[Model] Add support for moonshotai/Kimi-Audio-7B-Instruct (#36127)

Signed-off-by: tunglinwood <tunglinwood@gmail.com>
Signed-off-by: tunglinwood <tomwu.tunglin@gmail.com>
Signed-off-by: tunglinwood <113751333+tunglinwood@users.noreply.github.com>

vllm/renderers/kimi_audio.py

@@ -0,0 +1,49 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+from typing import Any, cast
+
+from vllm.config import VllmConfig
+from vllm.tokenizers.kimi_audio import KimiAudioTokenizer
+from vllm.tokenizers.registry import get_tokenizer
+
+from .hf import HfRenderer, HfTokenizer
+
+
+class KimiAudioRenderer(HfRenderer):
+    """Renderer for Kimi-Audio models.
+
+    This renderer uses HfRenderer internally with a custom TikToken tokenizer.
+    """
+
+    @classmethod
+    def from_config(  # type: ignore[override]
+        cls,
+        config: VllmConfig,
+        tokenizer_kwargs: dict[str, Any],
+    ) -> "HfRenderer":
+        """Create an HfRenderer instance for Kimi-Audio models."""
+        model_config = config.model_config
+        if model_config.skip_tokenizer_init:
+            tokenizer = None
+        else:
+            # Extract tokenizer_name from kwargs (already processed by
+            # tokenizer_args_from_config for ModelScope/GGUF/etc)
+            tokenizer_name = tokenizer_kwargs.pop(
+                "tokenizer_name", model_config.tokenizer
+            )
+            # Remove tokenizer_cls from kwargs to avoid duplicate argument
+            tokenizer_kwargs = {
+                k: v for k, v in tokenizer_kwargs.items() if k != "tokenizer_cls"
+            }
+            # Use get_tokenizer directly instead of cached_get_tokenizer
+            # (KimiAudioTokenizer doesn't work with get_cached_tokenizer)
+            tokenizer = cast(
+                HfTokenizer,
+                get_tokenizer(
+                    tokenizer_name,
+                    tokenizer_cls=KimiAudioTokenizer,  # type: ignore[arg-type]
+                    **tokenizer_kwargs,
+                ),
+            )
+
+        return HfRenderer(config, tokenizer)

vllm/renderers/registry.py

@@ -19,6 +19,7 @@ _VLLM_RENDERERS = {
     "deepseek_v32": ("deepseek_v32", "DeepseekV32Renderer"),
     "hf": ("hf", "HfRenderer"),
     "grok2": ("grok2", "Grok2Renderer"),
+    "kimi_audio": ("kimi_audio", "KimiAudioRenderer"),
     "mistral": ("mistral", "MistralRenderer"),
     "qwen_vl": ("qwen_vl", "QwenVLRenderer"),
     "terratorch": ("terratorch", "TerratorchRenderer"),
@@ -74,10 +75,18 @@ RENDERER_REGISTRY = RendererRegistry(
 
 def renderer_from_config(config: "VllmConfig", **kwargs):
     model_config = config.model_config
+
     tokenizer_mode, tokenizer_name, args, kwargs = tokenizer_args_from_config(
         model_config, **kwargs
     )
 
+    # Override tokenizer_mode for Kimi-Audio models
+    if model_config.architecture == "MoonshotKimiaForCausalLM":
+        tokenizer_mode = "kimi_audio"
+        # Update model_config so other components (e.g., multimodal registry)
+        # also use the correct tokenizer mode
+        model_config.tokenizer_mode = "kimi_audio"
+
     if (
         model_config.tokenizer_mode == "auto"
         and model_config.model_impl == "terratorch"