[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>

tests/models/multimodal/processing/test_common.py

@@ -198,13 +198,17 @@ def get_text_token_prompts(
             mm_counts,
             mm_options={},
         )
-        assert isinstance(inputs.prompt, str)
-
-        text_prompt = inputs.prompt
-        token_prompt = tokenizer.encode(
-            text_prompt,
-            add_special_tokens=_ADD_SPECIAL_TOKENS_OVERRIDES.get(model_type, True),
-        )
+        # Some models (e.g., Kimi-Audio) return token IDs directly instead of str
+        if isinstance(inputs.prompt, list):
+            text_prompt = None
+            token_prompt = inputs.prompt
+        else:
+            assert isinstance(inputs.prompt, str)
+            text_prompt = inputs.prompt
+            token_prompt = tokenizer.encode(
+                text_prompt,
+                add_special_tokens=_ADD_SPECIAL_TOKENS_OVERRIDES.get(model_type, True),
+            )
 
     return text_prompt, token_prompt
 

tests/models/registry.py

@@ -857,6 +857,15 @@ _MULTIMODAL_EXAMPLE_MODELS = {
         "Kwai-Keye/Keye-VL-1_5-8B",
         trust_remote_code=True,
     ),
+    "MoonshotKimiaForCausalLM": _HfExamplesInfo(
+        "moonshotai/Kimi-Audio-7B-Instruct",
+        tokenizer_mode="kimi_audio",
+        trust_remote_code=True,
+    ),
+    "KimiK25ForConditionalGeneration": _HfExamplesInfo(
+        "moonshotai/Kimi-K2.5",
+        trust_remote_code=True,
+    ),
     "KimiVLForConditionalGeneration": _HfExamplesInfo(
         "moonshotai/Kimi-VL-A3B-Instruct",
         extras={"thinking": "moonshotai/Kimi-VL-A3B-Thinking"},
@@ -870,10 +879,6 @@ _MULTIMODAL_EXAMPLE_MODELS = {
             )
         },
     ),
-    "KimiK25ForConditionalGeneration": _HfExamplesInfo(
-        "moonshotai/Kimi-K2.5",
-        trust_remote_code=True,
-    ),
     "LightOnOCRForConditionalGeneration": _HfExamplesInfo(
         "lightonai/LightOnOCR-1B-1025"
     ),

tests/models/test_initialization.py

@@ -103,6 +103,12 @@ def can_initialize(
             "pickle error when loading `transformers.models.auto.CONFIG_MAPPING`"
         )
 
+    if model_arch == "MoonshotKimiaForCausalLM":
+        pytest.skip(
+            "Kimi-Audio requires SpeechToTextConfig "
+            "which is not configured in test environment"
+        )
+
     if model_arch in ["DeepseekV32ForCausalLM", "GlmMoeDsaForCausalLM"]:
         from vllm.platforms import current_platform