Qwen3-Omni][Bugfix] Replace vLLM fused layers with HF-compatible numerics in code predictor

[LJH-LBJ]

PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTTOM) HAVE BEEN CONSIDERED.

Purpose

Resolve: #2286
Fixes audio quality degradation (noise after ~3000 tokens) in Qwen3-Omni's code predictor by replacing vLLM's fused kernels with plain PyTorch equivalents that match HuggingFace reference numerics.

This is the Qwen3-Omni counterpart of PR #2277 (which fixed the same issue for Qwen3-TTS). The Qwen3-Omni code predictor was written following PR #1617's approach and inherits the same precision bug.

Component	Before (precision issue)	After (HF-compatible)
RMSNorm	vLLM RMSNorm (bf16 variance)	Custom _RMSNorm (float32 variance)
RoPE	vLLM get_rope (bf16 cos/sin)	Custom _RotaryEmbedding (float32 cos/sin, torch.autocast(enabled=False))
QKV projection	QKVParallelLinear (fused)	Separate nn.Linear for q/k/v
MLP	MergedColumnParallelLinear + RowParallelLinear	Separate nn.Linear for gate/up/down
torch.compile	mode="default"	options={"epilogue_fusion": False}
position_ids	[bsz * seq_len] 1D flat	[bsz, seq_len] 2D (HF format)
load_weights	stacked_params_mapping remapping	Direct loading (weight names match HF)

Test Plan

curl -s http://localhost:46354/v1/chat/completions -H "Content-Type: application/json" -d '{
"model": "/workspace/models/Qwen3-Omni-30B-A3B-Instruct",
"messages": [{"role": "user", "content": "Please write a 5000-word novel."}],
"modalities": ["audio"]
}' | jq -r '.choices[0].message.audio.data' | base64 -d > output2.wav

Test Result

There is no noise in the audio. It cannot upload because its size is too big

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan. Please provide the test scripts & test commands. Please state the reasons if your codes don't require additional test scripts. For test file guidelines, please check the test style doc
• The test results. Please paste the results comparison before and after, or the e2e results.
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model. Please run mkdocs serve to sync the documentation editions to ./docs.
• (Optional) Release notes update. If your change is user-facing, please update the release notes draft.

BEFORE SUBMITTING, PLEASE READ https://github.com/vllm-project/vllm-omni/blob/main/CONTRIBUTING.md (anything written below this line will be removed by GitHub Actions)

[amy-why-3459]

Please add a nightly-test label. @Gaohan123 @david6666666 @gcanlin @hsliuustc0106

[amy-why-3459]

We are working on an accuracy benchmark. Could you add a long-output use case to monitor this scenario?

[amy-why-3459]

@Sy0307 PTAL

[LJH-LBJ]

We are working on an accuracy benchmark. Could you add a long-output use case to monitor this scenario?

Sure, I can do it in next pr.

[Sy0307]

LGTM.