointManager Save Finalize is syncing with other hosts...
I0421 11:58:16.876093 137016204396288 checkpoint_manager.py:2146] [process=4][thread=save_finalize][step=19] CheckpointManager Save Finalize is done on all hosts.
I0421 11:58:16.876275 137145608030016 checkpoint_manager.py:2032] [process=4][thread=MainThread][step=19][wait_until_finished] Done waiting for Save Finalize thread (save_finalize) running at step=19.
I0421 11:58:16.876425 137145608030016 checkpoint_manager.py:2009] [process=4][thread=MainThread][wait_until_finished] No Save Finalize thread to wait for. Returning.
I0421 11:58:16.877520 137145608030016 metric_logger.py:196] completed step: 19, seconds: 0.159, TFLOP/s/device: 85.682, Tokens/s/device: 12914.943, 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 11:58:29 UTC 2026
EXIT_CODE=0

Metrics: main (e27fc1e97) vs feat/nnx-post-train-fixes (e27fc1e97)

Metrics: `main (e27fc1e97)` vs `feat/nnx-post-train-fixes (e27fc1e97)`