FEAT: [model] qwen3-omni (#4137)

Author: qinxuye

xinference/model/llm/transformers/multimodal/qwen-omni.py

@@ -19,6 +19,8 @@
 from threading import Thread
 from typing import Any, Dict, Iterator, List, Optional, Tuple
 
+import torch
+
 from .....types import (
     ChatCompletion,
     ChatCompletionAudio,
@@ -35,20 +37,31 @@
 
 @register_transformer
 @register_non_default_model("qwen2.5-omni")
-class Qwen2_5OmniChatModel(PytorchMultiModalModel):
+@register_non_default_model("Qwen3-Omni-Thinking")
+@register_non_default_model("Qwen3-Omni-Instruct")
+class QwenOmniChatModel(PytorchMultiModalModel):
     DEFAULT_SYSTEM_PROMPT = (
         "You are Qwen, a virtual human developed by the Qwen Team, Alibaba Group, "
         "capable of perceiving auditory and visual inputs, as well as generating text and speech."
     )
 
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        # 2.5 or 3
+        model_family = self.model_family.model_family or self.model_family.model_name
+        self._omni_version = "2.5" if "2.5" in model_family else "3"
+
     @classmethod
     def match_json(
         cls, model_family: "LLMFamilyV2", model_spec: "LLMSpecV1", quantization: str
     ) -> bool:
         if model_spec.model_format not in ["pytorch", "gptq", "awq", "bnb"]:
             return False
         llm_family = model_family.model_family or model_family.model_name
-        if "qwen2.5-omni".lower() in llm_family.lower():
+        if (
+            "qwen2.5-omni".lower() in llm_family.lower()
+            or "qwen3-omni".lower() in llm_family.lower()
+        ):
             return True
         return False
 
@@ -58,15 +71,25 @@ def decide_device(self):
         self._device = device
 
     def load_processor(self):
-        from transformers import Qwen2_5OmniProcessor
+        if self._omni_version == "2.5":
+            from transformers import Qwen2_5OmniProcessor as QwenOminiProcessor
+        else:
+            from transformers import Qwen3OmniMoeProcessor as QwenOminiProcessor
 
-        self._processor = Qwen2_5OmniProcessor.from_pretrained(
+        self._processor = QwenOminiProcessor.from_pretrained(
             self.model_path, trust_remote_code=True
         )
         self._tokenizer = self._processor.tokenizer
 
     def load_multimodal_model(self):
-        from transformers import Qwen2_5OmniForConditionalGeneration
+        if self._omni_version == "2.5":
+            from transformers import (
+                Qwen2_5OmniForConditionalGeneration as QwenOmniForConditionalGeneration,
+            )
+        else:
+            from transformers import (
+                Qwen3OmniMoeForConditionalGeneration as QwenOmniForConditionalGeneration,
+            )
 
         # for multiple GPU, set back to auto to make multiple devices work
         device = "auto" if self._device == "cuda" else self._device
@@ -79,7 +102,7 @@ def load_multimodal_model(self):
         kwargs = self.apply_bnb_quantization(kwargs)
         logger.debug("Loading model with extra kwargs: %s", kwargs)
 
-        self._model = Qwen2_5OmniForConditionalGeneration.from_pretrained(
+        self._model = QwenOmniForConditionalGeneration.from_pretrained(
             self.model_path,
             torch_dtype="auto",
             device_map=device,
@@ -181,11 +204,37 @@ def generate_non_streaming(
         inputs = self.build_inputs_from_messages(messages, generate_config)  # type: ignore
         use_audio_in_video = generate_config.get("use_audio_in_video", True)
         gen_kwargs = dict(**inputs, **config, use_audio_in_video=use_audio_in_video)
-        generated_ids, audio = self._model.generate(**gen_kwargs)
-        generated_ids_trimmed = [
-            out_ids[len(in_ids) :]
-            for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
-        ]
+        # === Run model.generate() (handle both (ids, audio) and ids-only cases) ===
+        result = self._model.generate(**gen_kwargs)
+        if isinstance(result, tuple) and len(result) == 2:
+            # Qwen2.5-Omni returns (generated_ids, audio)
+            generated_ids, audio = result
+        else:
+            # Qwen3-Omni returns only generated_ids
+            generated_ids, audio = result, None
+        if hasattr(generated_ids, "sequences"):
+            generated_ids = generated_ids.sequences
+
+        # === Handle text decoding ===
+        input_len = inputs.input_ids.shape[1]
+        # Ensure we have a consistent 2D structure
+        # Normalize to list[list[int]]
+        if isinstance(generated_ids, torch.Tensor):
+            generated_ids = generated_ids.tolist()
+        elif isinstance(generated_ids, list) and all(
+            isinstance(x, int) for x in generated_ids
+        ):
+            # Single sequence as flat list of ints
+            generated_ids = [generated_ids]
+        elif isinstance(generated_ids, list) and all(
+            isinstance(x, list) for x in generated_ids
+        ):
+            pass  # already correct
+        else:
+            raise TypeError(f"Unexpected generated_ids type: {type(generated_ids)}")
+
+        # Remove prompt tokens
+        generated_ids_trimmed = [out_ids[input_len:] for out_ids in generated_ids]
         output_text = self._processor.batch_decode(
             generated_ids_trimmed,
             skip_special_tokens=True,