rebase: rename Qwen3-Omni audio attention projection to qkv_proj

vLLM 0.29 renamed the audio attention projection from qkv to qkv_proj, and
its stacked mapping now rewrites checkpoint q_proj/k_proj/v_proj onto that
name. Omni subclasses the upstream audio attention and mirrors its __init__
to add quant_config, so it kept building self.qkv and weight loading died
with "There is no module or parameter named layers.0.self_attn.qkv_proj in
Qwen3OmniMoeAudioEncoder ... available: self_attn.qkv.weight". The inherited
forward calls self.qkv_proj too, so the old name broke both paths.

This is a second, independent cause behind the Omni Accuracy Test failure:
the shutdown-timeout signature also appeared in that job, which is why the
earlier pass attributed it solely to the shutdown fix. Both builds 3009 and
3010 carry this weight-loading error, so it predates that pass.

Not reproducible locally: it needs the 30B checkpoint. The rename mirrors
upstream Qwen3OmniMoeAudioAttention exactly, and the 20 Qwen3-Omni unit tests
still pass.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0146Xu5Sgj4vXxLxvLjV5YPj
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>

vllm_omni/model_executor/models/qwen3_omni/qwen3_omni_moe_thinker.py

@@ -278,7 +278,12 @@ class Qwen3OmniMoeAudioAttention(_Qwen3OmniMoeAudioAttention):
 
         self.scaling = self.head_dim**-0.5
 
-        self.qkv = QKVParallelLinear(
+        # vLLM 0.29 renamed the audio attention projection to ``qkv_proj`` and
+        # its stacked mapping rewrites checkpoint ``q_proj``/``k_proj``/
+        # ``v_proj`` onto that name, so the module must match or weight loading
+        # fails with "no module or parameter named ...self_attn.qkv_proj".
+        # The inherited forward also calls ``self.qkv_proj``.
+        self.qkv_proj = QKVParallelLinear(
             hidden_size=self.embed_dim,
             head_size=self.head_dim,
             total_num_heads=self.num_heads,
@@ -286,7 +291,7 @@ class Qwen3OmniMoeAudioAttention(_Qwen3OmniMoeAudioAttention):
             bias=True,
             disable_tp=disable_tp,
             quant_config=quant_config,
-            prefix=f"{prefix}.qkv",
+            prefix=f"{prefix}.qkv_proj",
         )
 
         self.out_proj = RowParallelLinear(