Merge remote-tracking branch 'origin/main' into dev/vllm-align

Made-with: Cursor

# Conflicts:
#	vllm_omni/entrypoints/openai/serving_chat.py
#	vllm_omni/worker/gpu_model_runner.py

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

vllm_omni/model_executor/models/qwen3_omni/qwen3_omni.py

@@ -410,11 +410,16 @@ class Qwen3OmniMoeForConditionalGeneration(
             talker_hidden = model_outputs
             # merge the code_predictor_codes from the info_dict list into a single tensor
             multimodal_outputs: dict = None
-            # Here is the only place to use runtime_additional_information. After MTP in the
+            # Here is the only place to use model_intermediate_buffer. After MTP in the
             # preprocess function, the code_predictor_codes are stored in the info_dict list.
             # We need to merge the tensors from different requests into a single tensor.
             # In the future, we may allow user to custom an aggregated function.
-            info_dicts = kwargs.get("runtime_additional_information")
+            info_dicts = kwargs.get("model_intermediate_buffer")
+            if info_dicts is None:
+                info_dicts = kwargs.get("runtime_additional_information")
+
+            if "runtime_additional_information" in kwargs and "model_intermediate_buffer" not in kwargs:
+                logger.warning_once("runtime_additional_information is deprecated, use model_intermediate_buffer")
             code_predictor_codes = [info.get("code_predictor_codes") for info in info_dicts]
             multimodal_outputs = {"code_predictor_codes": torch.cat(code_predictor_codes, dim=0)}
             span_len = multimodal_outputs["code_predictor_codes"].shape[0]

vllm_omni/model_executor/models/qwen3_tts/qwen3_tts.py

@@ -13,6 +13,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import base64
+import copy
 import io
 import urllib.request
 from collections.abc import Iterable
@@ -151,9 +152,14 @@ class Qwen3TTSModelForGeneration(nn.Module):
 
         # Extract additional parameters from kwargs that the generation methods expect
 
-        runtime_additional_information = kwargs.get("runtime_additional_information", [{}])
+        runtime_additional_information = kwargs.get("model_intermediate_buffer")
+        if runtime_additional_information is None:
+            runtime_additional_information = kwargs.get("runtime_additional_information", [{}])
+        if "runtime_additional_information" in kwargs and "model_intermediate_buffer" not in kwargs:
+            logger.warning_once("runtime_additional_information is deprecated, use model_intermediate_buffer")
         if isinstance(runtime_additional_information, list) and len(runtime_additional_information) > 0:
             runtime_additional_information = runtime_additional_information[0]
+        runtime_additional_information = copy.deepcopy(runtime_additional_information)
         text = runtime_additional_information.pop("text", [""])[0]
         # Extract task_type from kwargs, default to self.task_type
         task_type = _normalize_task_type(runtime_additional_information.pop("task_type", [self.task_type])[0])

vllm_omni/model_executor/models/qwen3_tts/qwen3_tts_talker.py

@@ -332,7 +332,6 @@ class Qwen3TTSTalkerForConditionalGeneration(nn.Module):
         self.have_multimodal_outputs = True
         self.has_preprocess = True
         self.has_postprocess = True
-
         # Used by OmniGPUModelRunner for the GPU-side MTP fast-path.
         self.mtp_hidden_size = int(self.talker_config.hidden_size)
         # OmniGPUModelRunner will store talker_mtp output under this key in
@@ -441,7 +440,11 @@ class Qwen3TTSTalkerForConditionalGeneration(nn.Module):
             return model_outputs
 
         hidden = model_outputs
-        info_dicts = kwargs.get("runtime_additional_information") or []
+        info_dicts = kwargs.get("model_intermediate_buffer")
+        if info_dicts is None:
+            info_dicts = kwargs.get("runtime_additional_information") or []
+        if "runtime_additional_information" in kwargs and "model_intermediate_buffer" not in kwargs:
+            logger.warning_once("runtime_additional_information is deprecated, use model_intermediate_buffer")
         audio_codes_list: list[torch.Tensor] = []
         ref_code_len_list: list[torch.Tensor] = []
         codec_streaming_list: list[torch.Tensor] = []