[MaxText] Fix the MLPerf parallelism, precision and config_filename mllog disclosure for Lineage runs

[copybara-service]

[MaxText] Fix the MLPerf parallelism, precision and config_filename mllog disclosure for Lineage runs

mllog_utils.init_print derived the v6.1 disclosure keys from the named
ici_*_parallelism and quantization fields. The Lineage recipe
(deepseek3-671b-lineage.yml) uses neither: it runs on a physical
[dcn, x, y, z, core] mesh with explicit logical_axis_rules and quantizes
through lineage_quantization. The 2026-10-01 Lineage runs therefore logged
expert_parallelism=1, bfloat16 for linear/comm, and every run logged the
placeholder config_filename=config.yml.

For use_lineage runs:

• expert_parallelism is the mesh size of the exp rule
  (x * y * core = 32).
• tensor_parallelism is the mesh size of the activation_length rule
  (the TensorCore pair, 2). Lineage's MLA up/out projections are head-sharded
  across it, and the splash kernel sees the full sequence (Megatron TP + SP),
  so this is not context parallelism.
• precision falls back to lineage_quantization. fp8_full quantizes
  the routed-expert GMMs and the EP token all-gather
  (QuantConfig.routed_experts, dsv3_sparse_layer.ubatch_dispatch), so both
  linear and comm log fp8.

For all runs, config_filename is <model_name>.yml.

Non-Lineage parallelism and precision are unchanged.

[codecov]

Codecov Report

❌ Patch coverage is 90.90909% with 2 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/maxtext/utils/mllog_utils.py	90.90%	1 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!