fix(model): honor config epsilon in Post-LN row-parallel linear

[hokuyama0106]

What does this PR do ?

Pass config.layernorm_epsilon to the Post-LN built inside TERowParallelLinearLayerNorm, so that Gemma2 / Gemma3 / Gemma4's post-attention and post-feedforward RMSNorms use the model's epsilon (1e-6) instead of TENorm's 1e-5 default.

TENorm.__new__(cls, config, hidden_size, eps=1e-5, has_residual=False) does not read config.layernorm_epsilon; the epsilon has to be passed explicitly, as MCore itself does for the final layernorm (transformer_block.py: eps=self.config.layernorm_epsilon) and as this repo already does for the Gemma3-VL projector norm (modeling_gemma3_vl.py: TENorm(config, config.input_size, eps=config.layernorm_epsilon)).

Because TERowParallelLinearLayerNorm omitted it, the Post-LN modules were the only norms in the layer running at 1e-5:

Norm	epsilon before	epsilon after
input_layernorm (via TELayerNormColumnParallelLinear)	config.layernorm_epsilon	unchanged
pre_feedforward_layernorm (via TELayerNormColumnParallelLinear)	config.layernorm_epsilon	unchanged
final_layernorm (via MCore TransformerBlock)	config.layernorm_epsilon	unchanged
post_attention_layernorm (linear_proj)	1e-5	config.layernorm_epsilon
post_feedforward_layernorm (linear_fc2)	1e-5	config.layernorm_epsilon

Affected models (all users of TERowParallelLinearLayerNorm):

• Gemma2 (gemma2_provider.py: linear_proj, linear_fc2) — layernorm_epsilon = 1e-6 → behavior changes, now matches HF rms_norm_eps
• Gemma3 (gemma3_provider.py: linear_proj, linear_fc2) — layernorm_epsilon = 1e-6 → behavior changes, now matches HF rms_norm_eps
• Gemma4 (modeling_gemma4.py: linear_proj in gemma4_block_spec) — layernorm_epsilon = 1e-6 → behavior changes, now matches HF rms_norm_eps
• EXAONE 4.0 / 4.5 — rms_norm_eps = 1e-5, i.e. equal to the old hardcoded default, so no behavior change

The forward-pass difference is small (sqrt(mean(x^2) + eps) with eps 1e-5 vs 1e-6), but it applies to every Post-LN in every layer and shows up as a systematic HF <-> Megatron parity gap for the Gemma family.

Changelog

• src/megatron/bridge/models/common/te_layers.py: build the Post-LN as TENorm(config, output_size, eps=config.layernorm_epsilon)
• tests/unit_tests/models/common/test_te_layers.py: add test_post_layernorm_uses_config_epsilon, asserting the epsilon from the config reaches TENorm

GitHub Actions CI

Needs an NVIDIA developer to approve/trigger CI for this external contribution.

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation?
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
  • Reviewer: Does the PR have correct import guards for all optional libraries?

Additional Information

• The Post-LN epsilon has been unset since the module was introduced as TERowParallelLinearLayerNorm in gemma3_provider.py (Gemma3 Provider & Bridge #867) and moved unchanged into models/common/te_layers.py in feat: Add EXAONE 4.0 model bridge (LG AI Research) #2532.
• Gemma checkpoints converted/trained before this change will see a tiny numerical shift in the Post-LN outputs; no weights or checkpoint formats are affected.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.