[GLM-ASR] GLM-ASR Support (#15570)

Co-authored-by: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com>

python/sglang/srt/multimodal/processors/base_processor.py

@@ -299,6 +299,7 @@ class BaseMultimodalProcessor(ABC):
         if audios:
             if self._processor.__class__.__name__ in {
                 "Gemma3nProcessor",
+                "GlmasrProcessor",
                 "Qwen2AudioProcessor",
                 "Qwen3OmniMoeProcessor",
             }:
@@ -800,7 +801,6 @@ class BaseMultimodalProcessor(ABC):
         # Process items and get input_ids
         all_collected_items: list[MultimodalDataItem] = []
         input_ids = None
-
         # Handle raw items (need processing)
         if raw_images or raw_audios or raw_videos:
             collected_items, input_ids, ret = self._process_and_collect_mm_items(

python/sglang/srt/multimodal/processors/glmasr.py

@@ -0,0 +1,53 @@
+import re
+
+from sglang.srt.models.glmasr import GlmasrForConditionalGeneration
+from sglang.srt.multimodal.processors.base_processor import (
+    BaseMultimodalProcessor,
+    MultimodalSpecialTokens,
+)
+
+
+class GlmasrProcessor(BaseMultimodalProcessor):
+    models = [GlmasrForConditionalGeneration]
+
+    def __init__(self, hf_config, server_args, _processor, *args, **kwargs):
+        super().__init__(hf_config, server_args, _processor, *args, **kwargs)
+        self.AUDIO_TOKEN = "<|begin_of_audio|><|pad|><|end_of_audio|>"
+        self.AUDIO_TOKEN_REGEX = re.compile(
+            r"<\|begin_of_audio\|><\|pad\|><\|end_of_audio\|>"
+        )
+        # Collect special token ids
+        tokenizer = self._processor.tokenizer
+        self.audio_start_id = tokenizer.convert_tokens_to_ids("<|begin_of_audio|>")
+        self.audio_token_id = tokenizer.convert_tokens_to_ids("<|pad|>")
+        self.audio_end_id = tokenizer.convert_tokens_to_ids("<|end_of_audio|>")
+
+        self.mm_tokens = MultimodalSpecialTokens(
+            audio_token=self.AUDIO_TOKEN,
+            audio_token_regex=self.AUDIO_TOKEN_REGEX,
+            audio_token_id=self.audio_token_id,
+        ).build(_processor)
+
+    async def process_mm_data_async(
+        self,
+        audio_data,
+        input_text,
+        **kwargs,
+    ):
+        base_output = self.load_mm_data(
+            prompt=input_text,
+            audio_data=audio_data,
+            multimodal_tokens=self.mm_tokens,
+        )
+        if base_output is None:
+            return None
+        mm_items, input_ids, ret = self.process_and_combine_mm_data(
+            base_output, self.mm_tokens
+        )
+        return {
+            "mm_items": mm_items,
+            "input_ids": input_ids.tolist(),
+            "audio_start_id": self.audio_start_id,
+            "audio_token_id": self.audio_token_id,
+            "audio_end_id": self.audio_end_id,
+        }