[Bugfix] Fix the issue where DeepSeek v3.2 cannot use structured_output (#30371)

Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>

vllm/v1/structured_output/backend_xgrammar.py

@@ -10,7 +10,7 @@ import torch
 import vllm.envs
 from vllm.logger import init_logger
 from vllm.sampling_params import SamplingParams
-from vllm.tokenizers import MistralTokenizer
+from vllm.tokenizers import DeepseekV32Tokenizer, MistralTokenizer
 from vllm.utils.import_utils import LazyLoader
 from vllm.v1.structured_output.backend_types import (
     StructuredOutputBackend,
@@ -56,6 +56,27 @@ class XgrammarBackend(StructuredOutputBackend):
                 stop_token_ids=stop_token_ids,
                 add_prefix_space=True,
             )
+        elif isinstance(self.tokenizer, DeepseekV32Tokenizer):
+            # copy from xgr.TokenizerInfo.from_huggingface()
+            # because we are using a custom tokenizer wrapper here.
+            vocab_dict = self.tokenizer.get_vocab()
+            tokenizer_vocab_size = max(len(vocab_dict), self.tokenizer.max_token_id + 1)
+            vocab_size = self.vocab_size or tokenizer_vocab_size
+            # maintain tokenizer's indexing
+            encoded_vocab = [""] * vocab_size
+            for token, idx in vocab_dict.items():
+                if idx < vocab_size:
+                    encoded_vocab[idx] = token
+            stop_token_ids = [self.tokenizer.eos_token_id]
+            backend_str = self.tokenizer.tokenizer.backend_tokenizer.to_str()
+            metadata = xgr.TokenizerInfo._detect_metadata_from_hf(backend_str)
+            tokenizer_info = xgr.TokenizerInfo(
+                encoded_vocab=encoded_vocab,
+                vocab_type=metadata["vocab_type"],
+                vocab_size=vocab_size,
+                stop_token_ids=stop_token_ids,
+                add_prefix_space=metadata["add_prefix_space"],
+            )
         else:
             tokenizer_info = xgr.TokenizerInfo.from_huggingface(
                 self.tokenizer,