rebase: align vllm-omni with vLLM ac9126e58aa7

Signed-off-by: tzhouam <tzhouam@connect.ust.hk>

tests/worker_v2/test_capture_query_bound_gpu.py

@@ -44,7 +44,7 @@ def _check_prefill():
     spec = FullAttentionSpec(block_size=16, num_kv_heads=8, head_size=128, dtype=torch.bfloat16)
     kvconfig = KVCacheConfig(num_blocks=1024, kv_cache_tensors=[], kv_cache_groups=[KVCacheGroupSpec(["attn"], spec)])
     state = OmniModelState.__new__(OmniModelState)
-    state.max_model_len = 256
+    state.model_config = SimpleNamespace(max_model_len=256)
     state.supports_mm_inputs = False
     with set_current_vllm_config(c):
         group = AttentionGroup(FlashAttentionBackend, ["attn"], spec, 0)