ci: remove CosyVoice3 post-merge test

[linyueqian]

Summary

• Remove CosyVoice3 post-merge (advanced_model) e2e test from test-merge.yml
• Zero-shot voice cloning quality is too variable for ASR similarity checks
• Pre-merge core_model tests already verify the pipeline produces audio

Test plan

• CI passes without the post-merge test block

[chatgpt-codex-connector]

💡 Codex Review

vllm-omni/vllm_omni/model_executor/models/cosyvoice3/cosyvoice3.py

Lines 514 to 517 in c39621b

	weighted_scores = weighted_scores.clone()
	weighted_scores[top_id] = float("-inf")
	fallback_probs = weighted_scores.softmax(dim=0)
	top_id = int(cls._multinomial_sample(fallback_probs, generator=generator).item())

Skip RAS fallback when no alternative token exists

When repetition rejection triggers here, top_id is set to -inf and fallback sampling is always attempted. If constraints/logits processors leave only one valid token (for example an allowed_token_ids mask with a single id), this makes all logits -inf, so softmax becomes invalid and torch.multinomial can fail at runtime. This regression is introduced by the new model-side sampler path and can crash constrained CosyVoice3 decoding requests instead of returning a token.

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[Gaohan123]

LGTM