Fix missing rms_norm_eps in DeepseekV3 MLA layernorms

[mvanhorn]

What does this PR do?

Passes eps=config.rms_norm_eps to both q_a_layernorm and kv_a_layernorm in the DeepseekV3 MLA attention module. Without this, these layernorms default to eps=1e-5 instead of the config value (1e-6), causing precision differences compared to vLLM and SGLang implementations.

The fix was applied to modular_deepseek_v3.py and propagated to generated modeling files (deepseek_v3, glm4_moe_lite, longcat_flash, youtu) via make fix-repo.

Note: DeepseekV2 has the same issue but is left for a separate PR to keep this focused.

Fixes #44261

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case. [Bug/Discussion] MLA q_a_layernorm Missing config.rms_norm_eps, Causing 1e-5/1e-6 Precision Error #44261
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

Who can review?

@ArthurZucker @Cyrilvallez (text models, attention)

This contribution was developed with AI assistance (Claude Code).

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: deepseek_v3, glm4_moe_lite, longcat_flash, youtu

[alvinttang]

The fix is correct — without passing eps=config.rms_norm_eps, the MLA layernorms for both q_a_layernorm and kv_a_layernorm would silently use the RMSNorm default epsilon (typically 1e-6) instead of the model-configured value, which would cause subtle numerical divergence from reference implementations without any error. It is worth noting this same fix is applied consistently across all derived models (glm4_moe_lite, longcat_flash, youtu) and the modular source, which is the right approach. One open question: does the DeepseekV3RMSNorm default epsilon happen to match the typical config value, masking this bug in practice? If so, a unit test asserting the epsilon is correctly propagated would prevent future regressions of this class.

[Cyrilvallez]

cc @ArthurZucker, do you know if the norms in the Attention are supposed to use the config value as well or not?

[mvanhorn]

The MLA attention layernorms (q_a_layernorm and kv_a_layernorm) are the only layernorms in DeepseekV3 that weren't getting config.rms_norm_eps passed through - all other norms in the model (input, post-attention, MLP) already use it. This PR just makes them consistent.

The same pattern is applied across the derived models (glm4_moe_lite, longcat_flash, youtu) and the modular source, so it stays consistent everywhere.