Whisper model support & /v1/audio/transcriptions endpoint & benchmark (#16983)

Signed-off-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
Co-authored-by: MahmoudAshraf97 <hassouna97.ma@gmail.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

python/sglang/srt/parser/conversation.py

@@ -1027,6 +1027,23 @@ register_conv_template(
     )
 )
 
+# Whisper speech-to-text template
+# Whisper uses special tokens: <|startoftranscript|>, <|en|>, <|transcribe|>, etc.
+# Audio features are processed by encoder separately, not inserted into text
+# The decoder start tokens (task, language) should be set via generation config
+register_conv_template(
+    Conversation(
+        name="whisper",
+        system_template="",
+        system_message="",
+        roles=("", ""),
+        sep_style=SeparatorStyle.NO_COLON_SINGLE,
+        sep="",
+        stop_str=["<|endoftext|>"],
+        audio_token="",  # Empty - audio is handled by encoder, not as text token
+    )
+)
+
 MODEL_TYPE_TO_TEMPLATE = {
     "internvl_chat": "internvl-2-5",
     "deepseek_vl_v2": "deepseek-vl2",
@@ -1036,6 +1053,7 @@ MODEL_TYPE_TO_TEMPLATE = {
     "minicpmo": "minicpmo",
     "deepseek-ocr": "deepseek-ocr",
     "paddleocr_vl": "paddle-ocr",
+    "whisper": "whisper",
 }
 
 
@@ -1129,3 +1147,11 @@ def match_paddle_ocr(model_path: str):
         return "paddle-ocr"
     model_type = get_model_type(model_path)
     return MODEL_TYPE_TO_TEMPLATE.get(model_type)
+
+
+@register_conv_template_matching_function
+def match_whisper(model_path: str):
+    if "whisper" in model_path.lower():
+        return "whisper"
+    model_type = get_model_type(model_path)
+    return MODEL_TYPE_TO_TEMPLATE.get(model_type)