fix: regression in text generate with LTXAV model

[kijai]

Somehow missed the presence_penalty in the LTXAV text model specifically.

[coderabbitai]

📝 Walkthrough

Walkthrough

This change adds a presence_penalty parameter to the text generation pipeline in the text encoder module. The parameter is introduced to two method signatures: Gemma3_12BModel.generate() and LTXAVTEModel.generate(). The parameter is then threaded through the call chain, propagating from LTXAVTEModel through Gemma3_12BModel to the underlying transformer's generation method. This enables presence penalty control to be passed down to the transformer during text generation operations.

🚥 Pre-merge checks | ✅ 2 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (2 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and specifically describes the main change: fixing a regression in the LTXAV text model by restoring the presence_penalty parameter.
Description check	✅ Passed	The description is directly related to the changeset, explaining that the presence_penalty parameter was omitted from the LTXAV text model and is now being restored.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Actionable comments posted: 2

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@comfy/text_encoders/lt.py`:
- Around line 192-193: The generate method in com­fy/text_encoders/lt.py should
provide a default for presence_penalty to preserve backward compatibility;
update the generate signature (function name: generate) to set
presence_penalty=0.0 and keep the internal call to
self.gemma3_12b.generate(tokens["gemma3_12b"], do_sample, max_length,
temperature, top_k, top_p, min_p, repetition_penalty, seed, presence_penalty)
unchanged so callers that omit presence_penalty continue to work.
- Around line 94-98: The generate method signature in class/function generate
currently adds presence_penalty without a default causing callers to break;
update the signature to add a default (presence_penalty=0.0) so it matches
sibling implementations (e.g., llama.py and sd.py) and maintain backward
compatibility, leaving the internal usage of presence_penalty unchanged in the
call to self.transformer.generate.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 248df038-1ee1-4bb7-abb9-fb69362fd3aa

📥 Commits

Reviewing files that changed from the base of the PR and between 2a1f402 and 6ef043e.

📒 Files selected for processing (1)

• comfy/text_encoders/lt.py